Skip to content

User Guide

Command Line Interface

Ludwig provides six command line interface entry points

  • train
  • predict
  • test
  • experiment
  • visualize
  • collect_weights
  • collect_activations

They are described in detail below.

train

This command lets you train a model from your data. You can call it with:

ludwig train [options]

or with

python -m ludwig.train [options]

from within Ludwig's main directory.

These are the available arguments:

usage: ludwig train [options]

This script trains a model.

optional arguments:
  -h, --help            show this help message and exit
  --output_directory OUTPUT_DIRECTORY
                        directory that contains the results
  --experiment_name EXPERIMENT_NAME
                        experiment name
  --model_name MODEL_NAME
                        name for the model
  --data_csv DATA_CSV   input data CSV file. If it has a split column, it will
                        be used for splitting (0: train, 1: validation, 2:
                        test), otherwise the dataset will be randomly split
  --data_train_csv DATA_TRAIN_CSV
                        input train data CSV file
  --data_validation_csv DATA_VALIDATION_CSV
                        input validation data CSV file
  --data_test_csv DATA_TEST_CSV
                        input test data CSV file
  --data_hdf5 DATA_HDF5
                        input data HDF5 file. It is an intermediate preprocess
                        version of the input CSV created the first time a CSV
                        file is used in the same directory with the same name
                        and a hdf5 extension
  --data_train_hdf5 DATA_TRAIN_HDF5
                        input train data HDF5 file. It is an intermediate
                        preprocess version of the input CSV created the first
                        time a CSV file is used in the same directory with the
                        same name and a hdf5 extension
  --data_validation_hdf5 DATA_VALIDATION_HDF5
                        input validation data HDF5 file. It is an intermediate
                        preprocess version of the input CSV created the first
                        time a CSV file is used in the same directory with the
                        same name and a hdf5 extension
  --data_test_hdf5 DATA_TEST_HDF5
                        input test data HDF5 file. It is an intermediate
                        preprocess version of the input CSV created the first
                        time a CSV file is used in the same directory with the
                        same name and a hdf5 extension
  --train_set_metadata_json TRAIN_SET_METADATA_JSON
                        input metadata JSON file. It is an intermediate
                        preprocess file containing the mappings of the input
                        CSV created the first time a CSV file is used in the
                        same directory with the same name and a json extension
  -sspi, --skip_save_processed_input
                        skips saving intermediate HDF5 and JSON files
  -md MODEL_DEFINITION, --model_definition MODEL_DEFINITION
                        model definition
  -mdf MODEL_DEFINITION_FILE, --model_definition_file MODEL_DEFINITION_FILE
                        YAML file describing the model. Ignores
                        --model_hyperparameters
  -mlp MODEL_LOAD_PATH, --model_load_path MODEL_LOAD_PATH
                        path of a pretrained model to load as initialization
  -mrp MODEL_RESUME_PATH, --model_resume_path MODEL_RESUME_PATH
                        path of a the model directory to resume training of
  -ssm, --skip_save_model
                        disables saving weights each time the model imrpoves. By
                        default Ludwig saves weights after each epoch the
                        validation measure imrpvoes, but if the model is
                        really big that can be time consuming if you do not
                        want to keep the weights and just find out what
                        performance can a model get with a set of
                        hyperparameters, use this parameter to skip it.
  -ssp, --skip_save_progress
                        disables saving weights after each epoch. By default
                        ludwig saves weights after each epoch for enabling
                        resuming of training, but if the model is really big
                        that can be time consuming and will save twice as much
                        space, use this parameter to skip it.
  -ssl, --skip_save_log
                        disables saving TensorBoard logs. By default Ludwig
                        saves logs for the TensorBoard, but if it is not
                        needed turning it off can slightly increase the
                        overall speed.
  -rs RANDOM_SEED, --random_seed RANDOM_SEED
                        a random seed that is going to be used anywhere there
                        is a call to a random number generator: data
                        splitting, parameter initialization and training set
                        shuffling
  -g GPUS [GPUS ...], --gpus GPUS [GPUS ...]
                        list of gpus to use
  -gf GPU_FRACTION, --gpu_fraction GPU_FRACTION
                        fraction of gpu memory to initialize the process with
  -uh, --use_horovod    uses horovod for distributed training
  -dbg, --debug         enables debugging mode
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

When Ludwig trains a model it creates two intermediate files, one HDF5 and one JSON. The HDF5 file contains the data mapped to numpy ndarrays, while the JSON file contains the mappings from the values in the tensors to their original labels.

For instance, for a categorical feature with 3 possible values, the HDF5 file will contain integers from 0 to 3 (with 0 being a <UNK> category), while the JSON file will contain a idx2str list containing all tokens ([<UNK>, label_1, label_2, label_3]), a str2idx dictionary ({"<UNK>": 0, "label_1": 1, "label_2": 2, "label_3": 3}) and a str2freq dictionary ({"<UNK>": 0, "label_1": 93, "label_2": 55, "label_3": 24}).

The reason to have those intermediate files is two-fold: on one hand, if you are going to train your model again Ludwig will try to load them instead of recomputing all tensors, which saves a consistent amount of time, and on the other hand when you want to use your model to predict, data has to be mapped to tensors in exactly the same way it was mapped during training, so you'll be required to load the JSON metadata file in the predict command. The way this works is: the first time you provide a UTF-8 encoded CSV (--data_csv), the HDF5 and JSON files are created, from the second time on Ludwig will load them instead of the CSV even if you specify the CSV (it looks in the same directory for files names in the same way but with a different extension), finally you can directly specify the HDF5 and JSON files (--data_hdf5 and --metadata_json).

As the mapping from raw data to tensors depends on the type of feature that you specify in your model definition, if you change type (for instance from sequential to text) you also have to redo the preprocessing, which is achieved by deleting the HDF5 and JSON files. Alternatively you can skip saving the HDF5 and JSON files specifying --skip_save_processed_input.

Splitting between train, validation and test set can be done in several ways. This allows for a few possible input data scenarios:

  • one single UTF-8 encoded CSV file is provided (-data_csv). In this case if the CSV contains a split column with values 0 for training, 1 for validation and 2 for test, this split will be used. If you want to ignore the split column and perform a random split, use a force_split argument in the model definition. In the case when there is no split column, a random 70-20-10 split will be performed. You can set the percentages and specify if you want stratified sampling in the model definition preprocessing section.

  • you can provide separate UTF-8 encoded train, validation and test CSVs (--data_train_csv, --data_validation_csv, --data_test_csv).

  • the HDF5 and JSON file indications specified in the case of a single CSV file apply also in the multiple files case (--data_train_hdf5, --data_validation_hdf5, --data_test_hdf5), with the only difference that you need to specify only one JSON file (--metadata_json) instead of three. The validation set is optional, but if absent the training wil continue until the end of the training epochs, while when there's a validation set the default behavior is to perform early stopping after the validation measure does not improve for a a certain amount of epochs. The test set is optional too.

Other optional arguments are --output_directory, --experiment_name and --model name. By default the output directory is ./results. That directory will contain a directory named [experiment_name]_[model_name]_0 if model name and experiment name are specified. If the same combination of experiment and model name is used again, the integer at the end of the name wil be increased. If neither of them is specified the directory will be named run_0. The directory will contain

  • description.json - a file containing a description of the training process with all the information to reproduce it.
  • training_statistics.json which contains records of all measures and losses for each epoch.
  • model - a directory containing model hyperparameters, weights, checkpoints and logs (for TensorBoard).

The model definition can be provided either as a string (--model_definition) or as YAML file (--model_definition_file). Details on how to write your model definition are provided in the Model Definition section.

During training Ludwig saves two sets of weights for the model, one that is the weights at the end of the epoch where the best performance on the validation measure was achieved and one that is the weights at the end of the latest epoch. The reason for keeping the second set is to be able to resume training in case the training process gets interrupted somehow.

To resume training using the latest weights and the whole history of progress so far you have to specify the --model_resume_path argument. You can avoid saving the latest weights and the overall progress so far by using the argument --skip_save_progress, but you will not be able to resume it afterwards. Another available option is to load a previously trained model as an initialization for a new training process. In this case Ludwig will start a new training process, without knowing any progress of the previous model, no training statistics, nor the number of epochs the model has been trained on so far. It's not resuming training, just initializing training with a previously trained model with the same model definition, and it is accomplished through the --model_load_path argument.

You can specify a random sed to be used by the python environment, python random package, numpy and TensorFlow with the --random_seed argument. This is useful for reproducibility. Be aware that due to asynchronicity in the TensorFlow GPU execution, when training on GPU results may not be reproducible.

You can manage which GPUs on your machine are used with the --gpus argument, which accepts a string identical to the format of CUDA_VISIBLE_DEVICES environment variable, namely a list of integers separated by comma. You can also specify the fraction of the GPU memory that will be initially assigned to TensorFlow with --gpu_fraction. By default it is 1.0, but you can set it, for instance, to 0.2 to use only 1/5 of the available memory. If TensorFlow will need more GPU memory it will try to increase this amount.

Finally the --logging_level argument lets you set the amount of logging that you want to see during training and the --debug argument turns on TensorFlow's tfdbg. Be careful when doing so, as it will help in catching errors, in particular infs and NaNs but it will consume much more memory.

Example:

ludwig train --data_csv reuters-allcats.csv --model_definition "{input_features: [{name: text, type: text, encoder: parallel_cnn, level: word}], output_features: [{name: class, type: category}]}"

predict

This command lets you use a previously trained model to predict on new data. You can call it with:

ludwig predict [options]

or with

python -m ludwig.predict [options]

from within Ludwig's main directory.

These are the available arguments:

usage: ludwig predict [options]

This script loads a pretrained model and uses it to predict.

optional arguments:
  -h, --help            show this help message and exit
  --data_csv DATA_CSV   input data CSV file. If it has a split column, it will
                        be used for splitting (0: train, 1: validation, 2:
                        test), otherwise the dataset will be randomly split
  --data_hdf5 DATA_HDF5
                        input data HDF5 file. It is an intermediate preprocess
                        version of the input CSV created the first time a CSV
                        file is used in the same directory with the same name
                        and a hdf5 extension
  --train_set_metadata_json TRAIN_SET_METADATA_JSON
                        input metadata JSON file. It is an intermediate
                        preprocess file containing the mappings of the input
                        CSV created the first time a CSV file is used in the
                        same directory with the same name and a json extension
  -s {training,validation,test,full}, --split {training,validation,test,full}
                        the split to test the model on
  -m MODEL_PATH, --model_path MODEL_PATH
                        model to load
  -od OUTPUT_DIRECTORY, --output_directory OUTPUT_DIRECTORY
                        directory that contains the results
  -ssuo, --skip_save_unprocessed_output
                        skips saving intermediate NPY output files
  -bs BATCH_SIZE, --batch_size BATCH_SIZE
                        size of batches
  -g GPUS, --gpus GPUS  list of gpu to use
  -gf GPU_FRACTION, --gpu_fraction GPU_FRACTION
                        fraction of gpu memory to initialize the process with
  -uh, --use_horovod    uses horovod for distributed training
  -dbg, --debug         enables debugging mode
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

The same distinction between UTF-8 encoded CSV files and HDF5 / JSON files explained in the train section also applies here. In either case, the JSON metadata file obtained during training is needed in order to map the new data into tensors. If the new data contains a split column, you can specify which split to use to calculate the predictions with the --split argument. By default it's full which means all the splits will be used.

A model to load is needed, and you can specify its path with the --model_path argument. If you trained a model previously and got the results in, for instance, ./results/experiment_run_0, you have to specify ./results/experiment_run_0/model for using it to predict.

You can specify an output directory with the argument --output-directory, by default it will be ./result_0, with increasing numbers if a directory with the same name is present.

The directory will contain a prediction CSV file and a probability CSV file for each output feature, together with raw NPY files containing raw tensors. You can specify not to save the raw NPY output files with the argument skip_save_unprocessed_output. If the argument --evaluate_performance if provided, a predict_statistics.json file containing all prediction statistics will also be outputted. If this parameter is specified, the data must contain columns for each output feature with ground truth output values in order to compute the performance statistics. If you receive an error regarding a missing output feature column in your data, it means that the data does not contain the columns for each output feature to use as ground truth.

A specific batch size for speeding up the prediction can be specified using the argument --batch_size.

Finally the --logging_level, --debug and --gpus related arguments behave exactly like described in the train command section.

Example:

ludwig predict --data_csv reuters-allcats.csv --model_path results/experiment_run_0/model/

test

This command lets you use a previously trained model to predict on new data and evaluate the performance of the prediction compared to ground truth. You can call it with:

ludwig test [options]

or with

python -m ludwig.test_performance [options]

from within Ludwig's main directory.

These are the available arguments:

usage: ludwig predict [options]

This script loads a pretrained model and uses it to predict.

optional arguments:
  -h, --help            show this help message and exit
  --data_csv DATA_CSV   input data CSV file. If it has a split column, it will
                        be used for splitting (0: train, 1: validation, 2:
                        test), otherwise the dataset will be randomly split
  --data_hdf5 DATA_HDF5
                        input data HDF5 file. It is an intermediate preprocess
                        version of the input CSV created the first time a CSV
                        file is used in the same directory with the same name
                        and a hdf5 extension
  --train_set_metadata_json TRAIN_SET_METADATA_JSON
                        input metadata JSON file. It is an intermediate
                        preprocess file containing the mappings of the input
                        CSV created the first time a CSV file is used in the
                        same directory with the same name and a json extension
  -s {training,validation,test,full}, --split {training,validation,test,full}
                        the split to test the model on
  -m MODEL_PATH, --model_path MODEL_PATH
                        model to load
  -od OUTPUT_DIRECTORY, --output_directory OUTPUT_DIRECTORY
                        directory that contains the results
  -ssuo, --skip_save_unprocessed_output
                        skips saving intermediate NPY output files
  -bs BATCH_SIZE, --batch_size BATCH_SIZE
                        size of batches
  -g GPUS, --gpus GPUS  list of gpu to use
  -gf GPU_FRACTION, --gpu_fraction GPU_FRACTION
                        fraction of gpu memory to initialize the process with
  -uh, --use_horovod    uses horovod for distributed training
  -dbg, --debug         enables debugging mode
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

All parameters are the same of predict and the behavior is the same. The only difference isthat test requires the dataset to contain also columns with the same name of output features. This is needed because test compares the predictions produced by the model with the ground truth and will save all those statistics in a test_statistics.json file in the result directory.

Example:

ludwig test --data_csv reuters-allcats.csv --model_path results/experiment_run_0/model/

experiment

This command combines training and test into a single handy command. You can call it with:

ludwig experiment [options]

or with

python -m ludwig.experiment [options]

from within Ludwig's main directory.

These are the available arguments:

usage: ludwig experiment [options]

This script trains and tests a model.

optional arguments:
  -h, --help            show this help message and exit
  --output_directory OUTPUT_DIRECTORY
                        directory that contains the results
  --experiment_name EXPERIMENT_NAME
                        experiment name
  --model_name MODEL_NAME
                        name for the model
  --data_csv DATA_CSV   input data CSV file. If it has a split column, it will
                        be used for splitting (0: train, 1: validation, 2:
                        test), otherwise the dataset will be randomly split
  --data_train_csv DATA_TRAIN_CSV
                        input train data CSV file
  --data_validation_csv DATA_VALIDATION_CSV
                        input validation data CSV file
  --data_test_csv DATA_TEST_CSV
                        input test data CSV file
  --data_hdf5 DATA_HDF5
                        input data HDF5 file. It is an intermediate preprocess
                        version of the input CSV created the first time a CSV
                        file is used in the same directory with the same name
                        and a hdf5 extension
  --data_train_hdf5 DATA_TRAIN_HDF5
                        input train data HDF5 file. It is an intermediate
                        preprocess version of the input CSV created the first
                        time a CSV file is used in the same directory with the
                        same name and a hdf5 extension
  --data_validation_hdf5 DATA_VALIDATION_HDF5
                        input validation data HDF5 file. It is an intermediate
                        preprocess version of the input CSV created the first
                        time a CSV file is used in the same directory with the
                        same name and a hdf5 extension
  --data_test_hdf5 DATA_TEST_HDF5
                        input test data HDF5 file. It is an intermediate
                        preprocess version of the input CSV created the first
                        time a CSV file is used in the same directory with the
                        same name and a hdf5 extension
  --train_set_metadata_json TRAIN_SET_METADATA_JSON
                        input train set metadata JSON file. It is an intermediate
                        preprocess file containing the mappings of the input
                        CSV created the first time a CSV file is used in the
                        same directory with the same name and a json extension
  -sspi, --skip_save_processed_input
                        skips saving intermediate HDF5 and JSON files
  -ssuo, --skip_save_unprocessed_output
                        skips saving intermediate NPY output files
  -md MODEL_DEFINITION, --model_definition MODEL_DEFINITION
                        model definition
  -mdf MODEL_DEFINITION_FILE, --model_definition_file MODEL_DEFINITION_FILE
                        YAML file describing the model. Ignores
                        --model_hyperparameters
  -mlp MODEL_LOAD_PATH, --model_load_path MODEL_LOAD_PATH
                        path of a pretrained model to load as initialization
  -mrp MODEL_RESUME_PATH, --model_resume_path MODEL_RESUME_PATH
                        path of a the model directory to resume training of
  -ssp SKIP_SAVE_PROGRESS_WEIGHTS, --skip_save_progress SKIP_SAVE_PROGRESS_WEIGHTS
                        disables saving weights after each epoch. By default
                        Ludwig saves weights after each epoch for enabling
                        resuming of training, but if the model is really big
                        that can be time consuming and will use twice as much
                        storage space, use this parameter to skip it.
  -rs RANDOM_SEED, --random_seed RANDOM_SEED
                        a random seed that is going to be used anywhere there
                        is a call to a random number generator: data
                        splitting, parameter initialization and training set
                        shuffling
  -g GPUS [GPUS ...], --gpus GPUS [GPUS ...]
                        list of gpus to use
  -gf GPU_FRACTION, --gpu_fraction GPU_FRACTION
                        fraction of gpu memory to initialize the process with
  -dbg, --debug         enables debugging mode
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

The parameters combine parameters from both train and test so refer to those sections for an in depth explanation. The output directory will contain the outputs both commands produce.

Example:

ludwig experiment --data_csv reuters-allcats.csv --model_definition "{input_features: [{name: text, type: text, encoder: parallel_cnn, level: word}], output_features: [{name: class, type: category}]}"

visualize

This command lets you visualize training and prediction statistics, alongside with comparing different models performances and predictions. You can call it with:

ludwig visualize [options]

or with

python -m ludwig.visualize [options]

from within Ludwig's main directory.

These are the available arguments:

usage: ludwig visualize [options]

This script analyzes results and shows some nice plots.

optional arguments:
  -h, --help            show this help message and exit
  -d DATA_CSV, --data_csv DATA_CSV
                        raw data file
  -g GROUND_TRUTH, --ground_truth GROUND_TRUTH
                        ground truth file
  -gm GROUND_TRUTH_METADATA, --ground_truth_metadata GROUND_TRUTH_METADATA
                        input metadata JSON file
  -v {compare_performance,compare_classifiers_performance_from_prob,compare_classifiers_performance_from_pred,compare_classifiers_performance_changing_k,compare_classifiers_performance_subset,compare_classifiers_predictions,compare_classifiers_predictions_distribution,confidence_thresholding,confidence_thresholding_2thresholds_3d,confidence_thresholding_data_vs_acc,confidence_thresholding_2thresholds_2d,confidence_thresholding_data_vs_acc_subset,confidence_thresholding_data_vs_acc_subset_per_class,binary_threshold_vs_metric,roc_curves,roc_curves_from_test_statistics,data_vs_acc_subset,data_vs_acc_subset_per_class,calibration_1_vs_all,calibration_multiclass,confusion_matrix,compare_classifiers_multiclass_multimetric,frequency_vs_f1,learning_curves}, --visualization {compare_performance,compare_classifiers_performance_from_prob,compare_classifiers_performance_from_pred,compare_classifiers_performance_changing_k,compare_classifiers_performance_subset,compare_classifiers_predictions,compare_classifiers_predictions_distribution,confidence_thresholding,confidence_thresholding_2thresholds_3d,confidence_thresholding_data_vs_acc,confidence_thresholding_2thresholds_2d,confidence_thresholding_data_vs_acc_subset,confidence_thresholding_data_vs_acc_subset_per_class,binary_threshold_vs_metric,roc_curves,roc_curves_from_test_statistics,data_vs_acc_subset,data_vs_acc_subset_per_class,calibration_1_vs_all,calibration_multiclass,confusion_matrix,compare_classifiers_multiclass_multimetric,frequency_vs_f1,learning_curves}
                        type of visualization
  -f FIELD, --field FIELD
                        field containing ground truth
  -tf THRESHOLD_FIELDS [THRESHOLD_FIELDS ...], --threshold_fields THRESHOLD_FIELDS [THRESHOLD_FIELDS ...]
                        fields for 2d threshold
  -pred PREDICTIONS [PREDICTIONS ...], --predictions PREDICTIONS [PREDICTIONS ...]
                        predictions files
  -prob PROBABILITIES [PROBABILITIES ...], --probabilities PROBABILITIES [PROBABILITIES ...]
                        probabilities files
  -tes TRAINING_STATS [TRAINING_STATS ...], --training_statistics TRAINING_STATS [TRAINING_STATS ...]
                        training stats files
  -trs TEST_STATS [TEST_STATS ...], --test_statistics TEST_STATS [TEST_STATS ...]
                        test stats files
  -alg ALGORITHMS [ALGORITHMS ...], --algorithms ALGORITHMS [ALGORITHMS ...]
                        names of the algorithms (for better graphs)
  -tn TOP_N_CLASSES [TOP_N_CLASSES ...], --top_n_classes TOP_N_CLASSES [TOP_N_CLASSES ...]
                        number of classes to plot
  -k TOP_K, --top_k TOP_K
                        number of elements in the ranklist to consider
  -ll LABELS_LIMIT, --labels_limit LABELS_LIMIT
                        maximum numbers of labels. If labels in dataset are
                        higher than this number, "rare" label
  -ss {ground_truth,predictions}, --subset {ground_truth,predictions}
                        type of subset filtering
  -n, --normalize       normalize rows in confusion matrix
  -m METRICS [METRICS ...], --metrics METRICS [METRICS ...]
                        metrics to dispay in threshold_vs_metric
  -pl POSITIVE_LABEL, --positive_label POSITIVE_LABEL
                        label of the positive class for the roc curve
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

As the --visualization parameters suggests, there is a vast number of visualizations readily available. Each of them requires a different subset of this command's arguments, so they will be described one by one in the Visualizations section.

collect_weights

This command lets you load a pre-trained model and collect the tensors with a specific name in order to save them in a NPY format. This may be useful in order to visualize the learned weights (for instance collecting embedding matrices) and for some post-hoc analyses. You can call it with:

ludwig collect_weights [options]

or with

python -m ludwig.collect weights [options]

from within Ludwig's main directory.

These are the available arguments:

usage: ludwig collect_weights [options]

This script loads a pretrained model and uses it collect weights.

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL_PATH, --model_path MODEL_PATH
                        model to load
  -t TENSORS [TENSORS ...], --tensors TENSORS [TENSORS ...]
                        tensors to collect
  -od OUTPUT_DIRECTORY, --output_directory OUTPUT_DIRECTORY
                        directory that contains the results
  -dbg, --debug         enables debugging mode
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

The three most important arguments are --model_path where you have to specify the path of the model to load, --tensors that lets you specify a list of tensor names in the TensorFlow graph that contain the weights you want to collect, and finally --output_directory that lets you specify where the NPY files (one for each tensor name specified) will be saved.

In order to figure out the names fo the tensors containing the weights you want to collect, the best way is to inspect the graph of the model with TensorBoard.

tensorboard --logdir /path/to/model/log

collect_activations

This command lets you load a pre-trained model and input data and collects the values of activations contained in tensors with a specific name in order to save them in a NPY format. This may be useful in order to visualize the activations (for instance collecting last layer's activations as embeddings representations of the input datapoint) and for some post-hoc analyses. You can call it with:

ludwig collect_activations [options]

or with

python -m ludwig.collect activations [options]

from within Ludwig's main directory.

These are the available arguments:

usage: ludwig collect_activations [options]

This script loads a pretrained model and uses it collect tensors for each
datapoint in the dataset.

optional arguments:
  -h, --help            show this help message and exit
  --data_csv DATA_CSV   input data CSV file
  --data_hdf5 DATA_HDF5
                        input data HDF5 file
  -s {training,validation,test,full}, --split {training,validation,test,full}
                        the split to test the model on
  -m MODEL_PATH, --model_path MODEL_PATH
                        model to load
  -t TENSORS [TENSORS ..], --tensors TENSORS [TENSORS ..]
                        tensors to collect
  -od OUTPUT_DIRECTORY, --output_directory OUTPUT_DIRECTORY
                        directory that contains the results
  -bs BATCH_SIZE, --batch_size BATCH_SIZE
                        size of batches
  -g GPUS, --gpus GPUS  list of gpu to use
  -gf GPU_FRACTION, --gpu_fraction GPU_FRACTION
                        fraction of gpu memory to initialize the process with
  -dbg, --debug         enables debugging mode
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

The data related and runtime related arguments (GPUs, batch size, etc.) are the same used in predict, you can refer to that section for an explanation. The collect specific arguments --model_path, --tensors and --output_directory are the same used in collect_weights, you can refer to that section for an explanation.

In order to figure out the names fo the tensors containing the activations you want to collect, the best way is to inspect the graph of the model with TensorBoard.

tensorboard --logdir /path/to/model/log

Data Preprocessing

Ludwig data preprocessing maps raw data coming in UTF-8 encoded CSV format into an HDF5 file containing tensors and a JSON file containing mappings from strings to tensors when needed. This mapping is performed when a UTF-8 encoded CSV is provided as input and both HDF5 and JSON files are saved in the same directory as the input CSV, unless the argument --skip_save_processed_input is used (both in train and experiment commands). The reason to save those files is both to provide a cache and avoid performing the preprocessing again (as, depending on the type of features involved, it could be time consuming) and to provide the needed mappings to be able to map unseen data into tensors.

The preprocessing process is personalizable to fit the specifics of your data format, but the basic assumption is always that your UTF-8 encoded CSV files contains one row for each datapoint and one column for each feature (either input or output), and that you are able to determine the type of that column among the ones supported by Ludwig. The reason for that is that each data type is mapped into tensors in a different way and expects the content to be formatted in a specific way. Different datatypes may have different formatters that format the values of a cell.

For instance the value of a cell of a sequence feature column by default is managed by a space formatter, that splits the content of the value into a list of strings using space.

before formatter after formatter
"token3 token4 token2" [token3, token4, token2]
"token3 token1" [token3, token1]

Then a list idx2str and two dictionaries str2idx and str2freq are created containing all the tokens in all the lists obtained by splitting all the rows of the column and an integer id is assigned to each of them (in order of frequency).

{
    "column_name": {
        "idx2str": [
            "<PAD>",
            "<UNK>",
            "token3",
            "token2",
            "token4",
            "token1"
        ],
        "str2idx": {
            "<PAD>": 0,
            "<UNK>": 1,
            "token3": 2,
            "token2": 3,
            "token4": 4,
            "token1": 5
        },
        "str2freq": {
            "<PAD>":  0,
            "<UNK>":  0,
            "token3": 2,
            "token2": 1,
            "token4": 1,
            "token1": 1
        }
    }
}

Finally a numpy matrix is created with sizes n x l where n is the number of rows in the column and l is the minimum of the longest tokenized list and a max_length parameter that can be set. All sequences shorter than l are padded on the right (but this behavior may also be modified through a parameter).

after formatter numpy matrix
[token3, token4, token2] 2 4 3
[token3, token1] 2 5 0

The final result matrix is saved in the HDF5 the name of the original column in the CSV as key, while the mapping from token to integer ID (and its inverse mapping) is saved in the JSON file.

Each datatype is preprocessed in a different way, using different parameters and different formatters. Details on how to set those parameters for each feature type and for each specific feature will be described in the Model Definition - Preprocessing section.

Binary features are directly transformed into a binary valued vector of length n (where n is the size of the dataset) and added to HDF5 with a key that reflects the name of column in the CSV. No additional information about them is available in the JSON metadata file.

Numerical features are directly transformed into a float valued vector of length n (where n is the size of the dataset) and added to HDF5 with a key that reflects the name of column in the CSV. No additional information about them is available in the JSON metadata file.

Category features are transformed into an integer valued vector of size n (where n is the size of the dataset) and added to HDF5 with a key that reflects the name of column in the CSV. The way categories are mapped into integers consists in first collecting a dictionary of all the different category strings present in the column of the CSV, then rank them by frequency and then assign them an increasing integer ID from the most frequent to the most rare (with 0 being assigned to a <UNK> token). The column name is added to the JSON file, with an associated dictionary containing 1. the mapping from integer to string (idx2str) 2. the mapping from string to id (str2idx) 3. the mapping from string to frequency (str2freq) 4. the size of the set of all tokens (vocab_size) 4. additional preprocessing information (by default how to fill missing values and what token to use to fill missing values)

Set features are transformed into a binary (int8 actually) valued matrix of size n x l (where n is the size of the dataset and l is the minimum of the size of the biggest set and a max_size parameter) and added to HDF5 with a key that reflects the name of column in the CSV. The way sets are mapped into integers consists in first using a formatter to map from strings to sequences of set items (by default this is done by splitting on spaces). Then a a dictionary of all the different set item strings present in the column of the CSV is collected, then they are ranked by frequency and an increasing integer ID is assigned to them from the most frequent to the most rare (with 0 being assigned to <PAD> used for padding and 1 assigned to <UNK> item). The column name is added to the JSON file, with an associated dictionary containing 1. the mapping from integer to string (idx2str) 2. the mapping from string to id (str2idx) 3. the mapping from string to frequency (str2freq) 4. the maximum size of all sets (max_set_size) 5. additional preprocessing information (by default how to fill missing values and what token to use to fill missing values)

Bag features are treated in the same way of set features, with the only difference being that the matrix had float values (frequencies).

Sequence features are transformed into an integer valued matrix of size n x l (where n is the size of the dataset and l is the minimum of the length of the longest sequence and a sequence_length_limit parameter) and added to HDF5 with a key that reflects the name of column in the CSV. The way sets are mapped into integers consists in first using a formatter to map from strings to sequences of tokens (by default this is done by splitting on spaces). Then a a dictionary of all the different token strings present in the column of the CSV is collected, then they are ranked by frequency and an increasing integer ID is assigned to them from the most frequent to the most rare (with 0 being assigned to <PAD> used for padding and 1 assigned to <UNK> item). The column name is added to the JSON file, with an associated dictionary containing 1. the mapping from integer to string (idx2str) 2. the mapping from string to id (str2idx) 3. the mapping from string to frequency (str2freq) 4. the maximum length of all sequences (sequence_length_limit) 5. additional preprocessing information (by default how to fill missing values and what token to use to fill missing values)

Text features are treated in the same way of sequence features, with a couple differences. Two different formatting/splitting happen, one that splits at every character and one that uses a spaCy based tokenizer (and removes stopwords) are used, and two different key are added to the HDF5 file, one containing the matrix of characters and one containing the matrix of words. The same thing happens in the JSON file, where there are dictionaries for mapping characters to integers (and the inverse) and words to integers (and their inverse). In the model definition you are able to specify which level of representation to use, if the character level or the word level.

Timeseries features are treated in the same way of sequence features, with the only difference being that the matrix in the HDF5 file does not have integer values, but float values. Moreover, there is no need for any mapping in the JSON file.

Image features are transformed into a int8 valued tensor of size n x h x w x c (where n is the size of the dataset and h x w is a specific resizing of the image that can be set, and c is the number of color channels) and added to HDF5 with a key that reflects the name of column in the CSV. The column name is added to the JSON file, with an associated dictionary containing preprocessing information about the sizes of the resizing.

CSV Format

Ludwig uses Pandas under the hood to read the UTF-8 encoded CSV files. Pandas tries to automatically identify the separator (generally ',') from the data. The default escape character is '\'. For example, if ',' is the column separator and one of your data columns has a ',' in it, Pandas would fail to load the data properly. To handle such cases, we expect the values in the columns to be escaped with backslashes (replace ',' in the data with '\\,').

Data Postprocessing

The JSON file obtained from preprocessing is used also for postprocessing: Ludwig models return output predictions and, depending on their datatype they are mapped back into the original space. Numerical and timeseries are returned as they are, while category, set, sequence, and text features output integers, those integers are mapped back into the original tokens / names using the idx2str in the JSON file. When you run experiment or predict you will find both a CSV file for each output containing the mapped predictions, a probability CSV file containing the probability of that prediction, a probabilities CSV file containing the probabilities for all alternatives (for instance, the probabilities of all the categories in case of a categorical feature). You will also find the unmapped NPY files. If you don't need them you can use the --skip_save_unprocessed_output argument.

Model Definition

The model definition is the core of Ludwig. It is a dictionary that contains all the information needed to build and train a Ludwig model. It mixes ease of use, by means of reasonable defaults, with flexibility, by means of detailed control over the parameters of your model. It is provided to both experiment and train commands either as a string (--model_definition) or as a file (--model_definition_file). The string or the content of the file will be parsed by PyYAML into a dictionary in memory, so any style of YAML accepted by the parser is considered to be valid, so both multiline and oneline formats are accepted. For instance a list of dictionaries can be written both as

mylist: [{name: item1, score: 2}, {name: item2, score: 1}, {name: item3, score: 4}]

or as:

mylist:
    -
        name: item1
        score: 2
    -
        name: item2
        score: 1
    -
        name: item3
        score: 4

The structure of the model definition file is a dictionary with five keys:

input_features: []
combiner: {}
output_features: []
training: {}
preprocessing: {}

Only input_features and output_features are required, the other three fields have default values, but you are free to modify them.

Input features

The input_features list contains a list of dictionaries, each of them containing two required fields name and type. name is the name of the feature and is the same name of the column of the CSV input file, type is one of the supported datatypes. Input features may have different ways to be encoded and the parameter to decide it is encoder.

All the other parameters you specify in an input feature will be passed as parameters to the function that build the encoder, and each encoder can have different parameters.

For instance a sequence feature can be encoded by a stacked_cnn or by and rnn, but only the stacked_cnn will accept the parameter num_filters while only the rnn will accept the parameter bidirectional.

A list of all the encoders available for all the datatypes alongside with the description of all parameters will be provided in the datatype-specific sections. Some datatypes have only one type of encoder, so you are not required to specify it.

The role of the encoders is to map inputs into tensors, usually vectors in the case of datatype without a temporal / sequential aspect, matrices in case there is a temporal / sequential aspect or higher rank tensors in case there is a spatial or a spatio-temporal aspect to the input data.

Different configurations of the same encoder may return a tensor with different rank, for instance a sequential encoder may return a vector of size h that is either the final vector of a sequence or the result of pooling over the sequence length, or it can return a matrix of size l x h where l is the length of the sequence and h is the hidden dimension if you specify the pooling reduce operation (reduce_output) to be null. For the sake of simplicity you can imagine the output to be a vector in most of the cases, but there is a reduce_output parameter one can specify to change the default behavior.

An additional feature that ludwig provides is the option to have tied weights between different encoders. For instance if my model takes two sentences as input and return the probability of their entailment, I may want to encode both sentences with the same encoder. The way to do it is by specifying the tied-weights parameter of the second feature you define to be the name of the first feature you defined.

input_features:
    -
        name: sentence1
        type: text
    -
        name: sentence2
        type: text
        tied_weights: sentence1

If you specify a name of an input feature that has not been defined yet, it will result in an error. Also, in order to be able to have tied weights, all encoder parameters have to be identical between the two input features.

Combiner

Combiners are part of the model that take all the outputs of the different input features and combine them in a single representation that is passed to the outputs. You can specify which one to use in the combiner section of the model definition. Different combiners implement different combination logic, but the default one concat just concatenates all outputs of input feature encoders and optionally passes the concatenation through fully connected layers, with the output of the last layer being forwarded to the outputs decoders.

+-----------+
|Input      |
|Feature 1  +-+
+-----------+ |            +---------+
+-----------+ | +------+   |Fully    |
|...        +--->Concat+--->Connected+->
+-----------+ | +------+   |Layers   |
+-----------+ |            +---------+
|Input      +-+
|Feature N  |
+-----------+

For the sake of simplicity you can imagine the both inputs and outputs are vectors in most of the cases, but there are reduce_input and reduce_output parameters to specify to change the default behavior.

Output Features

The output_features list has the same structure of the input_features list: it is a list of dictionaries containing a name and a type. They represent outputs / targets that you want your model to predict. In most machine learning tasks you want to predict only one target variable, but in Ludwig you are allowed to specify as many outputs as you want and they are going to be optimized in a multi-task fashion, using a weighted sum of their losses as a combined loss to optimize.

Instead of having encoders, output features have decoders, but most of them have only one decoder so you don't have to specify it.

Decoders take the output of the combiner as input, process it further, for instance passing it through fully connected layers, and finally predict values and compute a loss and some measures (depending on the datatype different losses and measures apply).

Decoders have additional parameters, in particular loss that allows you to specify a different loss to optimize for this specific decoder, for instance numerical features support both mean_squared_error and mean_absolute_error as losses. Details about the available decoders and losses alongside with the description of all parameters will be provided in the datatype-specific sections.

For the sake of simplicity you can imagine the input coming from the combiner to be a vector in most of the cases, but there is a reduce_input parameter one can specify to change the default behavior.

Output Features Dependencies

An additional feature that Ludwig provides is the concept of dependency between output_features. You can specify a list of output features as dependencies when you write the dictionary of a specific feature. At model building time Ludwig checks that no cyclic dependency exists. If you do so Ludwig will concatenate all the final representations before the prediction of those output features to the original input of the decoder. The reason is that if different output features have a causal dependency, knowing which prediction has been made for one can help making the prediction of the other.

For instance if two output features are one coarse grained category and one fine-grained category that are in a hierarchical structure with each other, knowing the prediction made for coarse grained restricts the possible categories to predict for the fine-grained. In this case the following model definition structure can be used:

output_features:
    -
        name: coarse_class
        type: category
        num_fc_layers: 2
        fc_size: 64
    -
        name: fine_class
        type: category
        dependencies:
            - coarse_class
        num_fc_layers: 1
        fc_size: 64

Assuming the input coming from the combiner has hidden dimension h 128, there are two fully connected layers that return a vector with hidden size 64 at the end of the coarse_class decoder (that vector will be used for the final layer before projecting in the output coarse_class space) In the decoder of fine_class, the 64 dimensional vector of coarse_class will be concatenated to the combiner output vector, making a vector of hidden size 192 that will be passed through a fully connected layer and the 64 dimensional output will be used for the final layer before projecting in the output class space of the fine_class.

Training

The training section of the model definition lets you specify some parameters of the training process, like for instance the number of epochs or the learning rate.

These are the available training parameters:

  • batch_size (default 128): size of the batch used for training the model.
  • eval_batch_size (default 0): size of the batch used for evaluating the model. If it is 0, the same value of batch_size is used. This is usefult to speedup evaluation with a much bigger batch size than training, if enough memory is available, or to decrease the batch size when sampled_softmax_cross_entropy is used as loss for sequential and categorical features with big vocabulary sizes (evaluation needs to be performed on the full vocabulary, so a much smaller batch size may be needed to fit the activation tensors in memory).
  • epochs (default 100): number of epochs the training process will run for.
  • early_stop (default 5): if there's a validation set, number of epochs of patience without an improvement on the validation measure before the training is stopped.
  • optimizer (default {type: adam, beta1: 0.9, beta2: 0.999, epsilon: 1e-08}): which optimizer to use with the relative parameters. The available optimizers are: sgd (or stochastic_gradient_descent, gd, gradient_descent, they are all the same), adam, adadelta, adagrad, adagradda, momentum, ftrl, proximalgd, proximaladagrad, rmsprop. To know their parameters check TensorFlow's optimizer documentation.
  • learning_rate (default 0.001): the learning rate to use.
  • decay (default false): if to use exponential decay of the learning rate or not.
  • decay_rate (default 0.96): the rate of the exponential learning rate decay.
  • decay_steps (default 10000): the number of steps of the exponential learning rate decay.
  • staircase (default false): decays the learning rate at discrete intervals.
  • regularization_lambda (default 0): the lambda parameter used for adding a l2 regularization loss to the overall loss.
  • dropout_rate (default 0.0): the probability to drop neurons in dropout. The dropout_rate is used throughout the whole model, but to decide which parts of the model will use it, use the dropout boolean parameter available in each encoder, combiner and decoder.
  • reduce_learning_rate_on_plateau (default 0): if there's a validation set, how many times to reduce the learning rate when a plateau of validation measure is reached.
  • reduce_learning_rate_on_plateau_patience (default 5): if there's a validation set, number of epochs of patience without an improvement on the validation measure before reducing the learning rate.
  • reduce_learning_rate_on_plateau_rate (default 0.5): if there's a validation set, the reduction rate of the learning rate.
  • increase_batch_size_on_plateau (default 0): if there's a validation set, how many times to increase the batch size when a plateau of validation measure is reached.
  • increase_batch_size_on_plateau_patience (default 5): if there's a validation set, number of epochs of patience without an improvement on the validation measure before increasing the learning rate.
  • increase_batch_size_on_plateau_rate (default 2): if there's a validation set, the increase rate of the batch size.
  • increase_batch_size_on_plateau_max (default 512): if there's a validation set, the maximum value of batch size.
  • validation_field (default combined): when there is more than one output feature, which one to use for computing if there was an improvement on validation. The measure to use to determine if there was an improvement can be set with the validation_measure parameter. Different datatypes have different available measures, refer to the datatype-specific section for more details. combined indicates the use the combination of all features. For instance the combination of combined and loss as measure uses a decrease in the combined loss of all output features to check for improvement on validation, while combined and accuracy considers on how many datapoints the predictions for all output features were correct (but consider that for some features, for instance numeric there is no accuracy measure, so you should use accuracy only if all your output features have an accuracy measure).
  • validation_measure: (default accuracy): the measure to use to determine if there was an improvement. The measure is considered for the output feature specified in validation_field. Different datatypes have different available measures, refer to the datatype-specific section for more details.
  • bucketing_field (default null): when not null, when creating batches, instead of shuffling randomly, the length along the last dimension of the matrix of the specified input feature is used for bucketing datapoints and then randomly shuffled datapoints from the same bin are sampled. Padding is trimmed to the longest datapoint in the batch. The specified feature should be either a sequence or text feature and the encoder encoding it has to be rnn. When used, bucketing improves speed of rnn encoding up to 1.5x, depending on the length distribution of the inputs.

Optimizers details

Preprocessing

The preprocessing section of the model definition makes it possible to specify datatype specific parameters to perform data preprocessing. The preprocessing dictionary contains one key of each datatype, but you have to specify only the ones that apply to your case, the other ones will be kept as defaults. Moreover, the preprocessing dictionary contains parameters related to how to split the data that are not feature specific.

  • force_split (default false): if true the split column in the CSV data file is ignored and the dataset is randomly split. If false the split column is used if available.
  • split_probabilities (default [0.7, 0.1, 0.2]): the proportion of the CSV data to end up in training, validation and test. The three values have to sum up to one.
  • stratify (default null): if null the split is random, otherwise you can specify the name of a category feature and the split will be stratified on that feature.

Example preprocessing dictionary (showing default values):

preprocessing:
    force_split: false
    split_probabilities: [0.7, 0.1, 0.2]
    stratify: null
    category: {...}
    sequence: {...}
    text: {...}
    ...

The details about the preprocessing parameters that each datatype accepts will be provided in the datatype-specific sections.

It is important to point out that different features within the same datatype may require different preprocessing. For instance a document classification model may have two text input features, one for the title of the document and one for the body.

As the length of the title is much shorter than the length of the body, the parameter word_length_limit should be set to 10 for the title and 2000 for the body, but both of them share the same parameter most_common_words with value 10000.

The way to do this is adding a preprocessing key inside the title input_feature dictionary and one in the body input feature dictionary containing the desired parameter and value. The model definition will look like:

preprocessing:
    text:
        most_common_word: 10000
input_features:
    -
        name: title
        type: text
        preprocessing:
            word_length_limit: 20
    -
        name: body
        type: text
        preprocessing:
            word_length_limit: 2000

Binary Features

Binary Features Preprocessing

Binary features are directly transformed into a binary valued vector of length n (where n is the size of the dataset) and added to HDF5 with a key that reflects the name of column in the CSV. No additional information about them is available in the JSON metadata file.

The parameters available for preprocessing are

  • missing_value_strategy (default fill_with_const): what strategy to follow when there's a missing value in a binary column. The value should be one of fill_with_const (replaces the missing value with a specific value specified with the fill_value parameter), fill_with_mode (replaces the missing values with the most frequent value in the column), fill_with_mean (replaces the missing values with the mean of the values in the column), backfill (replaces the missing values with the next valid value).
  • fill_value (default 0): the value to replace the missing values with in case the missing_value_strategy is fill-value.

Binary Input Features and Encoders

Binary features have no encoder, the raw binary values coming from the input placeholders are just returned as outputs. By consequence there are no encoding parameters. Inputs are of size b while outputs are of size b x 1 where b is the batch size.

Example binary feature entry in the output features list:

name: binary_csv_column_name
type: binary

Binary Output Features and Decoders

Binary features can be used when a binary classification needs to be performed or when the output is a single probability. There is only one decoder available for binary features and it is a (potentially empty) stack of fully connected layers, followed by a projection into a single number followed by a sigmoid function.

These are the available parameters of a binary output feature

  • reduce_inputs (default sum): defines how to reduce an input that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • dependencies (default []): the output features this one is dependent on. For a detailed explanation refer to Output Features Dependencies.
  • reduce_dependencies (default sum): defines how to reduce the output of a dependent feature that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • loss (default {type: cross_entropy, confidence_penalty: 0, robust_lambda: 0}): is a dictionary containing a loss type and its hyperparameters. The only available loss type is cross_entropy (cross entropy), and the two optional parameters are confidence_penalty (an additional term that penalizes too confident predictions by adding a a * (max_entropy - entropy) / max_entropy term to the loss, where a is the value of this parameter) and robust_lambda (replaces the loss with (1 - robust_lambda) * loss + robust_lambda / 2 which is useful in case of noisy labels).

These are the available parameters of a binary output feature decoder

  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, dropout, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the decoder will be used instead.
  • num_fc_layers (default 0): this is the number of stacked fully connected layers that the input to the feature passes through. Their output is projected in the feature's output space.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the wights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • threshold (default 0.5): The threshold above (greater or equal) which the predicted output of the sigmoid will be mapped to 1.

Example binary feature entry (with default parameters) in the output features list:

name: binary_csv_column_name
type: binary
reduce_inputs: sum
dependencies: []
reduce_dependencies: sum
loss:
    type: cross_entropy
    confidence_penalty: 0
    robust_lambda: 0
fc_layers: null
num_fc_layers: 0
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true
threshold: 0.5

Binary Features Measures

The only measures that are calculated every epoch and are available for binary features are the accuracy and the loss itself. You can set either of them as validation_measure in the training section of the model definition if you set the validation_field to be the name of a binary feature.

Numerical Features

Numerical Features Preprocessing

Numerical features are directly transformed into a float valued vector of length n (where n is the size of the dataset) and added to HDF5 with a key that reflects the name of column in the CSV. No additional information about them is available in the JSON metadata file.

Parameters available for preprocessing are

  • missing_value_strategy (default fill_with_const): what strategy to follow when there's a missing value in a binary column. The value should be one of fill_with_const (replaces the missing value with a specific value specified with the fill_value parameter), fill_with_mode (replaces the missing values with the most frequent value in the column), fill_with_mean (replaces the missing values with the mean of the values in the column), backfill (replaces the missing values with the next valid value).
  • fill_value (default 0): the value to replace the missing values with in case the missing_value_strategy is fill-value.

Numerical Input Features and Encoders

Numerical features have one encoder, the raw float values coming from the input placeholders are passed through a single neuron for scaling purposes, (optionally) passed through a normalization layer (either null, batch_norm, or layer_norm) and returned as outputs. Inputs are of size b while outputs are fo size b x 1 where b is the batch size.

The available encoder parameters are:

  • norm' (default null): norm to apply after the single neuron. It can be null, batch or layer.
  • tied_weights (default null): name of the input feature to tie the weights the encoder with. It needs to be the name of a feature of the same type and with the same encoder parameters.

Example numerical feature entry in the output features list:

name: numerical_csv_column_name
type: numerical
norm: null
tied_weights: null

Numerical Output Features and Decoders

Numerical features can be used when a regression needs to be performed. There is only one decoder available for numerical features and it is a (potentially empty) stack of fully connected layers, followed by a projection into a single number.

These are the available parameters of a numerical output feature

  • reduce_inputs (default sum): defines how to reduce an input that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • dependencies (default []): the output features this one is dependent on. For a detailed explanation refer to Output Features Dependencies.
  • reduce_dependencies (default sum): defines how to reduce the output of a dependent feature that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • loss (default {type: mean_squared_error}): is a dictionary containing a loss type. The available losses type are mean_squared_error and mean_absolute_error.

These are the available parameters of a numerical output feature decoder

  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, dropout, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the decoder will be used instead.
  • num_fc_layers (default 0): this is the number of stacked fully connected layers that the input to the feature passes through. Their output is projected in the feature's output space.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the weights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • clip (default null): If not null it specifies a minimum and maximum value the predictions will be clipped to. The value can be either a list or a tuple of length 2, with the first value representing the minimum and the second the maximum. For instance (-5,5) will make it so that all predictions will be clipped in the [-5,5] interval.

Example numerical feature entry (with default parameters) in the output features list:

name: numerical_csv_column_name
type: numerical
reduce_inputs: sum
dependencies: []
reduce_dependencies: sum
loss:
    type: mean_squared_error
fc_layers: null
num_fc_layers: 0
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true

Numerical Features Measures

The measures that are calculated every epoch and are available for numerical features are mean_squared_error, mean_absolute_error, r2 and the loss itself. You can set either of them as validation_measure in the training section of the model definition if you set the validation_field to be the name of a numerical feature.

Category Features

Category Features Preprocessing

Category features are transformed into an integer valued vector of size n (where n is the size of the dataset) and added to HDF5 with a key that reflects the name of column in the CSV. The way categories are mapped into integers consists in first collecting a dictionary of all the different category strings present in the column of the CSV, then rank them by frequency and then assign them an increasing integer ID from the most frequent to the most rare (with 0 being assigned to a <UNK> token). The column name is added to the JSON file, with an associated dictionary containing 1. the mapping from integer to string (idx2str) 2. the mapping from string to id (str2idx) 3. the mapping from string to frequency (str2freq) 4. the size of the set of all tokens (vocab_size) 4. additional preprocessing information (by default how to fill missing values and what token to use to fill missing values)

The parameters available for preprocessing are

  • missing_value_strategy (default fill_with_const): what strategy to follow when there's a missing value in a binary column. The value should be one of fill_with_const (replaces the missing value with a specific value specified with the fill_value parameter), fill_with_mode (replaces the missing values with the most frequent value in the column), fill_with_mean (replaces the missing values with the mean of the values in the column), backfill (replaces the missing values with the next valid value).
  • fill_value (default "<UNK>"): the value to replace the missing values with in case the missing_value_strategy is fill-value.
  • lowercase (default false): if the string has to be lowercased before being handled by the formatter.
  • most_common (default 10000): the maximum number of most common tokens to be considered. if the data contains more than this amount, the most infrequent tokens will be treated as unknown.

Category Input Features and Encoders

Category features have one encoder, the raw integer values coming from the input placeholders are mapped to either dense or sparse embeddings (one-hot encodings) and returned as outputs. Inputs are of size b while outputs are fo size b x h where b is the batch size and h is the dimensionality of the embeddings.

The available encoder parameters are

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 256): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • dropout (default false): determines if there should be a dropout layer after embedding.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the embedding weights are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • tied_weights (default null): name of the input feature to tie the weights the encoder with. It needs to be the name of a feature of the same type and with the same encoder parameters.

Example category feature entry in the input features list:

name: category_csv_column_name
type: category
representation: dense
embedding_size: 256
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
dropout: false
initializer: null
regularize: true
tied_weights: null

Category Output Features and Decoders

Category features can be used when a multi-class classification needs to be performed. There is only one decoder available for category features and it is a (potentially empty) stack of fully connected layers, followed by a projection into a vector of size of the number of available classes, followed by a softmax.

+--------------+   +---------+   +-----------+
|Combiner      |   |Fully    |   |Projection |   +-------+
|Output        +--->Connected+--->into Output+--->Softmax|
|Representation|   |Layers   |   |Space      |   +-------+
+--------------+   +---------+   +-----------+

These are the available parameters of a category output feature

  • reduce_inputs (default sum): defines how to reduce an input that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • dependencies (default []): the output features this one is dependent on. For a detailed explanation refer to Output Features Dependencies.
  • reduce_dependencies (default sum): defines how to reduce the output of a dependent feature that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • loss (default {type: softmax_cross_entropy, class_similarities_temperature: 0, class_weights: 1, confidence_penalty: 0, distortion: 1, labels_smoothing: 0, negative_samples: 0, robust_lambda: 0, sampler: null, unique: false}): is a dictionary containing a loss type. The available losses type are softmax_cross_entropy and sampled_softmax_cross_entropy.

These are the loss parameters

  • confidence_penalty (default 0): penalizes overconfident predictions (low entropy) by adding an additional term that penalizes too confident predictions by adding a a * (max_entropy - entropy) / max_entropy term to the loss, where a is the value of this parameter. Useful in case of noisy labels.
  • robust_lambda (default 0): replaces the loss with (1 - robust_lambda) * loss + robust_lambda / c where c is the number of classes, which is useful in case of noisy labels.
  • class_weights (default 1): the value can be a vector of weights, one of each class, that is multiplied to the loss of the datapoints that have that class as ground truth. It is an alternative to oversampling in case of unbalanced class distribution. The ordering of the vector follows the category to integer ID mapping in the JSON metadata file (the <UNK> class needs to be included too).
  • class_similarities (default null): if not null it is a c x c matrix in the form of a list of lists that contains the mutual similarity of classes. It is used if class_similarities_temperature is greater than 0. The ordering of the vector follows the category to integer ID mapping in the JSON metadata file (the <UNK> class needs to be included too).
  • class_similarities_temperature (default 0): is the temperature parameter of the softmax that is performed on each row of class_similarities. The output of that softmax is used to determine the supervision vector to provide instead of the one hot vector that would be provided otherwise for each datapoint. The intuition behind it is that errors between similar classes are more tollerable than errors between really different classes.
  • labels_smoothing (default 0): If label_smoothing is nonzero, smooth the labels towards 1/num_classes: new_onehot_labels = onehot_labels * (1 - label_smoothing) + label_smoothing / num_classes.
  • negative_samples (default 0): if type is sampled_softmax_cross_entropy, this parameter indicates how many negative samples to use.
  • sampler (default null): options are fixed_unigram, uniform, log_uniform, learned_unigram. For a detailed description of the samplers refer to TensorFlow's documentation.
  • distortion (default 1): when loss is sampled_softmax_cross_entropy and the sampler is either unigram or learned_unigram this is used to skew the unigram probability distribution. Each weight is first raised to the distortion's power before adding to the internal unigram distribution. As a result, distortion = 1.0 gives regular unigram sampling (as defined by the vocab file), and distortion = 0.0 gives a uniform distribution.
  • unique (default false): Determines whether all sampled classes in a batch are unique.

These are the available parameters of a category output feature decoder

  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, dropout, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the decoder will be used instead.
  • num_fc_layers (default 0): this is the number of stacked fully connected layers that the input to the feature passes through. Their output is projected in the feature's output space.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the weights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • top_k (default 3): determines the parameter k, the number of categories to consider when computing the top_k measure. It computes accuracy but considering as a match if the true category appears in the first k predicted categories ranked by decoder's confidence.

Example category feature entry (with default parameters) in the output features list:

name: category_csv_column_name
type: category
reduce_inputs: sum
dependencies: []
reduce_dependencies: sum
loss:
    type: softmax_cross_entropy
    confidence_penalty: 0
    robust_lambda: 0
    class_weights: 1
    class_similarities: null
    class_similarities_temperature: 0
    labels_smoothing: 0
    negative_samples: 0
    sampler: null
    distortion: 1
    unique: false
fc_layers: null
num_fc_layers: 0
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true
top_k: 3

Category Features Measures

The measures that are calculated every epoch and are available for category features are accuracy, top_k (computes accuracy considering as a match if the true category appears in the first k predicted categories ranked by decoder's confidence) and the loss itself. You can set either of them as validation_measure in the training section of the model definition if you set the validation_field to be the name of a category feature.

Set Features

Set Features Preprocessing

Set features are transformed into a binary (int8 actually) valued matrix of size n x l (where n is the size of the dataset and l is the minimum of the size of the biggest set and a max_size parameter) and added to HDF5 with a key that reflects the name of column in the CSV. The way sets are mapped into integers consists in first using a formatter to map from strings to sequences of set items (by default this is done by splitting on spaces). Then a a dictionary of all the different set item strings present in the column of the CSV is collected, then they are ranked by frequency and an increasing integer ID is assigned to them from the most frequent to the most rare (with 0 being assigned to <PAD> used for padding and 1 assigned to <UNK> item). The column name is added to the JSON file, with an associated dictionary containing 1. the mapping from integer to string (idx2str) 2. the mapping from string to id (str2idx) 3. the mapping from string to frequency (str2freq) 4. the maximum size of all sets (max_set_size) 5. additional preprocessing information (by default how to fill missing values and what token to use to fill missing values)

The parameters available for preprocessing arehe parameters available for preprocessing are

  • missing_value_strategy (default fill_with_const): what strategy to follow when there's a missing value in a binary column. The value should be one of fill_with_const (replaces the missing value with a specific value specified with the fill_value parameter), fill_with_mode (replaces the missing values with the most frequent value in the column), fill_with_mean (replaces the missing values with the mean of the values in the column), backfill (replaces the missing values with the next valid value).
  • fill_value (default 0): the value to replace the missing values with in case the missing_value_strategy is fill-value.
  • format (default space): defines how to map from the raw string content of the CSV column to a set of elements. The default value space splits the string on spaces. Other options are: underscore (splits on underscore), comma(splits on comma), json (decodes the string into a set or a list through a JSON parser).
  • lowercase (default false): if the string has to be lowercased before being handled by the formatter.
  • most_common (default 10000): the maximum number of most common tokens to be considered. if the data contains more than this amount, the most infrequent tokens will be treated as unknown.

Set Input Features and Encoders

Set features have one encoder, the raw binary values coming from the input placeholders are first transformed in sparse integer lists, then they are mapped to either dense or sparse embeddings (one-hot encodings), finally they are aggregated and returned as outputs. Inputs are of size b while outputs are fo size b x h where b is the batch size and h is the dimensionally of the embeddings.

+-+
|0|          +-----+
|0|   +-+    |emb 2|   +-----------+
|1|   |2|    +-----+   |Aggregation|
|0+--->4+---->emb 4+--->Reduce     +->
|1|   |5|    +-----+   |Operation  |
|1|   +-+    |emb 5|   +-----------+
|0|          +-----+
+-+

The available encoder parameters are

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 50): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • dropout (default false): determines if there should be a dropout layer before returning the encoder output.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the embedding weights are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • reduce_output (default sum): describes the strategy to use to aggregate the embeddings of the items of the set. Possible values are sum, mean and sqrt (the weighted sum divided by the square root of the sum of the squares of the weights).
  • tied_weights (default null): name of the input feature to tie the weights the encoder with. It needs to be the name of a feature of the same type and with the same encoder parameters.

Example set feature entry in the output features list:

name: set_csv_column_name
type: set
representation: dense
embedding_size: 50
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
dropout: false
initializer: null
regularize: true
reduce_output: sum
tied_weights: null

Set Output Features and Decoders

Set features can be used when multi-label classification needs to be performed. There is only one decoder available for set features and it is a (potentially empty) stack of fully connected layers, followed by a projection into a vector of size of the number of available classes, followed by a sigmoid.

+--------------+   +---------+   +-----------+
|Combiner      |   |Fully    |   |Projection |   +-------+
|Output        +--->Connected+--->into Output+--->Sigmoid|
|Representation|   |Layers   |   |Space      |   +-------+
+--------------+   +---------+   +-----------+

These are the available parameters of the set output feature

  • reduce_inputs (default sum): defines how to reduce an input that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • dependencies (default []): the output features this one is dependent on. For a detailed explanation refer to Output Features Dependencies.
  • reduce_dependencies (default sum): defines how to reduce the output of a dependent feature that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • loss (default {type: sigmoid_cross_entropy}): is a dictionary containing a loss type. The available loss type is sigmoid_cross_entropy.

These are the available parameters of a set output feature decoder

  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, dropout, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the decoder will be used instead.
  • num_fc_layers (default 0): this is the number of stacked fully connected layers that the input to the feature passes through. Their output is projected in the feature's output space.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the wights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • threshold (default 0.5): The threshold above (greater or equal) which the predicted output of the sigmoid will be mapped to 1.

Example set feature entry (with default parameters) in the output features list:

name: set_csv_column_name
type: set
reduce_inputs: sum
dependencies: []
reduce_dependencies: sum
loss:
    type: sigmoid_cross_entropy
fc_layers: null
num_fc_layers: 0
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true
threshold: 0.5

Set Features Measures

The measures that are calculated every epoch and are available for category features are jaccard_index and the loss itself. You can set either of them as validation_measure in the training section of the model definition if you set the validation_field to be the name of a set feature.

Bag Features

Bag Features Preprocessing

Bag features are treated in the same way of set features, with the only difference being that the matrix had float values (frequencies).

Bag Input Features and Encoders

Bag features have one encoder, the raw float values coming from the input placeholders are first transformed in sparse integer lists, then they are mapped to either dense or sparse embeddings (one-hot encodings), they are aggregated as a weighted sum, where the weights are the original float values, and finally returned as outputs. Inputs are of size b while outputs are fo size b x h where b is the batch size and h is the dimensionality of the embeddings.

The parameters are the same used for set input features with the exception of reduce_output that does not apply in this case because the weighted sum already acts as a reducer.

Bag Output Features and Decoders

There is no bag decoder available yet.

Bag Features Measures

As there is no decoder there is also no measure available yet for bag feature.

Sequence Features

Sequence Features Preprocessing

Sequence features are transformed into an integer valued matrix of size n x l (where n is the size of the dataset and l is the minimum of the length of the longest sequence and a sequence_length_limit parameter) and added to HDF5 with a key that reflects the name of column in the CSV. The way sequences are mapped into integers consists in first using a formatter to map from strings to sequences of tokens (by default this is done by splitting on spaces). Then a a dictionary of all the different token strings present in the column of the CSV is collected, then they are ranked by frequency and an increasing integer ID is assigned to them from the most frequent to the most rare (with 0 being assigned to <PAD> used for padding and 1 assigned to <UNK> item). The column name is added to the JSON file, with an associated dictionary containing 1. the mapping from integer to string (idx2str) 2. the mapping from string to id (str2idx) 3. the mapping from string to frequency (str2freq) 4. the maximum length of all sequences (sequence_length_limit) 5. additional preprocessing information (by default how to fill missing values and what token to use to fill missing values)

The parameters available for preprocessing are

  • missing_value_strategy (default fill_with_const): what strategy to follow when there's a missing value in a binary column. The value should be one of fill_with_const (replaces the missing value with a specific value specified with the fill_value parameter), fill_with_mode (replaces the missing values with the most frequent value in the column), fill_with_mean (replaces the missing values with the mean of the values in the column), backfill (replaces the missing values with the next valid value).
  • fill_value (default ""): the value to replace the missing values with in case the missing_value_strategy is fill_value.
  • padding (default right): the direction of the padding. right and left are available options.
  • padding_symbol (default <PAD>): the string used as a padding symbol. Is is mapped to the integer ID 0 in the vocabulary.
  • unknown_symbol (default <UNK>): the string used as a unknown symbol. Is is mapped to the integer ID 1 in the vocabulary.
  • lowercase (default false): if the string has to be lowercase before being handled by the formatter.
  • format (default space): defines how to map from the raw string content of the CSV column to a sequence of elements. The default value space splits the string on spaces. Other options are: underscore (splits on underscore), comma(splits on comma), json (decodes the string into a set or a list through a JSON parser).
  • most_common (default 20000): the maximum number of most common tokens to be considered. if the data contains more than this amount, the most infrequent tokens will be treated as unknown.
  • sequence_length_limit (default 256): the maximum length of the sequence. Sequences that are longer than this value will be truncated, while sequences that are shorter will be padded.

Sequence Input Features and Encoders

Sequence features have several encoders and each of them has its own parameters. Inputs are of size b while outputs are fo size b x h where b is the batch size and h is the dimensionally of the output of the encoder. In case a representation for each element of the sequence is needed (for example for tagging them, or for using an attention mechanism), one can specify the parameter reduce_output to be null or None and the output will be a b x s x h tensor where s is the length of the sequence. Some encoders, because of their inner workings, may require additional parameters to be specified in order to obtain one representation for each element of the sequence. For instance the parallel_cnn encoder, by default pools and flattens the sequence dimension and then passes the flattened vector through fully connected layers, so in order to obtain the full tesnor one has to specify reduce_output: null.

Sequence input feature parameters are

  • encoder (default parallel_cnn): the name of the encoder to use to encode the sequence. The available ones are embed, parallel_cnn, stacked_cnn, stacked_parallel_cnn, rnn, cnnrnn and passthrough (equivalent to specify None or null).
  • tied_weights (default null): name of the input feature to tie the weights the encoder with. It needs to be the name of a feature of the same type and with the same encoder parameters.
Embed Encoder

The embed decoder simply maps each integer in the sequence to an embedding, creating a b x s x h tensor where b is the batch size, s is the length of the sequence and h is the embedding size. The tensor is reduced along the s dimension to obtain a single vector of size h for each element of the batch. If you want to output the full b x s x h tensor, you can specify reduce_output: null.

       +------+
       |Emb 12|
       +------+
+--+   |Emb 7 |
|12|   +------+
|7 |   |Emb 43|   +-----------+
|43|   +------+   |Aggregation|
|65+--->Emb 65+--->Reduce     +->
|23|   +------+   |Operation  |
|4 |   |Emb 23|   +-----------+
|1 |   +------+
+--+   |Emb 4 |
       +------+
       |Emb 1 |
       +------+

These are the parameters available for the embed encoder

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 50): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • dropout (default false): determines if there should be a dropout layer before returning the encoder output.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the embedding weights are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • reduce_output (default sum): defines how to reduce the output tensor along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension) and null or None (which does not reduce and returns the full tensor).

Example sequence feature entry in the output features list using an embed encoder:

name: sequence_csv_column_name
type: sequence
encoder: embed
tied_weights: null
representation: dense
embedding_size: 256
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
dropout: false
initializer: null
regularize: true
reduce_output: sum
Parallel CNN Encoder

The parallel cnn encoder is inspired by Yoon Kim's Convolutional Neural Network for Sentence Classification. It works by first mapping the input integer sequence b x s (where b is the batch size and s is the length of the sequence) into a sequence of embeddings, then it passes the embedding through a number of parallel 1d convolutional layers with different filter size (by default 4 layers with filter size 2, 3, 4 and 5), followed by max pooling and concatenation. This single vector concatenating the outputs of the parallel convolutional layers is then passed through a stack of fully connected layers and returned as a b x h tensor where h is the output size of the last fully connected layer. If you want to output the full b x s x h tensor, you can specify reduce_output: null.

                   +-------+   +----+
                +-->1D Conv+--->Pool+-+
       +------+ |  |Width 2|   +----+ |
       |Emb 12| |  +-------+          |
       +------+ |                     |
+--+   |Emb 7 | |  +-------+   +----+ |
|12|   +------+ +-->1D Conv+--->Pool+-+
|7 |   |Emb 43| |  |Width 3|   +----+ |           +---------+
|43|   +------+ |  +-------+          | +------+  |Fully    |
|65+--->Emb 65+-+                     +->Concat+-->Connected+->
|23|   +------+ |  +-------+   +----+ | +------+  |Layers   |
|4 |   |Emb 23| +-->1D Conv+--->Pool+-+           +---------+
|1 |   +------+ |  |Width 4|   +----+ |
+--+   |Emb 4 | |  +-------+          |
       +------+ |                     |
       |Emb 1 | |  +-------+   +----+ |
       +------+ +-->1D Conv+--->Pool+-+
                   |Width 5|   +----+
                   +-------+

These are the available for an parallel cnn encoder:

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 256): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • conv_layers (default null): it is a list of dictionaries containing the parameters of all the convolutional layers. The length of the list determines the number of parallel convolutional layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: filter_size, num_filters, pool, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both conv_layers and num_conv_layers are null, a default list will be assigned to conv_layers with the value [{filter_size: 2}, {filter_size: 3}, {filter_size: 4}, {filter_size: 5}].
  • num_conv_layers (default null): if conv_layers is null, this is the number of parallel convolutional layers.
  • filter_size (default 3): if a filter_size is not already specified in conv_layers this is the default filter_size that will be used for each layer. It indicates how wide is the 1d convolutional filter.
  • num_filters (default 256): if a num_filters is not already specified in conv_layers this is the default num_filters that will be used for each layer. It indicates the number of filters, and by consequence the output channels of the 1d convolution.
  • pool_size (default null): if a pool_size is not already specified in conv_layers this is the default pool_size that will be used for each layer. It indicates the size of the max pooling that will be performed along the s sequence dimension after the convolution operation.
  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both fc_layers and num_fc_layers are null, a default list will be assigned to fc_layers with the value [{fc_size: 512}, {fc_size: 256}]. (only applies if reduce_output is not null).
  • num_fc_layers (default null): if fc_layers is null, this is the number of stacked fully connected layers (only applies if reduce_output is not null).
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in conv_layers or fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in conv_layers or fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null it uses glorot_uniform. Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if a regularize is not already specified in conv_layers or fc_layers this is the default regularize that will be used for each layer. It indicates if the layer weights should be considered when computing a regularization loss.
  • reduce_output (default sum): defines how to reduce the output tensor along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the sequence dimension), last (returns the last vector of the sequence dimension) and null or None (which does not reduce and returns the full tensor).

Example sequence feature entry in the output features list using a parallel cnn encoder:

name: sequence_csv_column_name
type: sequence
encoder: parallel_cnn
tied_weights: null
representation: dense
embedding_size: 256
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
conv_layers: null
num_conv_layers: null
filter_size: 3
num_filters: 256
pool_size: null
fc_layers: null
num_fc_layers: null
fc_size: 256
activation: relu
norm: null
dropout: false
regularize: true
reduce_output: sum
Stacked CNN Encoder

The stacked cnn encoder is inspired by Xiang Zhang at all's Character-level Convolutional Networks for Text Classification. It works by first mapping the input integer sequence b x s (where b is the batch size and s is the length of the sequence) into a sequence of embeddings, then it passes the embedding through a stack of 1d convolutional layers with different filter size (by default 6 layers with filter size 7, 7, 3, 3, 3 and 3), followed by an optional final pool and by a flatten operation. This single flatten vector is then passed through a stack of fully connected layers and returned as a b x h tensor where h is the output size of the last fully connected layer. If you want to output the full b x s x h tensor, you can specify the pool_size of all your conv_layers to be null and reduce_output: null, while if pool_size has a value different from null and reduce_output: null the returned tensor will be of shape b x s' x h, where s' is width of the output of the last convolutional layer.

       +------+
       |Emb 12|
       +------+
+--+   |Emb 7 |
|12|   +------+
|7 |   |Emb 43|   +----------------+  +---------+
|43|   +------+   |1D Conv         |  |Fully    |
|65+--->Emb 65+--->Layers          +-->Connected+->
|23|   +------+   |Different Widths|  |Layers   |
|4 |   |Emb 23|   +----------------+  +---------+
|1 |   +------+
+--+   |Emb 4 |
       +------+
       |Emb 1 |
       +------+

These are the parameters available for the stack cnn encoder:

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 256): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • conv_layers (default null): it is a list of dictionaries containing the parameters of all the convolutional layers. The length of the list determines the number of stacked convolutional layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: filter_size, num_filters, pool_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both conv_layers and num_conv_layers are null, a default list will be assigned to conv_layers with the value [{filter_size: 7, pool_size: 3, regularize: false}, {filter_size: 7, pool_size: 3, regularize: false}, {filter_size: 3, pool_size: null, regularize: false}, {filter_size: 3, pool_size: null, regularize: false}, {filter_size: 3, pool_size: null, regularize: true}, {filter_size: 3, pool_size: 3, regularize: true}].
  • num_conv_layers (default null): if conv_layers is null, this is the number of stacked convolutional layers.
  • filter_size (default 3): if a filter_size is not already specified in conv_layers this is the default filter_size that will be used for each layer. It indicates how wide is the 1d convolutional filter.
  • num_filters (default 256): if a num_filters is not already specified in conv_layers this is the default num_filters that will be used for each layer. It indicates the number of filters, and by consequence the output channels of the 1d convolution.
  • pool_size (default null): if a pool_size is not already specified in conv_layers this is the default pool_size that will be used for each layer. It indicates the size of the max pooling that will be performed along the s sequence dimension after the convolution operation.
  • reduce_output (default max): defines how to reduce the output tensor of the convolutional layers along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension) and null or None (which does not reduce and returns the full tensor).
  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both fc_layers and num_fc_layers are null, a default list will be assigned to fc_layers with the value [{fc_size: 512}, {fc_size: 256}]. (only applies if reduce_output is not null).
  • num_fc_layers (default null): if fc_layers is null, this is the number of stacked fully connected layers (only applies if reduce_output is not null).
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in conv_layers or fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in conv_layers or fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null it uses glorot_uniform. Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if a regularize is not already specified in conv_layers or fc_layers this is the default regularize that will be used for each layer. It indicates if the layer weights should be considered when computing a regularization loss.
  • reduce_output (default sum): defines how to reduce the output tensor along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension) and null or None (which does not reduce and returns the full tensor).

Example sequence feature entry in the output features list using a parallel cnn encoder:

name: sequence_csv_column_name
type: sequence
encoder: stacked_cnn
tied_weights: null
representation: dense
embedding_size: 256
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
conv_layers: null
num_conv_layers: null
filter_size: 3
num_filters: 256
pool_size: null
fc_layers: null
num_fc_layers: null
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true
reduce_output: max
Stacked Parallel CNN Encoder

The stacked parallel cnn encoder is a combination of the Parallel CNN and the Stacked CNN encoders where each layer of the stack is a composed of parallel convolutional layers. It works by first mapping the input integer sequence b x s (where b is the batch size and s is the length of the sequence) into a sequence of embeddings, then it passes the embedding through a stack of several parallel 1d convolutional layers with different filter size, followed by an optional final pool and by a flatten operation. This single flatten vector is then passed through a stack of fully connected layers and returned as a b x h tensor where h is the output size of the last fully connected layer. If you want to output the full b x s x h tensor, you can specify reduce_output: null.

                   +-------+                      +-------+
                +-->1D Conv+-+                 +-->1D Conv+-+
       +------+ |  |Width 2| |                 |  |Width 2| |
       |Emb 12| |  +-------+ |                 |  +-------+ |
       +------+ |            |                 |            |
+--+   |Emb 7 | |  +-------+ |                 |  +-------+ |
|12|   +------+ +-->1D Conv+-+                 +-->1D Conv+-+
|7 |   |Emb 43| |  |Width 3| |                 |  |Width 3| |                   +---------+
|43|   +------+ |  +-------+ | +------+  +---+ |  +-------+ | +------+  +----+  |Fully    |
|65+--->Emb 65+-+            +->Concat+-->...+-+            +->Concat+-->Pool+-->Connected+->
|23|   +------+ |  +-------+ | +------+  +---+ |  +-------+ | +------+  +----+  |Layers   |
|4 |   |Emb 23| +-->1D Conv+-+                 +-->1D Conv+-+                   +---------+
|1 |   +------+ |  |Width 4| |                 |  |Width 4| |
+--+   |Emb 4 | |  +-------+ |                 |  +-------+ |
       +------+ |            |                 |            |
       |Emb 1 | |  +-------+ |                 |  +-------+ |
       +------+ +-->1D Conv+-+                 +-->1D Conv+-+
                   |Width 5|                      |Width 5|
                   +-------+                      +-------+

These are the available parameters for the stack parallel cnn encoder:

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 256): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • stacked_layers (default null): it is a of lists of list of dictionaries containing the parameters of the stack of parallel convolutional layers. The length of the list determines the number of stacked parallel convolutional layers, length of the sub-lists determines the number of parallel conv layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: filter_size, num_filters, pool_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both stacked_layers and num_stacked_layers are null, a default list will be assigned to stacked_layers with the value [[{filter_size: 2}, {filter_size: 3}, {filter_size: 4}, {filter_size: 5}], [{filter_size: 2}, {filter_size: 3}, {filter_size: 4}, {filter_size: 5}], [{filter_size: 2}, {filter_size: 3}, {filter_size: 4}, {filter_size: 5}]].
  • num_stacked_layers (default null): if stacked_layers is null, this is the number of elements in the stack of parallel convolutional layers.
  • filter_size (default 3): if a filter_size is not already specified in conv_layers this is the default filter_size that will be used for each layer. It indicates how wide is the 1d convolutional filter.
  • num_filters (default 256): if a num_filters is not already specified in conv_layers this is the default num_filters that will be used for each layer. It indicates the number of filters, and by consequence the output channels of the 1d convolution.
  • pool_size (default null): if a pool_size is not already specified in conv_layers this is the default pool_size that will be used for each layer. It indicates the size of the max pooling that will be performed along the s sequence dimension after the convolution operation.
  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both fc_layers and num_fc_layers are null, a default list will be assigned to fc_layers with the value [{fc_size: 512}, {fc_size: 256}]. (only applies if reduce_output is not null).
  • num_fc_layers (default null): if fc_layers is null, this is the number of stacked fully connected layers (only applies if reduce_output is not null).
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • norm (default null): if a norm is not already specified in conv_layers or fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output.
  • activation (default relu): if an activation is not already specified in conv_layers or fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • regularize (default true): if a regularize is not already specified in conv_layers or fc_layers this is the default regularize that will be used for each layer. It indicates if the layer weights should be considered when computing a regularization loss.
  • reduce_output (default sum): defines how to reduce the output tensor along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension) and null or None (which does not reduce and returns the full tensor).

Example sequence feature entry in the output features list using a parallel cnn encoder:

name: sequence_csv_column_name
type: sequence
encoder: stacked_parallel_cnn
tied_weights: null
representation: dense
embedding_size: 256
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
stacked_layers: null
num_stacked_layers: null
filter_size: 3
num_filters: 256
pool_size: null
fc_layers: null
num_fc_layers: null
fc_size: 256
norm: null
activation: relu
regularize: true
reduce_output: max
RNN Encoder

The rnn encoder works by first mapping the input integer sequence b x s (where b is the batch size and s is the length of the sequence) into a sequence of embeddings, then it passes the embedding through a stack of recurrent layers (by default 1 layer), followed by a reduce operation that by default only returns the last output, but can perform other reduce functions. If you want to output the full b x s x h where h is the size of the output of the last rnn layer, you can specify reduce_output: null.

       +------+
       |Emb 12|
       +------+
+--+   |Emb 7 |
|12|   +------+
|7 |   |Emb 43|                 +---------+
|43|   +------+   +----------+  |Fully    |
|65+--->Emb 65+--->RNN Layers+-->Connected+->
|23|   +------+   +----------+  |Layers   |
|4 |   |Emb 23|                 +---------+
|1 |   +------+
+--+   |Emb 4 |
       +------+
       |Emb 1 |
       +------+

These are the available parameters for the rnn encoder:

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 256): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • num_layers (default 1): the number of stacked recurrent layers.
  • cell_type (default rnn): the type of recurrent cell to use. Available values are: rnn, lstm, lstm_block, lstm, ln, lstm_cudnn, gru, gru_block, gru_cudnn. For reference about the differences between the cells please refer to TensorFlow's documentation. We suggest to use the block variants on CPU and the cudnn variants on GPU because of their increased speed.
  • state_size (default 256): the size of the state of the rnn.
  • bidirectional (default false): if true two recurrent networks will perform encoding in the forward and backward direction and their outputs will be concatenated.
  • dropout (default false): determines if there should be a dropout layer before returning the encoder output.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the embedding weights are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • reduce_output (default last): defines how to reduce the output tensor along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension) and null or None (which does not reduce and returns the full tensor).

Example sequence feature entry in the output features list using a parallel cnn encoder:

name: sequence_csv_column_name
type: sequence
encoder: rnn
tied_weights: null
representation: dense
embedding_size: 256
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
num_layers: 1
cell_type: rnn
state_size: 256
bidirectional: false
dropout: false
initializer: null
regularize: true
reduce_output: sum
CNN RNN Encoder

The cnn rnn encoder works by first mapping the input integer sequence b x s (where b is the batch size and s is the length of the sequence) into a sequence of embeddings, then it passes the embedding through a stack of convolutional layers (by default 2), that is followed by a stack of recurrent layers (by default 1), followed by a reduce operation that by default only returns the last output, but can perform other reduce functions. If you want to output the full b x s x h where h is the size of the output of the last rnn layer, you can specify reduce_output: null.

       +------+
       |Emb 12|
       +------+
+--+   |Emb 7 |
|12|   +------+
|7 |   |Emb 43|                                +---------+
|43|   +------+   +----------+   +----------+  |Fully    |
|65+--->Emb 65+--->CNN Layers+--->RNN Layers+-->Connected+->
|23|   +------+   +----------+   +----------+  |Layers   |
|4 |   |Emb 23|                                +---------+
|1 |   +------+
+--+   |Emb 4 |
       +------+
       |Emb 1 |
       +------+

These are the available parameters of the cnn rnn encoder:

  • representation' (default dense): the possible values are dense and sparse. dense means the embeddings are initialized randomly, sparse means they are initialized to be one-hot encodings.
  • embedding_size (default 256): it is the maximum embedding size, the actual size will be min(vocabulary_size, embedding_size) for dense representations and exactly vocabulary_size for the sparse encoding, where vocabulary_size is the number of different strings appearing in the training set in the column the feature is named after (plus 1 for <UNK>).
  • embeddings_on_cpu (default false): by default embeddings matrices are stored on GPU memory if a GPU is used, as it allows for faster access, but in some cases the embedding matrix may be really big and this parameter forces the placement of the embedding matrix in regular memory and the CPU is used to resolve them, slightly slowing down the process as a result of data transfer between CPU and GPU memory.
  • pretrained_embeddings (default null): by default dense embeddings are initialized randomly, but this parameter allow to specify a path to a file containing embeddings in the GloVe format. When the file containing the embeddings is loaded, only the embeddings with labels present in the vocabulary are kept, the others are discarded. If the vocabulary contains strings that have no match in the embeddings file, their embeddings are initialized with the average of all other embedding plus some random noise to make them different from each other. This parameter has effect only if representation is dense.
  • embeddings_trainable (default true): If true embeddings are trained during the training process, if false embeddings are fixed. It may be useful when loading pretrained embeddings for avoiding finetuning them. This parameter has effect only for representation is dense as sparse one-hot encodings are not trainable.
  • conv_layers (default null): it is a list of dictionaries containing the parameters of all the convolutional layers. The length of the list determines the number of stacked convolutional layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: filter_size, num_filters, pool_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both conv_layers and num_conv_layers are null, a default list will be assigned to conv_layers with the value [{filter_size: 7, pool_size: 3, regularize: false}, {filter_size: 7, pool_size: 3, regularize: false}, {filter_size: 3, pool_size: null, regularize: false}, {filter_size: 3, pool_size: null, regularize: false}, {filter_size: 3, pool_size: null, regularize: true}, {filter_size: 3, pool_size: 3, regularize: true}].
  • num_conv_layers (default null): if conv_layers is null, this is the number of parallel convolutional layers.
  • filter_size (default 3): if a filter_size is not already specified in conv_layers this is the default filter_size that will be used for each layer. It indicates how wide is the 1d convolutional filter.
  • num_filters (default 256): if a num_filters is not already specified in conv_layers this is the default num_filters that will be used for each layer. It indicates the number of filters, and by consequence the output channels of the 1d convolution.
  • pool_size (default null): if a pool_size is not already specified in conv_layers this is the default pool_size that will be used for each layer. It indicates the size of the max pooling that will be performed along the s sequence dimension after the convolution operation.
  • num_rec_layers (default 1): the number of stacked recurrent layers.
  • cell_type (default rnn): the type of recurrent cell to use. Available values are: rnn, lstm, lstm_block, lstm, ln, lstm_cudnn, gru, gru_block, gru_cudnn. For reference about the differences between the cells please refer to TensorFlow's documentstion. We suggest to use the block variants on CPU and the cudnn variants on GPU because of their increased speed.
  • state_size (default 256): the size of the state of the rnn.
  • bidirectional (default false): if true two recurrent networks will perform encoding in the forward and backward direction and their outputs will be concatenated.
  • dropout (default false): determines if there should be a dropout layer between conv_layers and before returning the encoder output.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the embedding weights are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • reduce_output (default last): defines how to reduce the output tensor along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension) and null or None (which does not reduce and returns the full tensor).

Example sequence feature entry in the output features list using a parallel cnn encoder:

name: sequence_csv_column_name
type: sequence
encoder: cnn_rnn
tied_weights: null
representation: dense
embedding_size: 256
embeddings_on_cpu: false
pretrained_embeddings: null
embeddings_trainable: true
conv_layers: null
num_conv_layers: null
filter_size: 3
num_filters: 256
pool_size: null
norm: null
activation: relu
num_rec_layers: 1
cell_type: rnn
state_size: 256
bidirectional: false
dropout: false
initializer: null
regularize: true
reduce_output: last
Passthrough Encoder

The passthrough decoder simply transforms each input value into a fleat value and adds a dimension to the input tensor, creating a b x s x 1 tensor where b is the batch size and s is the length of the sequence. The tensor is reduced along the s dimension to obtain a single vector of size h for each element of the batch. If you want to output the full b x s x h tensor, you can specify reduce_output: null. This encoder is not really useful for sequence or text features, but may be useful for timeseries features, as it allows for using them without any processing in later stages of the model, like in a sequence combiner for instance.

+--+   
|12|   
|7 |                    +-----------+
|43|   +------------+   |Aggregation|
|65+--->Cast float32+--->Reduce     +->
|23|   +------------+   |Operation  |
|4 |                    +-----------+
|1 |   
+--+   

These are the parameters available for the passthrough encoder

  • reduce_output (default null): defines how to reduce the output tensor along the s sequence length dimension if the rank of the tensor is greater than 2. Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension) and null or None (which does not reduce and returns the full tensor).

Example sequence feature entry in the output features list using an embed encoder:

name: sequence_csv_column_name
type: sequence
encoder: passthrough
reduce_output: null

Sequence Output Features and Decoders

Sequential features can be used when sequence tagging (classifying each element of an input sequence) or sequence generation needs to be performed. There are two decoders available for those to tasks names tagger and generator.

These are the available parameters of a sequence output feature

  • reduce_inputs (default sum): defines how to reduce an input that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • dependencies (default []): the output features this one is dependent on. For a detailed explanation refer to Output Features Dependencies.
  • reduce_dependencies (default sum): defines how to reduce the output of a dependent feature that is not a vector, but a matrix or a higher order tensor, on the first dimension 9second if you count the batch dimension). Available values are: sum, mean or avg, max, concat (concatenates along the first dimension), last (returns the last vector of the first dimension).
  • loss (default {type: softmax_cross_entropy, class_similarities_temperature: 0, class_weights: 1, confidence_penalty: 0, distortion: 1, labels_smoothing: 0, negative_samples: 0, robust_lambda: 0, sampler: null, unique: false}): is a dictionary containing a loss type. The available losses type are softmax_cross_entropy and sampled_softmax_cross_entropy. For details on both losses, please refer to the category feature output feature section.
Tagger Decoder

In the case of tagger the decoder is a (potentially empty) stack of fully connected layers, followed by a projection into a tensor of size b x s x c, where b is the batch size, s is the length of the sequence and c is the number of classes, followed by a softmax_cross_entropy. This decoder requires its input to be shaped as b x s x h, where h is an hidden dimension, which is the output of a sequence, text or timeseries input feature without reduced outputs or the output of a sequence-based combiner. If a b x h input is provided instead, an error will be raised during model building.

Combiner
Output

+---+                 +----------+   +-------+
|emb|   +---------+   |Projection|   |Softmax|
+---+   |Fully    |   +----------+   +-------+
|...+--->Connected+--->...       +--->...    |
+---+   |Layers   |   +----------+   +-------+
|emb|   +---------+   |Projection|   |Softmax|
+---+                 +----------+   +-------+

These are the available parameters of a tagger decoder:

  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, dropout, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the decoder will be used instead.
  • num_fc_layers (default 0): this is the number of stacked fully connected layers that the input to the feature passes through. Their output is projected in the feature's output space.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the weights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).

Example sequence feature entry using a tagger decoder (with default parameters) in the output features list:

name: sequence_csv_column_name
type: sequence
reduce_inputs: sum
dependencies: []
reduce_dependencies: sum
loss:
    type: softmax_cross_entropy
    confidence_penalty: 0
    robust_lambda: 0
    class_weights: 1
    class_similarities: null
    class_similarities_temperature: 0
    labels_smoothing: 0
    negative_samples: 0
    sampler: null
    distortion: 1
    unique: false
fc_layers: null
num_fc_layers: 0
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true
Generator Decoder

In the case of generator the decoder is a (potentially empty) stack of fully connected layers, followed by an rnn that generates outputs feeding on its own previous predictions and generates a tensor of size b x s' x c, where b is the batch size, s' is the length of the generated sequence and c is the number of classes, followed by a softmax_cross_entropy. By default a generator expects a b x h shaped input tensor, where h is a hidden dimension. The h vectors are (after an optional stack of fully connected layers) fed into the rnn generator. One exception is when the generator uses attention, as in that case the expected size of the input tensor is b x s x h, which is the output of a sequence, text or timeseries input feature without reduced outputs or the output of a sequence-based combiner. If a b x h input is provided to a generator decoder using an rnn with attention instead, an error will be raised during model building.

                            Output     Output
                               1  +-+    ... +--+    END
                               ^    |     ^     |     ^
+--------+   +---------+       |    |     |     |     |
|Combiner|   |Fully    |   +---+--+ | +---+---+ | +---+--+
|Output  +--->Connected+---+RNN   +--->RNN... +--->RNN   |
|        |   |Layers   |   +---^--+ | +---^---+ | +---^--+
+--------+   +---------+       |    |     |     |     |
                              GO    +-----+     +-----+

These are the available parameters of a tagger decoder:

  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, dropout, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the decoder will be used instead.
  • num_fc_layers (default 0): this is the number of stacked fully connected layers that the input to the feature passes through. Their output is projected in the feature's output space.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the weights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
  • cell_type (default rnn): the type of recurrent cell to use. Available values are: rnn, lstm, lstm_block, lstm, ln, lstm_cudnn, gru, gru_block, gru_cudnn. For reference about the differences between the cells please refer to TensorFlow's documentstion. We suggest to use the block variants on CPU and the cudnn variants on GPU because of their increased speed.
  • state_size (default 256): the size of the state of the rnn.
  • tied_embeddings (default null): if null the embeddings of the targets are initialized randomly, while if the values is the name of an input feature, the embeddings of that input feature will be used as embeddings of the target. The vocabulary_size of that input feature has to be the same of the output feature one and it has to have an embedding matrix (binary and numerical features will not have one, fo instance). In this case the embedding_size will be the same as the state_size. This is useful for implementing autoencoders where the encoding and decoding part of the model share parameters.
  • embedding_size (default 256): if tied_target_embeddings is false, the input embeddings and the weights of the softmax_cross_entropy weights before the softmax_cross_entropy are not tied together and can have different sizes, this parameter describes the size of the embeddings of the inputs of the generator.
  • beam_width (default 1): sampling from the rnn generator is performed using beam search. By default, with a beam of one, only a greedy sequence using always the most probably next token is generated, but the beam size can be increased. This usually leads to better performance at the expense of more computation and slower generation.
  • attention_mechanism (default null): the recurrent generator may use an attention mechanism. The available ones are bahdanau and luong (for more information refer to TensorFlow's documentation). When attention is not null the expected size of the input tensor is b x s x h, which is the output of a sequence, text or timeseries input feature without reduced outputs or the output of a sequence-based combiner. If a b x h input is provided to a generator decoder using an rnn with attention instead, an error will be raised during model building.

Example sequence feature entry using a tagger decoder (with default parameters) in the output features list:

name: sequence_csv_column_name
type: sequence
reduce_inputs: sum
dependencies: []
reduce_dependencies: sum
loss:
    type: softmax_cross_entropy
    confidence_penalty: 0
    robust_lambda: 0
    class_weights: 1
    class_similarities: null
    class_similarities_temperature: 0
    labels_smoothing: 0
    negative_samples: 0
    sampler: null
    distortion: 1
    unique: false
fc_layers: null
num_fc_layers: 0
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true
cell_type: rnn
state_size: 256
tied_target_embeddings: true
embedding_size: 256
beam_width: 1
attention_mechanism: null

Sequence Features Measures

The measures that are calculated every epoch and are available for category features are accuracy (counts the number of datapoints where all the elements of the predicted sequence are correct over the number of all datapoints), token_accuracy (computes the number of elements in all the sequences that are correctly predicted over the number of all the elements in all the sequences), last_accuracy (accuracy considering only the last element of the sequence, it is useful for being sure special end-of-sequence tokens are generated or tagged), edit_distance (the levenshtein distance between the predicted and ground truth sequence), perplexity (the perplexity of the ground truth sequence according to the model) and the loss itself. You can set either of them as validation_measure in the training section of the model definition if you set the validation_field to be the name of a sequence feature.

Text Features

Text Features Preprocessing

Text features are treated in the same way of sequence features, with a couple differences. Two different formattings/splittings happen, one that splits at every character and one that uses a spaCy based tokenizer (and removes stopwords) are used, and two different key are added to the HDF5 file, one containing the matrix of characters and one containing the matrix of words. The same thing happens in the JSON file, where there are dictionaries for mapping characters to integers (and the inverse) and words to integers (and their inverse). In the model definition you are able to specify which level of representation to use, if the character level or the word level.

The parameters available for preprocessing are:

  • missing_value_strategy (default fill_with_const): what strategy to follow when there's a missing value in a binary column. The value should be one of fill_with_const (replaces the missing value with a specific value specified with the fill_value parameter), fill_with_mode (replaces the missing values with the most frequent value in the column), fill_with_mean (replaces the missing values with the mean of the values in the column), backfill (replaces the missing values with the next valid value).
  • fill_value (default ""): the value to replace the missing values with in case the missing_value_strategy is fill-value.
  • padding (default right): the direction of the padding. right and left are available options.
  • padding_symbol (default <PAD>): the string used as a padding symbol. Is is mapped to the integer ID 0 in the vocabulary.
  • unknown_symbol (default <UNK>): the string used as a unknown symbol. Is is mapped to the integer ID 1 in the vocabulary.
  • lowercase (default false): if the string has to be lowercased before being handled by the formatter.
  • word_sequence_length_limit (default 256): the maximum length of the text in words. Texts that are longer than this value will be truncated, while texts that are shorter will be padded.
  • word_format (default space_punct): defines how to map from the raw string content of the CSV column to a sequence of words. The default value space_punct splits the string using a regular expression that separates also punctuation. Other options are: space (splits on space), underscore (splits on underscore), comma(splits on comma), json (decodes the string into a set or a list through a JSON parser), and a set of format functions that rely on spaCy.
  • word_most_common (default 20000): the maximum number of most common words to be considered. If the data contains more than this amount, the most infrequent words will be treated as unknown.
  • char_sequence_length_limit (default 1024): the maximum length of the text in characters. Texts that are longer than this value will be truncated, while sequences that are shorter will be padded.
  • char_format (default characters): defines how to map from the raw string content of the CSV column to a sequence of characters. The default value and only available option is characters and the behavior is to split the string at each character.
  • char_most_common (default 70): the maximum number of most common characters to be considered. if the data contains more than this amount, the most infrequent characters will be treated as unknown.
spaCy based word format options

The spaCy based word_format options are functions that use the powerful tokenization and NLP preprocessing models provided the library. Several languages are available: English (code en), Italian (code it), Spanish (code es), German (code de), French (code fr), Portuguese (code pt), Dutch (code nl), Greek (code el) and Multi (code xx, useful in case you have a dataset of different languages). For each language different functions are available: - tokenize: uses spaCy tokenizer, - tokenize_filter: uses spaCy tokenizer and filters out punctuation, numbers, stopwords and words shorter than 3 characters, - tokenize_remove_stopwords: uses spaCy tokenizer and filters out stopwords, - lemmatize: uses spaCy lemmatizer, - lemmatize_filter: uses spaCy lemmatizer and filters out punctuation, numbers, stopwords and words shorter than 3 characters, - lemmatize_remove_stopwords: uses spaCy lemmatize and filters out stopwords.

In order to use these options, you have to download the the spaCy model:

python -m spacy download <language_code>

and provide <language>_<function> as word_format like: english_tokenizer, italian_lemmatize_filter, multi_tokenize_filter and so on. More details on the models can be found in the spaCy documentation.

Text Input Features and Encoders

The encoders are the same used for the Sequence Features. The only difference is that you can specify an additional level parameter with possible values word or char to force to use the text words or characters as inputs (by default the encoder will use word).

Text Output Features and Decoders

The decoders are the same used for the Sequence Features. The only difference is that you can specify an additional level parameter with possible values word or char to force to use the text words or characters as inputs (by default the encoder will use word).

Text Features Measures

The measures are the same used for the Sequence Features.

Time Series Features

Time Series Features Preprocessing

Timeseries features are treated in the same way of sequence features, with the only difference being that the matrix in the HDF5 file does not have integer values, but float values. Moreover, there is no need for any mapping in the JSON file.

Time Series Input Features and Encoders

The encoders are the same used for the Sequence Features. The only difference is that time series features don't have an embedding layer at the beginning, so the b x s placeholders (where b is the batch size and s is the sequence length) are directly mapped to a b x s x 1 tensor and then passed to the different sequential encoders.

Time Series Output Features and Decoders

There are no time series decoders at the moment (WIP), so time series cannot be used as output features.

Time Series Features Measures

As no time series decoders are available at the moment, there are also no time series measures.

Image Features

Image Features Preprocessing

Ludwig supports both grayscale and color images, the number of channels is inferred, but make sure all your images have the same number of channels. During preprocessing raw image files are transformed into numpy ndarrays and saved in the hdf5 format. Images should have the same size. If they have different sizes they can be converted to the same size which should be set in the feature preprocessing parameters.

  • in_memory (default true): defines whether image dataset will reside in memory during the training process or will be dynamically fetched from disk (useful for large datasets). In the latter case a training batch of input images will be fetched from disk each training iteration.
  • resize_method (default crop_or_pad): available options: crop_or_pad - crops larger images to the desired size or pads smalled images using edge padding; interpolate - uses interpolation.
  • height (default null): image height in pixels, must be set if resizing is required
  • width (default null): image width in pixels, must be set if resizing is required
  • num_channels (default null): number of channels in the images. By default, if the value is null, the number of channels of the first image of the dataset will be used and if there is an image in the dataset with a different number of channels, an error will be reported. If the value specified is not null, images in the dataset will be adapted to the specified size. If the value is 1, all images with more then one channel will be greyscaled and reduced to one channel (trasparecy will be lost). If the value is 3 all images with 1 channel will be repeated 3 times to obtain 3 channels, while images with 4 channels will lose the transparecy channel. If the value is 4, all the images with less than 4 channels will have the remaining channels filled with zeros.

Depending on the application, do not to exceed a size of 256 x 256 as bigger sizes will, in most cases, not provide much advantage and considerably slow down trainin and inference and also make both forward and backward passes consume a lot of memory leading to memory overflow on machines with limited amounts of RAM or on GPUs with limited amounts of VRAM.

Example of a preprocessing specification:

name: image_feature_name
type: image
preprocessing:
  heights: 128
  width: 128
  resize_method: crop_or_pad

Image Input Features and Encoders

Input image features are transformed into a float valued tensors of size N x H x W x C (where N is the size of the dataset and H x W is a specific resizing of the image that can be set, and C is the number of channels) and added to HDF5 with a key that reflects the name of column in the CSV. The column name is added to the JSON file, with an associated dictionary containing preprocessing information about the sizes of the resizing.

Currently there are two encoders supported for images: Convolutional Stack Encoder and ResNet encoder which can be set by setting encoder parameter to stacked_cnn or resnet in the input feature dictionary in the model definition (stacked_cnn is the default one).

Convolutional Stack Encoder

Convolutional Stack Encoder takes the following optional parameters:

  • conv_layers (default null): it is a list of dictionaries containing the parameters of all the convolutional layers. The length of the list determines the number of stacked convolutional layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: filter_size, num_filters, pool_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both conv_layers and num_conv_layers are null, a default list will be assigned to conv_layers with the value [{filter_size: 7, pool_size: 3, regularize: false}, {filter_size: 7, pool_size: 3, regularize: false}, {filter_size: 3, pool_size: null, regularize: false}, {filter_size: 3, pool_size: null, regularize: false}, {filter_size: 3, pool_size: null, regularize: true}, {filter_size: 3, pool_size: 3, regularize: true}].
  • num_conv_layers (default null): if conv_layers is null, this is the number of stacked convolutional layers.
  • filter_size (default 3): if a filter_size is not already specified in conv_layers this is the default filter_size that will be used for each layer. It indicates how wide is the 1d convolutional filter.
  • num_filters (default 256): if a num_filters is not already specified in conv_layers this is the default num_filters that will be used for each layer. It indicates the number of filters, and by consequence the output channels of the 2d convolution.
  • pool_stride (default 1): if a pool_stride is not already specified in conv_layers this is the default pool_stride that will be used for each layer.
  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both fc_layers and num_fc_layers are null, a default list will be assigned to fc_layers with the value [{fc_size: 512}, {fc_size: 256}]. (only applies if reduce_output is not null).
  • num_fc_layers (default 1): This is the number of stacked fully connected layers.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • norm (default null): if a norm is not already specified in fc_layers or conv_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • activation (default relu): if an activation is not already specified in fc_layers or conv_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the weights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).
ResNet Encoder

ResNet Encoder takes the following optional parameters:

  • resnet_size (default 50): A single integer for the size of the ResNet model. If has to be one of the following values: 8, 14, 18, 34, 50, 101, 152, 200.
  • num_filters (default 16): It indicates the number of filters, and by consequence the output channels of the 2d convolution.
  • kernel_size (default 3): The kernel size to use for convolution.
  • conv_stride (default 1): Stride size for the initial convolutional layer.
  • first_pool_size (default null): Pool size to be used for the first pooling layer. If none, the first pooling layer is skipped.
  • batch_norm_momentum (default 0.9): Momentum of the batch norm running statistics. The suggested parameter in TensorFlow's implementation is 0.997, but that leads to a big discrepancy between the normalization at training time and test time, so the default value is a more conservative 0.9.
  • batch_norm_epsilon (default 0.001): Epsilon of the batch norm. The suggested parameter in TensorFlow's implementation is 1e-5, but that leads to a big discrepancy between the normalization at training time and test time, so the default value is a more conservative 0.001.
  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the encoder will be used instead. If both fc_layers and num_fc_layers are null, a default list will be assigned to fc_layers with the value [{fc_size: 512}, {fc_size: 256}]. (only applies if reduce_output is not null).
  • num_fc_layers (default 1): This is the number of stacked fully connected layers.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • norm (default null): if a norm is not already specified in fc_layers or conv_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • activation (default relu): if an activation is not already specified in fc_layers or conv_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the weights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).

Image Output Features and Decoders

There are no image decoders at the moment (WIP), so image cannot be used as output features.

Image Features Measures

As no image decoders are available at the moment, there are also no image measures.

Combiners

Combiners are the part of the model that take the outputs of the encoders of all input features and combine them before providing the combined representation to the different output decoders. If you don't specify a combiner, the concat combiner will be used.

Concat Combiner

The concat combiner assumes all outputs from encoders are tensors of size b x h where b is the batch size and h is the hidden dimension, which can be different for each input. It concatenates along the h dimension, and then (optionally) passes the concatenated tensor through a stack of fully connected layers. It returns the final b x h' tensor where h' is the size of the last fully connected layer or the sum of the sizes of the h of all inputs in the case there are no fully connected layers. If there's only one input feature and no fully connected layers are specified, the output of the input feature is just passed through as output.

+-----------+
|Input      |
|Feature 1  +-+
+-----------+ |            +---------+
+-----------+ | +------+   |Fully    |
|...        +--->Concat+--->Connected+->
+-----------+ | +------+   |Layers   |
+-----------+ |            +---------+
|Input      +-+
|Feature N  |
+-----------+

These are the available parameters of a concat combiner

  • fc_layers (default null): it is a list of dictionaries containing the parameters of all the fully connected layers. The length of the list determines the number of stacked fully connected layers and the content of each dictionary determines the parameters for a specific layer. The available parameters for each layer are: fc_size, norm, activation, dropout, initializer and regularize. If any of those values is missing from the dictionary, the default one specified as a parameter of the decoder will be used instead.
  • num_fc_layers (default 0): this is the number of stacked fully connected layers that the input to the feature passes through. Their output is projected in the feature's output space.
  • fc_size (default 256): if a fc_size is not already specified in fc_layers this is the default fc_size that will be used for each layer. It indicates the size of the output of a fully connected layer.
  • activation (default relu): if an activation is not already specified in fc_layers this is the default activation that will be used for each layer. It indicates the activation function applied to the output.
  • norm (default null): if a norm is not already specified in fc_layers this is the default norm that will be used for each layer. It indicates the norm of the output and it can be null, batch or layer.
  • dropout (default false): determines if there should be a dropout layer after each layer.
  • initializer (default null): the initializer to use. If null, the default initialized of each variable is used (glorot_uniform in most cases). Options are: constant, identity, zeros, ones, orthogonal, normal, uniform, truncated_normal, variance_scaling, glorot_normal, glorot_uniform, xavier_normal, xavier_uniform, he_normal, he_uniform, lecun_normal, lecun_uniform. Alternatively it is possible to specify a dictionary with a key type that identifies the type of initializer and other keys for its parameters, e.g. {type: normal, mean: 0, stddev: 0}. To know the parameters of each initializer, please refer to TensorFlow's documentation.
  • regularize (default true): if true the weights of the layers are added to the set of weights that get regularized by a regularization loss (if the regularization_lambda in training is greater than 0).

Example concat combiner in the model definition:

type: concat
fc_layers: null
num_fc_layers: 0
fc_size: 256
activation: relu
norm: null
dropout: false
initializer: null
regularize: true

Sequence Concat Combiner

The sequence concat combiner assumes at least one output from encoders is a tensors of size b x s x h where b is the batch size, s is the length of the sequence and h is the hidden dimension. The sequence / text / sequential input can be specified with the main_sequence_feature parameter that should have the name of the sequential feature as value. If no main_sequence_feature is specified, the combiner will look through all the features in the order they are defined in the model definition and will look for a feature with a rank 3 tensor output (sequence, text or time series). If it cannot find one it will raise an exception, otherwise the output of that feature will be used for concatenating the other features along the sequence s dimension.

If there are other input features with a rank 3 output tensor, the combiner will concatenate them alongside the s dimension, which means that all of them must have identical s dimension, otherwise an error will be thrown. Specifically, as the placeholders of the sequential features are of dimension [None, None] in order to make the BucketedBatcher trim longer sequences to their actual length, the check if the sequences are of the same length cannot be performed at model building time, and a dimension mismatch error will be returned during training when a datapoint with two sequential features of different lengths are provided.

Other features that have a b x h rank 2 tensor output will be replicated s times and concatenated to the s dimension. The final output is a b x s x h' tensor where h' is the size of the concatenation of the h dimensions of all input features.

+-----------+
|Input      |
|Feature 1  +-+
+-----------+ |            +---------+
+-----------+ | +------+   |Fully    |
|...        +--->Concat+--->Connected+->
+-----------+ | +------+   |Layers   |
+-----------+ |            +---------+
|Input      +-+
|Feature N  |
+-----------+

These are the available parameters of a sequence concat combiner

  • main_sequence_feature (default null): name fo the sequence / text/ time series feature to concatenate the outputs of the other features to. If no main_sequence_feature is specified, the combiner will look through all the features in the order they are defined in the model definition and will look for a feature with a rank 3 tensor output (sequence, text or time series). If it cannot find one it will raise an exception, otherwise the output of that feature will be used for concatenating the other features along the sequence s dimension. If there are other input features with a rank 3 output tensor, the combiner will concatenate them alongside the s dimension, which means that all of them must have identical s dimension, otherwise an error will be thrown.

Example sequence concat combiner in the model definition:

type: sequence_concat
main_sequence_feature: null

Sequence Combiner

The sequence combiner stacks a sequence concat combiner with a sequence encoder one on top of each other. All the considerations about inputs tensor ranks describer for the sequence concat combiner apply also in this case, but the main difference is that this combiner uses the b x s x h' output of the sequence concat combiner, where b is the batch size, s is the sequence length and h' is the sum of the hidden dimensions of all input features, as input fo any of the sequence encoders described in the sequence features encoders section. Refer to that section for more detailed information about the sequence encoders and their parameters. Also all the considerations on the shape of the outputs done for the sequence encoders apply in this case too.

Sequence
Feature
Output

+---------+
|emb seq 1|
+---------+
|...      +--+
+---------+  |  +-----------------+
|emb seq n|  |  |emb seq 1|emb oth|   +--------+
+---------+  |  +-----------------+   |Sequence|
             +-->...      |...    +-->+Encoder +->
Other        |  +-----------------+   |        |
Feature      |  |emb seq n|emb oth|   +--------+
Output       |  +-----------------+
             |
+-------+    |
|emb oth+----+
+-------+

Example sequence concat combiner in the model definition:

type: sequence
main_sequence_feature: null
encoder: parallel_cnn
... encoder parameters ...

Distributed Training

You can distribute the training and prediction of your models using Horovod, which allows to train on a single machine with multiple GPUs as well as on multiple machines with multiple GPUs.

In order to use distributed training you have to install Horovod as detailed in Horovod's installation instructions (which include installing OpenMPI or other MPI implementations) and then install the two packages:

pip install horovod mpi4py

Horovod works by, in practice, increasing the batch size and distributing a part of each batch to a different node and collecting the gradients from all the nodes in a smart and scalable way. It also adjusts the learning rate to counter balance the increase in the batch size. The advantage is that training speed scales almost linearly with the number of nodes.

experiment, train and predict commands accept a --horovod argument that instructs the model building, training and prediction phases to be conducted using Horovod in a distributed way. An MPI command specifying which machines and / or GPUs to use, together with a few more parameters, must be provided before the call to Ludwig's command. For instance, in order to train a Ludwig model on a local machine with four GPUs one you can run:

mpirun -np 4 \
    -H localhost:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    -mca pml ob1 -mca btl ^openib \
    ludwig train --horovod ...other Ludwig parameters...

While for training on four remote machines with four GPUs each you can run:

mpirun -np 16 \
    -H server1:4,server2:4,server3:4,server4:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    -mca pml ob1 -mca btl ^openib \
    ludwig train --horovod ...other Ludwig parameters...

The same applies to experiment and predict.

More details on the installation of MPI and how to run Horovod can be found in Horovod's documentation.

Programmatic API

Ludwig functionalities can also be accessed through a programmatic API. The API consists of one LudwigModel class that can be initialized with a model definition dictionary and then can be trained with data coming in the form of a dataframe or a CSV file. Pretrained models can be loaded and can be used to obtain predictions on new data, again either in dataframe or CSV format.

A detailed documentation of all the functions available in LudwigModel is provided in the API documentation.

Training a Model

To train a model one has first to initialize it using the initializer LudwigModel() and a model definition dictionary, and then calling the train() function using either a dataframe or a CSV file.

from ludwig.api import LudwigModel

model_definition = {...}
model = LudwigModel(model_definition)
train_stats = model.train(data_csv=csv_file_path)
# or
train_stats = model.train(data_df=dataframe)

model_definition is a dictionary that has the same key-value structure of a model definition YAML file, as it's technically equivalent as parsing the YAML file into a Python dictionary. train_statistics will be a dictionary containing statistics about the training. The contents are exactly the same of the training_statistics.json file produced by the experiment and train commands.

Loading a Pre-trained Model

In order to load a pre-trained Ludwig model you have to call the static function load() of the LudwigModel class providing the path containing the model.

from ludwig.api import LudwigModel

model = LudwigModel.load(model_path)

Predicting

Either a newly trained model or a pre-trained loaded model can be used for predicting on new data using the predict() function of the model object. The CSV / dataframe has to contain columns with the same names of all the input features of the model.

predictions = model.predict(dataset_csv=csv_file_path)
#or
predictions = model.predict(dataset_df=dataframe)

predictions will be a dataframe containing the prediction and confidence score / probability of all output features.

If you want to compute also measures on the quality of the predictions you can run:

predictions, test_stats = model.test(dataset_csv=csv_file_path)
#or
predictions, test_stats = model.test(dataset_df=dataframe)

In this case the CSV / dataframe should also contain columns with the same names of all the output features, as their content is going to be used as ground truth to compare the predictions against and compute the measures and test_statistics will be a dictionary containing several measures of quality depending on the type of each output feature (e.g. category features will have an accuracy measure and a confusion matrix, among other measures, associated to them, while numerical features will have measures like mean squared loss and R2 among others).

Visualizations

Several visualization can be obtained from the result files from both train, predict and experiment by using the visualize command. The command has several parameters, but not all the visualizations use all of them. Let's first present the parameters of the general script, and then, for each available visualization, we will discuss about the specific parameters needed and what visualization they produce.

usage: ludwig visualize [options]

This script analyzes results and shows some nice plots.

optional arguments:
  -h, --help            show this help message and exit
  -d DATA_CSV, --data_csv DATA_CSV
                        raw data file
  -g GROUND_TRUTH, --ground_truth GROUND_TRUTH
                        ground truth file
  -gm GROUND_TRUTH_METADATA, --ground_truth_metadata GROUND_TRUTH_METADATA
                        input metadata JSON file
  -v {learning_curves,compare_performance,compare_classifiers_performance_from_prob,compare_classifiers_performance_from_pred,compare_classifiers_performance_subset,compare_classifiers_performance_changing_k,compare_classifiers_multiclass_multimetric,compare_classifiers_predictions,compare_classifiers_predictions_distribution,confidence_thresholding,confidence_thresholding_data_vs_acc,confidence_thresholding_data_vs_acc_subset,confidence_thresholding_data_vs_acc_subset_per_class,confidence_thresholding_2thresholds_2d,confidence_thresholding_2thresholds_3d,binary_threshold_vs_metric,roc_curves,roc_curves_from_test_statistics,calibration_1_vs_all,calibration_multiclass,confusion_matrix,frequency_vs_f1}, --visualization {learning_curves,compare_performance,compare_classifiers_performance_from_prob,compare_classifiers_performance_from_pred,compare_classifiers_performance_subset,compare_classifiers_performance_changing_k,compare_classifiers_multiclass_multimetric,compare_classifiers_predictions,compare_classifiers_predictions_distribution,confidence_thresholding,confidence_thresholding_data_vs_acc,confidence_thresholding_data_vs_acc_subset,confidence_thresholding_data_vs_acc_subset_per_class,confidence_thresholding_2thresholds_2d,confidence_thresholding_2thresholds_3d,binary_threshold_vs_metric,roc_curves,roc_curves_from_test_statistics,calibration_1_vs_all,calibration_multiclass,confusion_matrix,frequency_vs_f1}
                        type of visualization
  -f FIELD, --field FIELD
                        field containing ground truth
  -tf THRESHOLD_FIELDS [THRESHOLD_FIELDS ...], --threshold_fields THRESHOLD_FIELDS [THRESHOLD_FIELDS ...]
                        fields for 2d threshold
  -pred PREDICTIONS [PREDICTIONS ...], --predictions PREDICTIONS [PREDICTIONS ...]
                        predictions files
  -prob PROBABILITIES [PROBABILITIES ...], --probabilities PROBABILITIES [PROBABILITIES ...]
                        probabilities files
  -trs TRAINING_STATS [TRAINING_STATS ...], --training_statistics TRAINING_STATS [TRAINING_STATS ...]
                        training stats files
  -tes TEST_STATS [TEST_STATS ...], --test_statistics TEST_STATS [TEST_STATS ...]
                        test stats files
  -mn MODEL_NAMES [MODEL_NAMES ...], --model_names MODEL_NAMES [MODEL_NAMES ...]
                        names of the models to use as labels
  -tn TOP_N_CLASSES [TOP_N_CLASSES ...], --top_n_classes TOP_N_CLASSES [TOP_N_CLASSES ...]
                        number of classes to plot
  -k TOP_K, --top_k TOP_K
                        number of elements in the ranklist to consider
  -ll LABELS_LIMIT, --labels_limit LABELS_LIMIT
                        maximum numbers of labels. If labels in dataset are
                        higher than this number, "rare" label
  -ss {ground_truth,predictions}, --subset {ground_truth,predictions}
                        type of subset filtering
  -n, --normalize       normalize rows in confusion matrix
  -m METRICS [METRICS ...], --metrics METRICS [METRICS ...]
                        metrics to dispay in threshold_vs_metric
  -pl POSITIVE_LABEL, --positive_label POSITIVE_LABEL
                        label of the positive class for the roc curve
  -l {critical,error,warning,info,debug,notset}, --logging_level {critical,error,warning,info,debug,notset}
                        the level of logging to use

Some additional information on the parameters:

  • The list parameters are considered to be aligned, meaning predictions, probabilities, training_statistics, test_statistics and model_names are indexed altogether, for instance the name of the model producing the second predictions in the list will be the second in the model names.
  • data_csv is intended to be the data the model(s) were trained on.
  • ground_truth and ground_truth_metadata are respectively the HDF5 and JSON file obtained during training preprocessing. If you plan to use the visualizations then be sure not to use the skip_save_preprocessing when training. Those files are needed because they contain the split performed at preprocessing time, so it is easy to extract the test set from them.
  • field is the output feature to use for creating the visualization.

Other parameters will be detailed for each visualization as different ones use them differently.

Learning Curves

learning_curves

This visualization uses the training_statistics and model_names parameters. For each model (in the aligned lists of training_statistics and model_names) and for each output feature and measure of the model, it produces a line plot showing how that measure changed over the course of the epochs of training on the training and validation sets.

Learning Curves Loss

Learning Curves Accuracy

Confusion Matrix

confusion_matrix

This visualization uses the top_n_classes, normalize, ground_truth_metadata, test_statistics and model_names parameters. For each model (in the aligned lists of test_statistics and model_names) it produces a heatmap of the confusion matrix in the predictions for each field that has a confusion matrix in test_statistics. The value of top_n_classes limits the heatmap to the n most frequent classes.

Confusion Matrix

The second plot produced, is a barplot showing the entropy of each class, ranked from most entropic to least entropic.

Confusion Matrix Entropy

Compare Performance

compare_performance

This visualization uses the field, test_statistics and model_names parameters. For each model (in the aligned lists of test_statistics and model_names) it produces bars in a bar plot, one for each overall metric available in the test_statistics file for the specified field.

Compare Classifiers Performance

compare_classifiers_performance_from_prob

This visualization uses the ground_truth, field, probabilities and model_names parameters. field needs to be a category. For each model (in the aligned lists of probabilities and model_names) it produces bars in a bar plot, one for each overall metric computed on the fly from the probabilities of predictions for the specified field.

Compare Classifiers Performance from Probabilities

compare_classifiers_performance_from_pred

This visualization uses the ground_truth, ground_truth_metadata, field, predictions and model_names parameters. field needs to be a category. For each model (in the aligned lists of predictions and model_names) it produces bars in a bar plot, one for each overall metric computed on the fly from the predictions for the specified field.

Compare Classifiers Performance from Predictions

compare_classifiers_performance_subset

This visualization uses the top_n_classes, subset, ground_truth, ground_truth_metadata, field, probabilities and model_names parameters. field needs to be a category. For each model (in the aligned lists of predictions and model_names) it produces bars in a bar plot, one for each overall metric computed on the fly from the probabilities predictions for the specified field, considering only a subset of the full training set. The way the subset is obtained is using the top_n_classes and subset parameters.

If the values of subset is ground_truth, then only datapoints where the ground truth class is within the top n most frequent ones will be considered as test set, and the percentage of datapoints that have been kept from the original set will be displayed.

Compare Classifiers Performance Subset Ground Truth

If the values of subset is predictions, then only datapoints where the the model predicts a class that is within the top n most frequent ones will be considered as test set, and the percentage of datapoints that have been kept from the original set will be displayed for each model.

Compare Classifiers Performance Subset Ground Predictions

compare_classifiers_performance_changing_k

This visualization uses the top_k, ground_truth_metadata, field, probabilities and model_names parameters. field needs to be a category. For each model (in the aligned lists of probabilities and model_names) it produces a line plot that shows the Hits@K measure (that counts a prediction as correct if the model produces it among the first k) while changing k from 1 to top_k for the specified field.

Compare Classifiers Performance Changing K

compare_classifiers_multiclass_multimetric

This visualization uses the top_n_classes, ground_truth_metadata, field, test_statistics and model_names parameters. field needs to be a category. For each model (in the aligned lists of test_statistics and model_names) it produces four plots that show the precision, recall and F1 of the model on several classes for the specified field.

The first one show the measures on the n most frequent classes.

Multiclass Multimetric top k

The second one shows the measures on the n classes where the model performs the best.

Multiclass Multimetric best k

The third one shows the measures on the n classes where the model performs the worst.

Multiclass Multimetric worst k

The fourth one shows the measures on all the classes, sorted by their frequency. This could become unreadable in case the number of classes is really high.

Multiclass Multimetric sorted

Compare Classifier Predictions

compare_classifiers_predictions

This visualization uses the ground_truth, ground_truth_metadata, field, predictions and model_names parameters. field needs to be a category and there must be two and only two models (in the aligned lists of predictions and model_names). This visualization produces a pie chart comparing the predictions of the two models for the specified field.

Compare Classifiers Predictions

compare_classifiers_predictions_distribution

This visualization uses the ground_truth, ground_truth_metadata, field, predictions and model_names parameters. field needs to be a category. This visualization produces a radar plot comparing the distributions of predictions of the models for the first 10 classes of the specified field.

Compare Classifiers Predictions Distribution

Confidence_Thresholding

confidence_thresholding

This visualization uses the ground_truth, field, probabilities and model_names parameters. field needs to be a category. For each model (in the aligned lists of probabilities and model_names) it produces a pair of lines indicating the accuracy of the model and the data coverage while increasing a threshold (x axis) on the probabilities of predictions for the specified field.

Confidence_Thresholding

confidence_thresholding_data_vs_acc

This visualization uses the ground_truth, field, probabilities and model_names parameters. field needs to be a category. For each model (in the aligned lists of probabilities and model_names) it produces a line indicating the accuracy of the model and the data coverage while increasing a threshold on the probabilities of predictions for the specified field. The difference with confidence_thresholding is that it uses two axes instead of three, not visualizing the threshold and having coverage as x axis instead of the threshold.

Confidence_Thresholding Data vs Accuracy

confidence_thresholding_data_vs_acc_subset

This visualization uses the top_n_classes, subset, ground_truth, field, probabilities and model_names parameters. field needs to be a category. For each model (in the aligned lists of probabilities and model_names) it produces a line indicating the accuracy of the model and the data coverage while increasing a threshold on the probabilities of predictions for the specified field, considering only a subset of the full training set. The way the subset is obtained is using the top_n_classes and subset parameters.. The difference with confidence_thresholding is that it uses two axes instead of three, not visualizing the threshold and having coverage as x axis instead of the threshold.

If the values of subset is ground_truth, then only datapoints where the ground truth class is within the top n most frequent ones will be considered as test set, and the percentage of datapoints that have been kept from the original set will be displayed. If the values of subset is predictions, then only datapoints where the the model predicts a class that is within the top n most frequent ones will be considered as test set, and the percentage of datapoints that have been kept from the original set will be displayed for each model.

Confidence_Thresholding Data vs Accuracy Subset

confidence_thresholding_data_vs_acc_subset_per_class

This visualization uses the top_n_classes, subset, ground_truth, ground_truth_metadata, field, probabilities and model_names parameters. field needs to be a category. For each model (in the aligned lists of probabilities and model_names) it produces a line indicating the accuracy of the model and the data coverage while increasing a threshold on the probabilities of predictions for the specified field, considering only a subset of the full training set. The way the subset is obtained is using the top_n_classes and subset parameters.. The difference with confidence_thresholding is that it uses two axes instead of three, not visualizing the threshold and having coverage as x axis instead of the threshold.

If the values of subset is ground_truth, then only datapoints where the ground truth class is within the top n most frequent ones will be considered as test set, and the percentage of datapoints that have been kept from the original set will be displayed. If the values of subset is predictions, then only datapoints where the the model predicts a class that is within the top n most frequent ones will be considered as test set, and the percentage of datapoints that have been kept from the original set will be displayed for each model.

The difference with confidence_thresholding_data_vs_acc_subset is that it produces one plot per class within the top_n_classes.

Confidence_Thresholding Data vs Accuracy Subset per class 1

Confidence_Thresholding Data vs Accuracy Subset per class 4

confidence_thresholding_2thresholds_2d

This visualization uses the ground_truth, threshold_fields, probabilities and model_names parameters. threshold_fields need to be exactly two, either category or binary. probabilities need to be exactly two, aligned with threshold_fields. model_names has to be exactly one. Three plots are produced.

The first plot shows several semi transparent lines. They summarize the 3d surfaces displayed by confidence_thresholding_2thresholds_3d that have thresholds on the confidence of the predictions of the two threshold_fields as x and y axes and either the data coverage percentage or the accuracy as z axis. Each line represents a slice of the data coverage surface projected onto the accuracy surface.

Confidence_Thresholding two thresholds 2D Multiline

The second plot shows the max of all the lines displayed in the first plot.

Confidence_Thresholding two thresholds 2D Maxline

The third plot shows the max line and the values of the thresholds that obtained a specific data coverage vs accuracy pair of values.

Confidence_Thresholding two thresholds 2D Accuracy and Thresholds

confidence_thresholding_2thresholds_3d

This visualization uses the ground_truth, threshold_fields, probabilities and model_names parameters. threshold_fields need to be exactly two, either category or binary. probabilities need to be exactly two, aligned with threshold_fields. model_names has to be exactly one. The plot shows the 3d surfaces displayed by confidence_thresholding_2thresholds_3d that have thresholds on the confidence of the predictions of the two threshold_fields as x and y axes and either the data coverage percentage or the accuracy as z axis.

Confidence_Thresholding two thresholds 3D

Binary Threshold vs. Metric

binary_threshold_vs_metric

This visualization uses the positive_label, metrics, ground_truth, ground_truth_metadata, field, probabilities and model_names parameters. field can be a category or binary feature. For each metric specified in metrics (options are f1, precision, recall, accuracy), this visualization produces a line chart plotting a threshold on the confidence of the model against the metric for the specified field. If field is a category feature, positive_label indicates which is the class to be considered positive class and all the others will be considered negative. It needs to be an integer, to figure out the association between classes and integers check the ground_truth_metadata JSON file.

Binary_Threshold_vs_Metric

ROC Curves

roc_curves

This visualization uses the positive_label, ground_truth, ground_truth_metadata, field, probabilities and model_names parameters. field can be a category or binary feature. This visualization produces a line chart plotting the roc curves for the specified field. If field is a category feature, positive_label indicates which is the class to be considered positive class and all the others will be considered negative. It needs to be an integer, to figure out the association between classes and integers check the ground_truth_metadata JSON file.

ROC Curves

roc_curves_from_test_statistics

This visualization uses the field, test_statistics and model_names parameters. field needs to be binary feature. This visualization produces a line chart plotting the roc curves for the specified field.

ROC Curves from Prediction Statistics

Calibration Plot

calibration_1_vs_all

This visualization uses the top_k, ground_truth, field, probabilities and model_names parameters. field needs to be a category or binary. For each class or each of the k most frequent classes if top_k is specified, it produces two plots computed on the fly from the probabilities of predictions for the specified field.

The first plot is a calibration curve that shows the calibration of the predictions considering the current class to be the true one and all others to be a false one, drawing one line for each model (in the aligned lists of probabilities and model_names).

Calibration 1 vs All Curve

The second plot shows the distributions of the predictions considering the current class to be the true one and all others to be a false one, drawing the distribution for each model (in the aligned lists of probabilities and model_names).

Calibration 1 vs All Counts

calibration_multiclass

This visualization uses the ground_truth, field, probabilities and model_names parameters. field needs to be a category. For each class, produces two plots computed on the fly from the probabilities of predictions for the specified field.

The first plot is a calibration curve that shows the calibration of the predictions considering al classes, drawing one line for each model (in the aligned lists of probabilities and model_names).

Calibration Multiclass Curve

The second plot shows a bar plot of the brier score (that calculates how calibrated are the probabilities of the predictions of a model), drawing one bar for each model (in the aligned lists of probabilities and model_names).

Calibration Multiclass Brier

Class Frequency vs. F1 score

frequency_vs_f1

This visualization uses the ground_truth_metadata, field, test_statistics and model_names parameters. field needs to be a category. For each model (in the aligned lists of test_statistics and model_names), produces two plots statistics of predictions for the specified field.

The first plot is a line plot with one x axis representing the different classes and two vertical axes colored in orange and blue respectively. The orange one is the frequency of the class and an orange line is plotted to show the trend. The blue one is the F1 score for that class and a blue line is plotted to show the trend. The classes on the x axis are sorted by f1 score.

Frequency vs F1 sorted by F1

The second plot has the same structure of the first one, but the axes are flipped and the classes on the x axis are sorted by frequency.

Frequency vs F1 sorted by Frequency