def initialize_model(self,
train_set_metadata=None,
train_set_metadata_json=None,
gpus=None,
gpu_fraction=1,
random_seed=default_random_seed,
debug=False,
**kwargs):
"""This function initializes a model. It is need for performing online
learning, so it has to be called before `train_online`.
`train` initialize the model under the hood, so there is no need to call
this function if you don't use `train_online`.
# Inputs
:param train_set_metadata: (dict) it contains metadata information for
the input and output features the model is going to be trained
on. It's the same content of the metadata json file that is
created while training.
:param train_set_metadata_json: (string) path to the JSON metadata file
created while training. it contains metadata information for the
input and output features the model is going to be trained on
:param gpus: (string, default: `None`) list of GPUs to use (it uses the
same syntax of CUDA_VISIBLE_DEVICES)
:param gpu_fraction: (float, default `1.0`) fraction of GPU memory to
initialize the process with
:param random_seed: (int, default`42`) 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
:param debug: (bool, default: `False`) enables debugging mode
"""
if train_set_metadata is None and train_set_metadata_json is None:
raise ValueError(
'train_set_metadata or train_set_metadata_json must not None.')
if train_set_metadata_json is not None:
train_set_metadata = load_metadata(train_set_metadata_json)
# update model definition with metadata properties
update_model_definition_with_metadata(self.model_definition,
train_set_metadata)
# build model
model = Model(self.model_definition['input_features'],
self.model_definition['output_features'],
self.model_definition['combiner'],
self.model_definition['training'],
self.model_definition['preprocessing'],
random_seed=random_seed,
debug=debug)
model.initialize_session(gpus=gpus, gpu_fraction=gpu_fraction)
# set parameters
self.model = model
self.train_set_metadata = train_set_metadata
def experiment(model_definition,
model_definition_file=None,
data_csv=None,
data_train_csv=None,
data_validation_csv=None,
data_test_csv=None,
data_hdf5=None,
data_train_hdf5=None,
data_validation_hdf5=None,
data_test_hdf5=None,
train_set_metadata_json=None,
experiment_name='experiment',
model_name='run',
model_load_path=None,
model_resume_path=None,
skip_save_model=False,
skip_save_progress=False,
skip_save_log=False,
skip_save_processed_input=False,
skip_save_unprocessed_output=False,
output_directory='results',
gpus=None,
gpu_fraction=1.0,
use_horovod=False,
random_seed=default_random_seed,
debug=False,
**kwargs):
"""Trains a model on a dataset's training and validation splits and
uses it to predict on the test split.
It saves the trained model and the statistics of training and testing.
:param model_definition: Model definition which defines the different
parameters of the model, features, preprocessing and training.
:type model_definition: Dictionary
:param model_definition_file: The file that specifies the model definition.
It is a yaml file.
:type model_definition_file: filepath (str)
:param data_csv: A CSV file contanining the input data which is used to
train, validate and test a model. The CSV either contains a
split column or will be split.
:type data_csv: filepath (str)
:param data_train_csv: A CSV file contanining the input data which is used
to train a model.
:type data_train_csv: filepath (str)
:param data_validation_csv: A CSV file contanining the input data which is used
to validate a model..
:type data_validation_csv: filepath (str)
:param data_test_csv: A CSV file contanining the input data which is used
to test a model.
:type data_test_csv: filepath (str)
:param data_hdf5: If the dataset is in the hdf5 format, this is used instead
of the csv file.
:type data_hdf5: filepath (str)
:param data_train_hdf5: If the training set is in the hdf5 format, this is
used instead of the csv file.
:type data_train_hdf5: filepath (str)
:param data_validation_hdf5: If the validation set is in the hdf5 format,
this is used instead of the csv file.
:type data_validation_hdf5: filepath (str)
:param data_test_hdf5: If the test set is in the hdf5 format, this is
used instead of the csv file.
:type data_test_hdf5: filepath (str)
:param train_set_metadata_json: If the dataset is in hdf5 format, this is
the associated json file containing metadata.
:type train_set_metadata_json: filepath (str)
:param experiment_name: The name for the experiment.
:type experiment_name: Str
:param model_name: Name of the model that is being used.
:type model_name: Str
:param model_load_path: If this is specified the loaded model will be used
as initialization (useful for transfer learning).
:type model_load_path: filepath (str)
:param model_resume_path: Resumes training of the model from the path
specified. The difference with model_load_path is that also training
statistics like the current epoch and the loss and performance so
far are also resumed effectively cotinuing a previously interrupted
training process.
:type model_resume_path: filepath (str)
:param skip_save_model: Disables
saving model weights and hyperparameters each time the model
improves. By default Ludwig saves model 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,
but the model will not be loadable later on.
:type skip_save_model: Boolean
:param skip_save_progress: Disables saving
progress each epoch. By default Ludwig saves weights and stats
after each epoch for enabling resuming of training, but if
the model is really big that can be time consuming and will uses
twice as much space, use this parameter to skip it, but training
cannot be resumed later on.
:type skip_save_progress: Boolean
:param 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..
:type skip_save_log: Boolean
:param skip_save_processed_input: If a CSV dataset is provided it is
preprocessed and then saved as an hdf5 and json to avoid running
the preprocessing again. If this parameter is False,
the hdf5 and json file are not saved.
:type skip_save_processed_input: Boolean
:param skip_save_unprocessed_output: By default predictions and
their probabilities are saved in both raw unprocessed numpy files
contaning tensors and as postprocessed CSV files
(one for each output feature). If this parameter is True,
only the CSV ones are saved and the numpy ones are skipped.
:type skip_save_unprocessed_output: Boolean
:param output_directory: The directory that will contanin the training
statistics, the saved model and the training procgress files.
:type output_directory: filepath (str)
:param gpus: List of GPUs that are available for training.
:type gpus: List
:param gpu_fraction: Fraction of the memory of each GPU to use at
the beginning of the training. The memory may grow elastically.
:type gpu_fraction: Integer
:param random_seed: Random seed used for weights initialization,
splits and any other random function.
:type random_seed: Integer
:param debug: If true turns on tfdbg with inf_or_nan checks.
:type debug: Boolean
"""
# set input features defaults
if model_definition_file is not None:
with open(model_definition_file, 'r') as def_file:
model_definition = merge_with_defaults(yaml.load(def_file))
else:
model_definition = merge_with_defaults(model_definition)
# setup directories and file names
experiment_dir_name = None
if model_resume_path is not None:
if os.path.exists(model_resume_path):
experiment_dir_name = model_resume_path
else:
if is_on_master():
logging.info('Model resume path does not exists, '
'starting training from scratch')
model_resume_path = None
if model_resume_path is None:
if is_on_master():
experiment_dir_name = get_experiment_dir_name(
output_directory, experiment_name, model_name)
else:
experiment_dir_name = '/'
description_fn, training_stats_fn, model_dir = get_file_names(
experiment_dir_name)
# save description
description = get_experiment_description(
model_definition, data_csv, data_train_csv, data_validation_csv,
data_test_csv, data_hdf5, data_train_hdf5, data_validation_hdf5,
data_test_hdf5, train_set_metadata_json, random_seed)
if is_on_master():
save_json(description_fn, description)
# print description
logging.info('Experiment name: {}'.format(experiment_name))
logging.info('Model name: {}'.format(model_name))
logging.info('Output path: {}'.format(experiment_dir_name))
logging.info('')
for key, value in description.items():
logging.info('{}: {}'.format(key, pformat(value, indent=4)))
logging.info('')
# preprocess
(training_set, validation_set, test_set,
train_set_metadata) = preprocess_for_training(
model_definition,
data_csv=data_csv,
data_train_csv=data_train_csv,
data_validation_csv=data_validation_csv,
data_test_csv=data_test_csv,
data_hdf5=data_hdf5,
data_train_hdf5=data_train_hdf5,
data_validation_hdf5=data_validation_hdf5,
data_test_hdf5=data_test_hdf5,
train_set_metadata_json=train_set_metadata_json,
skip_save_processed_input=skip_save_processed_input,
preprocessing_params=model_definition['preprocessing'],
random_seed=random_seed)
if is_on_master():
logging.info('Training set: {0}'.format(training_set.size))
if validation_set is not None:
logging.info('Validation set: {0}'.format(validation_set.size))
if test_set is not None:
logging.info('Test set: {0}'.format(test_set.size))
# update model definition with metadata properties
update_model_definition_with_metadata(model_definition, train_set_metadata)
# run the experiment
model, training_results = train(training_set=training_set,
validation_set=validation_set,
test_set=test_set,
model_definition=model_definition,
save_path=model_dir,
model_load_path=model_load_path,
resume=model_resume_path is not None,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
gpus=gpus,
gpu_fraction=gpu_fraction,
use_horovod=use_horovod,
random_seed=random_seed,
debug=debug)
(train_trainset_stats, train_valisest_stats,
train_testset_stats) = training_results
if is_on_master():
if not skip_save_model:
# save train set metadata
save_json(os.path.join(model_dir, TRAIN_SET_METADATA_FILE_NAME),
train_set_metadata)
# grab the results of the model with highest validation test performance
validation_field = model_definition['training']['validation_field']
validation_measure = model_definition['training']['validation_measure']
validation_field_result = train_valisest_stats[validation_field]
best_function = get_best_function(validation_measure)
# print results of the model with highest validation test performance
if is_on_master():
if validation_set is not None:
# max or min depending on the measure
epoch_best_vali_measure, best_vali_measure = best_function(
enumerate(validation_field_result[validation_measure]),
key=lambda pair: pair[1])
logging.info('Best validation model epoch: {0}'.format(
epoch_best_vali_measure + 1))
logging.info(
'Best validation model {0} on validation set {1}: {2}'.format(
validation_measure, validation_field, best_vali_measure))
if test_set is not None:
best_vali_measure_epoch_test_measure = train_testset_stats[
validation_field][validation_measure][
epoch_best_vali_measure]
logging.info(
'Best validation model {0} on test set {1}: {2}'.format(
validation_measure, validation_field,
best_vali_measure_epoch_test_measure))
# save training statistics
if is_on_master():
save_json(
training_stats_fn, {
'train': train_trainset_stats,
'validation': train_valisest_stats,
'test': train_testset_stats
})
if test_set is not None:
# predict
test_results = predict(test_set,
train_set_metadata,
model,
model_definition,
model_definition['training']['batch_size'],
only_predictions=False,
gpus=gpus,
gpu_fraction=gpu_fraction,
debug=debug)
# postprocess
postprocessed_output = postprocess(
test_results, model_definition['output_features'],
train_set_metadata, experiment_dir_name,
skip_save_unprocessed_output or not is_on_master())
if is_on_master():
print_prediction_results(test_results)
save_prediction_outputs(postprocessed_output, experiment_dir_name)
save_prediction_statistics(test_results, experiment_dir_name)
model.close_session()
if is_on_master():
logging.info('\nFinished: {0}_{1}'.format(experiment_name, model_name))
logging.info('Saved to: {}'.format(experiment_dir_name))
return experiment_dir_name
def train(self,
data_df=None,
data_train_df=None,
data_validation_df=None,
data_test_df=None,
data_csv=None,
data_train_csv=None,
data_validation_csv=None,
data_test_csv=None,
data_hdf5=None,
data_train_hdf5=None,
data_validation_hdf5=None,
data_test_hdf5=None,
train_set_metadata_json=None,
dataset_type='generic',
model_name='run',
model_load_path=None,
model_resume_path=None,
skip_save_model=False,
skip_save_progress=False,
skip_save_log=False,
skip_save_processed_input=False,
output_directory='results',
gpus=None,
gpu_fraction=1.0,
random_seed=42,
logging_level=logging.ERROR,
debug=False,
**kwargs):
"""This function is used to perform a full training of the model on the
specified dataset.
# Inputs
:param data_df: (DataFrame) dataframe containing data. 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
:param data_train_df: (DataFrame) dataframe containing training data
:param data_validation_df: (DataFrame) dataframe containing validation
data
:param data_test_df: (DataFrame dataframe containing test data
:param data_csv: (string) 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
:param data_train_csv: (string) input train data CSV file
:param data_validation_csv: (string) input validation data CSV file
:param data_test_csv: (string) input test data CSV file
:param data_hdf5: (string) 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
:param data_train_hdf5: (string) 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
:param data_validation_hdf5: (string) 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
:param data_test_hdf5: (string) 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
:param train_set_metadata_json: (string) 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
:param dataset_type: (string, default: `'default'`) determines the type
of preprocessing will be applied to the data. Only `generic` is
available at the moment
:param model_name: (string) a name for the model, user for the save
directory
:param model_load_path: (string) path of a pretrained model to load as
initialization
:param model_resume_path: (string) path of a the model directory to
resume training of
:param skip_save_model: (bool, default: `False`) disables
saving model weights and hyperparameters each time the model
improves. By default Ludwig saves model 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,
but the model will not be loadable later on.
:param skip_save_progress: (bool, default: `False`) disables saving
progress each epoch. By default Ludwig saves weights and stats
after each epoch for enabling resuming of training, but if
the model is really big that can be time consuming and will uses
twice as much space, use this parameter to skip it, but training
cannot be resumed later on.
:param skip_save_log: (bool, default: `False`) 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.
:param skip_save_processed_input: (bool, default: `False`) skips saving
intermediate HDF5 and JSON files
:param output_directory: (string, default: `'results'`) directory that
contains the results
:param gpus: (string, default: `None`) list of GPUs to use (it uses the
same syntax of CUDA_VISIBLE_DEVICES)
:param gpu_fraction: (float, default `1.0`) fraction of gpu memory to
initialize the process with
:param random_seed: (int, default`42`) 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
:param debug: (bool, default: `False`) enables debugging mode
:param logging_level: (int, default: `logging.ERROR`) logging level to
use for logging. Use logging constants like `logging.DEBUG`,
`logging.INFO` and `logging.ERROR`. By default only errors will
be printed.
There are three ways to provide data: by dataframes using the `_df`
parameters, by CSV using the `_csv` parameters and by HDF5 and JSON,
using `_hdf5` and `_json` parameters.
The DataFrame approach uses data previously obtained and put in a
dataframe, the CSV approach loads data from a CSV file, while HDF5 and
JSON load previously preprocessed HDF5 and JSON files (they are saved in
the same directory of the CSV they are obtained from).
For all three approaches either a full dataset can be provided (which
will be split randomly according to the split probabilities defined in
the model definition, by default 70% training, 10% validation and 20%
test) or, if it contanins a plit column, it will be plit according to
that column (interpreting 0 as training, 1 as validation and 2 as test).
Alternatively separated dataframes / CSV / HDF5 files can beprovided
for each split.
During training the model and statistics will be saved in a directory
`[output_dir]/[experiment_name]_[model_name]_n` where all variables are
resolved to user spiecified ones and `n` is an increasing number
starting from 0 used to differentiate different runs.
# Return
:return: (dict) a dictionary containing training statistics for each
output feature containing loss and measures values for each epoch.
"""
logging.getLogger().setLevel(logging_level)
if logging_level in {logging.WARNING, logging.ERROR, logging.CRITICAL}:
set_disable_progressbar(True)
# setup directories and file names
experiment_dir_name = None
if model_resume_path is not None:
if os.path.exists(model_resume_path):
experiment_dir_name = model_resume_path
else:
logging.info('Model resume path does not exists,'
' starting training from scratch')
model_resume_path = None
if model_resume_path is None:
experiment_dir_name = get_experiment_dir_name(
output_directory, '', model_name)
description_fn, training_stats_fn, model_dir = get_file_names(
experiment_dir_name)
# save description
description = get_experiment_description(
self.model_definition,
dataset_type,
data_csv=data_csv,
data_train_csv=data_train_csv,
data_validation_csv=data_validation_csv,
data_test_csv=data_test_csv,
data_hdf5=data_hdf5,
data_train_hdf5=data_train_hdf5,
data_validation_hdf5=data_validation_hdf5,
data_test_hdf5=data_test_hdf5,
metadata_json=train_set_metadata_json,
random_seed=random_seed)
save_json(description_fn, description)
# print description
logging.info('Model name: {}'.format(model_name))
logging.info('Output path: {}'.format(experiment_dir_name))
logging.info('\n')
for key, value in description.items():
logging.info('{0}: {1}'.format(key, pformat(value, indent=4)))
logging.info('\n')
# preprocess
if data_df is not None or data_train_df is not None:
(training_set, validation_set, test_set,
train_set_metadata) = preprocess_for_training(
self.model_definition,
dataset_type,
data_df=data_df,
data_train_df=data_train_df,
data_validation_df=data_validation_df,
data_test_df=data_test_df,
train_set_metadata_json=train_set_metadata_json,
skip_save_processed_input=True,
preprocessing_params=self.model_definition['preprocessing'],
random_seed=random_seed)
else:
(training_set, validation_set, test_set,
train_set_metadata) = preprocess_for_training(
self.model_definition,
dataset_type,
data_csv=data_csv,
data_train_csv=data_train_csv,
data_validation_csv=data_validation_csv,
data_test_csv=data_test_csv,
data_hdf5=data_hdf5,
data_train_hdf5=data_train_hdf5,
data_validation_hdf5=data_validation_hdf5,
data_test_hdf5=data_test_hdf5,
train_set_metadata_json=train_set_metadata_json,
skip_save_processed_input=skip_save_processed_input,
preprocessing_params=self.model_definition['preprocessing'],
random_seed=random_seed)
logging.info('Training set: {0}'.format(training_set.size))
if validation_set is not None:
logging.info('Validation set: {0}'.format(validation_set.size))
if test_set is not None:
logging.info('Test set: {0}'.format(test_set.size))
# update model definition with metadata properties
update_model_definition_with_metadata(self.model_definition,
train_set_metadata)
if not skip_save_model:
os.makedirs(model_dir, exist_ok=True)
train_set_metadata_path = os.path.join(
model_dir, TRAIN_SET_METADATA_FILE_NAME)
save_json(train_set_metadata_path, train_set_metadata)
# run the experiment
model, result = train(training_set=training_set,
validation_set=validation_set,
test_set=test_set,
model_definition=self.model_definition,
save_path=model_dir,
model_load_path=model_load_path,
resume=model_resume_path is not None,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
gpus=gpus,
gpu_fraction=gpu_fraction,
random_seed=random_seed,
debug=debug)
train_trainset_stats, train_valisest_stats, train_testset_stats = result
train_stats = {
'train': train_trainset_stats,
'validation': train_valisest_stats,
'test': train_testset_stats
}
# save training and test statistics
save_json(training_stats_fn, train_stats)
# grab the results of the model with highest validation test performance
md_training = self.model_definition['training']
validation_field = md_training['validation_field']
validation_measure = md_training['validation_measure']
validation_field_result = train_valisest_stats[validation_field]
best_function = get_best_function(validation_measure)
# print results of the model with highest validation test performance
if validation_set is not None:
# max or min depending on the measure
epoch_best_vali_measure, best_vali_measure = best_function(
enumerate(validation_field_result[validation_measure]),
key=lambda pair: pair[1])
logging.info('Best validation model epoch: {0}'.format(
epoch_best_vali_measure + 1))
logging.info(
'Best validation model {0} on validation set {1}: {2}'.format(
validation_measure, validation_field, best_vali_measure))
if test_set is not None:
best_vali_measure_epoch_test_measure = train_testset_stats[
validation_field][validation_measure][
epoch_best_vali_measure]
logging.info(
'Best validation model {0} on test set {1}: {2}'.format(
validation_measure, validation_field,
best_vali_measure_epoch_test_measure))
logging.info('Finished: {0}'.format(model_name))
logging.info('Saved to {0}:'.format(experiment_dir_name))
# set parameters
self.model = model
self.train_set_metadata = train_set_metadata
return train_stats
def initialize_model(self,
train_set_metadata=None,
train_set_metadata_json=None,
gpus=None,
gpu_memory_limit=None,
allow_parallel_threads=True,
random_seed=default_random_seed,
debug=False,
**kwargs):
"""This function initializes a model. It is need for performing online
learning, so it has to be called before `train_online`.
`train` initialize the model under the hood, so there is no need to call
this function if you don't use `train_online`.
# Inputs
:param train_set_metadata: (dict) it contains metadata information for
the input and output features the model is going to be trained
on. It's the same content of the metadata json file that is
created while training.
:param train_set_metadata_json: (string) path to the JSON metadata file
created while training. it contains metadata information for the
input and output features the model is going to be trained on
:param gpus: (string, default: `None`) list of GPUs to use (it uses the
same syntax of CUDA_VISIBLE_DEVICES)
:param gpu_memory_limit: (int: default: `None`) maximum memory in MB to allocate
per GPU device.
:param allow_parallel_threads: (bool, default: `True`) allow TensorFlow to use
multithreading parallelism to improve performance at the cost of
determinism.
:param random_seed: (int, default`42`) 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
:param debug: (bool, default: `False`) enables debugging mode
"""
if train_set_metadata is None and train_set_metadata_json is None:
raise ValueError(
'train_set_metadata or train_set_metadata_json must not None.')
if train_set_metadata_json is not None:
train_set_metadata = load_metadata(train_set_metadata_json)
# update model definition with metadata properties
update_model_definition_with_metadata(self.model_definition,
train_set_metadata)
# build model
model = Trainer(self.model_definition['input_features'],
self.model_definition['output_features'],
self.model_definition['combiner'],
self.model_definition[TRAINING],
self.model_definition['preprocessing'],
gpus=gpus,
gpu_memory_limit=gpu_memory_limit,
allow_parallel_threads=allow_parallel_threads,
random_seed=random_seed,
debug=debug)
# set parameters
self.model = model
self.train_set_metadata = train_set_metadata
