def put(self, training_set, test_set, validation_set,
training_set_metadata):
logger.info("Writing preprocessed training set cache")
training_set = self.dataset_manager.save(
self.cache_map[TRAINING],
training_set,
self.config,
training_set_metadata,
TRAINING,
)
if test_set is not None:
logger.info("Writing preprocessed test set cache")
test_set = self.dataset_manager.save(
self.cache_map[TEST],
test_set,
self.config,
training_set_metadata,
TEST,
)
if validation_set is not None:
logger.info("Writing preprocessed validation set cache")
validation_set = self.dataset_manager.save(
self.cache_map[VALIDATION],
validation_set,
self.config,
training_set_metadata,
VALIDATION,
)
logger.info("Writing train set metadata")
data_utils.save_json(self.cache_map[META], training_set_metadata)
return training_set, test_set, validation_set, training_set_metadata
def create_metrics_report(experiment_name: str) -> Tuple[Dict[str, Any], str]:
"""Compiles performance and non-performance metrics.
`experiment_name`: name referring to the experiment.
Returns a full report and the path where it's saved.
"""
full_report = dict()
os.makedirs(os.path.join(os.getcwd(), experiment_name, "metrics_report"), exist_ok=True)
for tag in [TRAIN_TAG, EVAL_TAG]:
if tag == TRAIN_TAG:
resource_usage_path = os.path.join(os.getcwd(), experiment_name, CACHE, "train_resource_usage_metrics.json")
performance_path = os.path.join(os.getcwd(), experiment_name, EXPERIMENT_RUN, "training_statistics.json")
elif tag == EVAL_TAG:
resource_usage_path = os.path.join(
os.getcwd(), experiment_name, CACHE, "evaluate_resource_usage_metrics.json"
)
performance_path = os.path.join(os.getcwd(), experiment_name, EXPERIMENT_RUN, "test_statistics.json")
else:
raise ValueError("Tag unrecognized. Please choose 'train' or 'evaluate'.")
resource_usage_metrics = load_json(resource_usage_path)
performance_metrics = load_json(performance_path)
full_report[tag] = merge_dict(performance_metrics, resource_usage_metrics)
merged_file_path = os.path.join(os.getcwd(), experiment_name, "metrics_report", "{}.json".format("full_report"))
save_json(merged_file_path, full_report)
return full_report, merged_file_path
def save(self, save_path):
"""This function allows to save models on disk
# Inputs
:param save_path: (string) path to the directory where the model is
going to be saved. Both a JSON file containing the model
architecture hyperparameters and checkpoints files containing
model weights will be saved.
# Example usage
```python
ludwig_model.save(save_path)
```
"""
self._check_initialization()
# save model definition
self.save_model_definition(save_path)
# save model weights
model_weights_path = os.path.join(save_path, MODEL_WEIGHTS_FILE_NAME)
self.model.save_weights(model_weights_path)
# save training set metadata
training_set_metadata_path = os.path.join(
save_path,
TRAIN_SET_METADATA_FILE_NAME
)
save_json(training_set_metadata_path, self.training_set_metadata)
def save_model_definition(self, save_path):
os.makedirs(save_path, exist_ok=True)
model_hyperparameters_path = os.path.join(
save_path,
MODEL_HYPERPARAMETERS_FILE_NAME
)
save_json(model_hyperparameters_path, self.model_definition)
def save(self, save_path):
"""This function allows to save models on disk
# Inputs
:param save_path: (string) path to the directory where the model is
going to be saved. Both a JSON file containing the model
architecture hyperparameters and checkpoints files containing
model weights will be saved.
# Example usage
```python
ludwig_model.save(save_path)
```
"""
if self.model is None or self.model_definition is None or self.train_set_metadata is None:
raise ValueError('Model has not been initialized or loaded')
model_weights_path = os.path.join(save_path, MODEL_WEIGHTS_FILE_NAME)
model_hyperparameters_path = os.path.join(
save_path, MODEL_HYPERPARAMETERS_FILE_NAME)
self.model.model.save_weights(model_weights_path)
train_set_metadata_path = os.path.join(save_path,
TRAIN_SET_METADATA_FILE_NAME)
save_json(train_set_metadata_path, self.train_set_metadata)
self.model.save_hyperparameters(self.model._hyperparameters,
model_hyperparameters_path)
def kfold_cross_validate_cli(
k_fold,
config=None,
config_file=None,
dataset=None,
data_format=None,
output_directory='results',
random_seed=default_random_seed,
skip_save_k_fold_split_indices=False,
**kwargs
):
"""Wrapper function to performs k-fold cross validation.
# Inputs
:param k_fold: (int) number of folds to create for the cross-validation
:param config: (dict, default: None) a dictionary containing
information needed to build a model. Refer to the [User Guide]
(http://ludwig.ai/user_guide/#model-config) for details.
:param config_file: (string, optional, default: `None`) path to
a YAML file containing the config. If available it will be
used instead of the config dict.
:param data_csv: (string, default: None)
:param output_directory: (string, default: 'results')
:param random_seed: (int) Random seed used k-fold splits.
:param skip_save_k_fold_split_indices: (boolean, default: False) Disables
saving k-fold split indices
:return: None
"""
if config is None and config_file is None:
raise ValueError(
"No config is provided 'config' or "
"'config_file' must be provided."
)
elif config is not None and config_file is not None:
raise ValueError(
"Cannot specify both 'config' and 'config_file'"
", proivde only one of the parameters."
)
(kfold_cv_stats,
kfold_split_indices) = kfold_cross_validate(
k_fold,
config=config if config is not None else
config_file,
dataset=dataset,
data_format=data_format,
output_directory=output_directory,
random_seed=random_seed
)
# save k-fold cv statistics
save_json(os.path.join(output_directory, 'kfold_training_statistics.json'),
kfold_cv_stats)
# save k-fold split indices
if not skip_save_k_fold_split_indices:
save_json(os.path.join(output_directory, 'kfold_split_indices.json'),
kfold_split_indices)
def save_prediction_outputs(
postprocessed_output,
output_directory,
backend,
):
postprocessed_output, column_shapes = flatten_df(postprocessed_output, backend)
postprocessed_output.to_parquet(os.path.join(output_directory, "predictions.parquet"))
save_json(os.path.join(output_directory, "predictions.shapes.json"), column_shapes)
def on_trainer_train_setup(self, trainer, save_path, is_coordinator):
if not is_coordinator:
return
# When running on a remote worker, the model metadata files will only have been
# saved to the driver process, so re-save it here before uploading.
training_set_metadata_path = os.path.join(
save_path, TRAIN_SET_METADATA_FILE_NAME)
if not os.path.exists(training_set_metadata_path):
save_json(training_set_metadata_path, self.training_set_metadata)
model_hyperparameters_path = os.path.join(
save_path, MODEL_HYPERPARAMETERS_FILE_NAME)
if not os.path.exists(model_hyperparameters_path):
save_json(model_hyperparameters_path, self.config)
def save_prediction_outputs(
postprocessed_output,
output_features,
output_directory,
backend,
):
postprocessed_output, column_shapes = flatten_df(postprocessed_output, backend)
postprocessed_output.to_parquet(os.path.join(output_directory, PREDICTIONS_PARQUET_FILE_NAME))
save_json(os.path.join(output_directory, PREDICTIONS_SHAPES_FILE_NAME), column_shapes)
if not backend.df_engine.partitioned:
# csv can only be written out for unpartitioned df format (i.e., pandas)
postprocessed_dict = convert_to_dict(postprocessed_output, output_features)
csv_filename = os.path.join(output_directory, "{}_{}.csv")
for output_field, outputs in postprocessed_dict.items():
for output_type, values in outputs.items():
save_csv(csv_filename.format(output_field, output_type), values)
def __exit__(self, exc_type, exc_val, exc_tb) -> None:
"""Waits for monitoring process to exit.
Computes and postprocesses more metrics. Saves report.
"""
self.queue.put(STOP_MESSAGE)
if torch.cuda.is_available():
torch.cuda.synchronize()
self.p.join()
self.info = load_json(
os.path.join(self.output_dir, self.info["tag"] + "_temp.json"))
os.remove(
os.path.join(self.output_dir, self.info["tag"] + "_temp.json"))
self.info["end_time"] = time.time()
self.info[f"{self.tag}_total_duration"] = self.info[
"end_time"] - self.info["start_time"]
if self.num_examples:
self.info["examples_per_second"] = self.num_examples / self.info[
f"{self.tag}_total_duration"]
self.info["end_disk_usage"] = shutil.disk_usage(
os.path.expanduser("~")).used
self.info["disk_footprint"] = self.info["end_disk_usage"] - self.info[
"start_disk_usage"]
for key in self.info["system"]:
if "gpu_" in key:
self.info["system"][key]["max_memory_used"] = max(
self.info["system"][key]["memory_used"])
self.info["system"]["max_cpu_utilization"] = max(
self.info["system"]["cpu_utilization"], default=None)
self.info["system"]["max_ram_utilization"] = max(
self.info["system"]["ram_utilization"], default=None)
if self.info["system"]["cpu_utilization"]:
self.info["system"]["average_cpu_utilization"] = mean(
self.info["system"]["cpu_utilization"])
if self.info["system"]["ram_utilization"]:
self.info["system"]["average_ram_utilization"] = mean(
self.info["system"]["ram_utilization"])
save_json(
os.path.join(self.output_dir,
self.info["tag"] + "_resource_usage_metrics.json"),
self.info)
def put_dataset(self, input_fname, config, processed,
skip_save_processed_input):
if not self.can_cache(input_fname, config, skip_save_processed_input):
return processed
training_set, test_set, validation_set, training_set_metadata = processed
key = training_set_metadata.get(CHECKSUM)
if not key:
key = self.get_cache_key(input_fname, config)
training_set_metadata[CHECKSUM] = key
logger.info('Writing preprocessed training set cache')
training_set = self.save(
self.get_cache_path(input_fname, key, TRAINING),
training_set,
config,
training_set_metadata,
TRAINING,
)
if test_set is not None:
logger.info('Writing preprocessed test set cache')
test_set = self.save(
self.get_cache_path(input_fname, key, TEST),
test_set,
config,
training_set_metadata,
TEST,
)
if validation_set is not None:
logger.info('Writing preprocessed validation set cache')
validation_set = self.save(
self.get_cache_path(input_fname, key, VALIDATION),
validation_set,
config,
training_set_metadata,
VALIDATION,
)
logger.info('Writing train set metadata')
data_utils.save_json(
self.get_cache_path(input_fname, key, 'meta', 'json'),
training_set_metadata)
return training_set, test_set, validation_set, training_set_metadata
def kfold_cross_validate_cli(k_fold,
model_definition=None,
model_definition_file=None,
data_csv=None,
output_directory='results',
random_seed=default_random_seed,
skip_save_k_fold_split_indices=False,
**kwargs):
"""Wrapper function to performs k-fold cross validation.
# Inputs
:param k_fold: (int) number of folds to create for the cross-validation
:param model_definition: (dict, default: None) a dictionary containing
information needed to build a model. Refer to the [User Guide]
(http://ludwig.ai/user_guide/#model-definition) for details.
:param model_definition_file: (string, optional, default: `None`) path to
a YAML file containing the model definition. If available it will be
used instead of the model_definition dict.
:param data_csv: (string, default: None)
:param output_directory: (string, default: 'results')
:param random_seed: (int) Random seed used k-fold splits.
:param skip_save_k_fold_split_indices: (boolean, default: False) Disables
saving k-fold split indices
:return: None
"""
model_definition = check_which_model_definition(model_definition,
model_definition_file)
(kfold_cv_stats, kfold_split_indices) = kfold_cross_validate(
k_fold,
model_definition=model_definition,
data_csv=data_csv,
output_directory=output_directory,
random_seed=random_seed)
# save k-fold cv statistics
save_json(os.path.join(output_directory, 'kfold_training_statistics.json'),
kfold_cv_stats)
# save k-fold split indices
if not skip_save_k_fold_split_indices:
save_json(os.path.join(output_directory, 'kfold_split_indices.json'),
kfold_split_indices)
def kfold_cross_validate_cli(
k_fold,
config=None,
dataset=None,
data_format=None,
output_directory="results",
random_seed=default_random_seed,
skip_save_k_fold_split_indices=False,
**kwargs,
):
"""Wrapper function to performs k-fold cross validation.
# Inputs
:param k_fold: (int) number of folds to create for the cross-validation
:param config: (Union[str, dict], default: None) a dictionary or file path
containing model configuration. Refer to the [User Guide]
(http://ludwig.ai/user_guide/#model-config) for details.
:param dataset: (string, default: None)
:param output_directory: (string, default: 'results')
:param random_seed: (int) Random seed used k-fold splits.
:param skip_save_k_fold_split_indices: (boolean, default: False) Disables
saving k-fold split indices
:return: None
"""
(kfold_cv_stats, kfold_split_indices) = kfold_cross_validate(
k_fold,
config=config,
dataset=dataset,
data_format=data_format,
output_directory=output_directory,
random_seed=random_seed,
)
# save k-fold cv statistics
save_json(os.path.join(output_directory, "kfold_training_statistics.json"),
kfold_cv_stats)
# save k-fold split indices
if not skip_save_k_fold_split_indices:
save_json(os.path.join(output_directory, "kfold_split_indices.json"),
kfold_split_indices)
def monitor(queue: multiprocessing.Queue, info: Dict[str, Any],
output_dir: str, logging_interval: int) -> None:
"""Monitors hardware resource use as part of a separate process.
Populate `info` with system specific metrics (GPU, CPU, RAM) at a `logging_interval` interval and saves the output
in `output_dir`.
Args:
queue: queue from which we can push and retrieve messages sent to the child process.
info: dictionary containing system resource usage information about the parent process.
output_dir: directory where the contents of `info` will be saved.
logging_interval: time interval at which we will poll the system for usage metrics.
"""
for key in info["system"]:
if "gpu_" in key:
info["system"][key]["memory_used"] = []
info["system"]["cpu_utilization"] = []
info["system"]["ram_utilization"] = []
while True:
try:
message = queue.get(block=False)
if isinstance(message, str):
if message == STOP_MESSAGE:
save_json(
os.path.join(output_dir, info["tag"] + "_temp.json"),
info)
return
else:
queue.put(message)
except EmptyQueueException:
pass
if torch.cuda.is_available():
gpu_infos = GPUStatCollection.new_query()
for i, gpu_info in enumerate(gpu_infos):
gpu_key = f"gpu_{i}"
info["system"][gpu_key]["memory_used"].append(
gpu_info.memory_used)
info["system"]["cpu_utilization"].append(psutil.cpu_percent())
info["system"]["ram_utilization"].append(
psutil.virtual_memory().percent)
time.sleep(logging_interval)
def on_trainer_train_setup(self, trainer, save_path, is_coordinator):
if not is_coordinator:
return
# When running on a remote worker, the model metadata files will only have been
# saved to the driver process, so re-save it here before uploading.
training_set_metadata_path = os.path.join(save_path, TRAIN_SET_METADATA_FILE_NAME)
if not os.path.exists(training_set_metadata_path):
save_json(training_set_metadata_path, self.training_set_metadata)
model_hyperparameters_path = os.path.join(save_path, MODEL_HYPERPARAMETERS_FILE_NAME)
if not os.path.exists(model_hyperparameters_path):
save_json(model_hyperparameters_path, self.config)
if self.save_in_background:
save_queue = queue.Queue()
self.save_fn = lambda args: save_queue.put(args)
self.save_thread = threading.Thread(target=_log_mlflow_loop, args=(save_queue,))
self.save_thread.start()
else:
self.save_fn = lambda args: _log_mlflow(*args)
def full_train(model_definition,
model_definition_file=None,
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,
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,
output_directory='results',
should_close_session=True,
gpus=None,
gpu_fraction=1.0,
use_horovod=False,
random_seed=42,
debug=False,
**kwargs):
"""*full_train* defines the entire training procedure used by Ludwig's
internals. Requires most of the parameters that are taken into the model.
Builds a full ludwig model and performs the training.
:param data_test_df:
:param data_df:
:param data_train_df:
:param data_validation_df:
: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_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_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_progress: 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
:returns: None
"""
# 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.safe_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():
logger.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 = '.'
# if model_load_path is not None, load its train_set_metadata
if model_load_path is not None:
train_set_metadata_json = os.path.join(model_load_path,
TRAIN_SET_METADATA_FILE_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_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)
if is_on_master():
save_json(description_fn, description)
# print description
logger.info('Experiment name: {}'.format(experiment_name))
logger.info('Model name: {}'.format(model_name))
logger.info('Output path: {}'.format(experiment_dir_name))
logger.info('\n')
for key, value in description.items():
logger.info('{}: {}'.format(key, pformat(value, indent=4)))
logger.info('\n')
# preprocess
preprocessed_data = preprocess_for_training(
model_definition,
data_df=data_df,
data_train_df=data_train_df,
data_validation_df=data_validation_df,
data_test_df=data_test_df,
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)
(training_set, validation_set, test_set,
train_set_metadata) = preprocessed_data
if is_on_master():
logger.info('Training set: {0}'.format(training_set.size))
if validation_set is not None:
logger.info('Validation set: {0}'.format(validation_set.size))
if test_set is not None:
logger.info('Test set: {0}'.format(test_set.size))
# update model definition with metadata properties
update_model_definition_with_metadata(model_definition, train_set_metadata)
if is_on_master():
if not skip_save_model:
# save train set metadata
os.makedirs(model_dir, exist_ok=True)
save_json(os.path.join(model_dir, TRAIN_SET_METADATA_FILE_NAME),
train_set_metadata)
# run the experiment
model, result = 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 = result
train_stats = {
'train': train_trainset_stats,
'validation': train_valisest_stats,
'test': train_testset_stats
}
if should_close_session:
model.close_session()
if is_on_master():
# save training and test statistics
save_json(training_stats_fn, train_stats)
# 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)
# results of the model with highest validation test performance
if is_on_master() and validation_set is not None:
epoch_best_vali_measure, best_vali_measure = best_function(
enumerate(validation_field_result[validation_measure]),
key=lambda pair: pair[1])
logger.info(
'Best validation model epoch: {0}'.format(epoch_best_vali_measure +
1))
logger.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]
logger.info(
'Best validation model {0} on test set {1}: {2}'.format(
validation_measure, validation_field,
best_vali_measure_epoch_test_measure))
logger.info('\nFinished: {0}_{1}'.format(experiment_name, model_name))
logger.info('Saved to: {0}'.format(experiment_dir_name))
contrib_command("train_save", experiment_dir_name)
return (model, preprocessed_data, experiment_dir_name, train_stats,
model_definition)
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 save_hyperopt_stats(hyperopt_stats, hyperopt_dir_name):
hyperopt_stats_fn = os.path.join(hyperopt_dir_name,
'hyperopt_statistics.json')
save_json(hyperopt_stats_fn, hyperopt_stats)
def save_hyperopt_stats(hyperopt_stats, hyperopt_dir_name):
hyperopt_stats_fn = os.path.join(hyperopt_dir_name,
HYPEROPT_STATISTICS_FILE_NAME)
save_json(hyperopt_stats_fn, hyperopt_stats)
def write_meta(df, path, compression_type):
meta = {'size': len(df.index), 'compression_type': compression_type or ''}
save_json(os.path.join(path, "meta.json"), meta)
def kfold_cross_validate(k_fold,
model_definition=None,
model_definition_file=None,
data_csv=None,
output_directory='results',
random_seed=default_random_seed,
skip_save_k_fold_split_indices=False,
**kwargs):
"""Performs k-fold cross validation.
# Inputs
:param k_fold: (int) number of folds to create for the cross-validation
:param model_definition: (dict, default: None) a dictionary containing
information needed to build a model. Refer to the [User Guide]
(http://ludwig.ai/user_guide/#model-definition) for details.
:param model_definition_file: (string, optional, default: `None`) path to
a YAML file containing the model definition. If available it will be
used instead of the model_definition dict.
:param data_csv: (string, default: None)
:param output_directory: (string, default: 'results')
:param random_seed: (int) Random seed used k-fold splits.
:param skip_save_k_fold_split_indices: (boolean, default: False) Disables
saving k-fold split indices
:return: None
"""
# check for model_definition and model_definition_file
if model_definition is None and model_definition_file is None:
raise ValueError(
'Either model_definition of model_definition_file have to be'
'not None to initialize a LudwigModel')
if model_definition is not None and model_definition_file is not None:
raise ValueError('Only one between model_definition and '
'model_definition_file can be provided')
# check for k_fold
if k_fold is None:
raise ValueError('k_fold parameter must be specified')
logger.info('starting {:d}-fold cross validation'.format(k_fold))
# create output_directory if not available
if not os.path.isdir(output_directory):
os.mkdir(output_directory)
# read in data to split for the folds
data_df = pd.read_csv(data_csv)
# place each fold in a separate directory
data_dir = os.path.dirname(data_csv)
kfold_training_stats = {}
kfold_split_indices = {}
for train_indices, test_indices, fold_num in \
generate_kfold_splits(data_df, k_fold, random_seed):
with tempfile.TemporaryDirectory(dir=data_dir) as temp_dir_name:
curr_train_df = data_df.iloc[train_indices]
curr_test_df = data_df.iloc[test_indices]
if not skip_save_k_fold_split_indices:
kfold_split_indices['fold_' + str(fold_num)] = {
'training_indices': train_indices,
'test_indices': test_indices
}
# train and validate model on this fold
if model_definition_file is not None:
with open(model_definition_file, 'r') as def_file:
model_definition = \
merge_with_defaults(yaml.safe_load(def_file))
logger.info("training on fold {:d}".format(fold_num))
(model, preprocessed_data, _, train_stats,
model_definition) = full_train(model_definition,
data_train_df=curr_train_df,
data_test_df=curr_test_df,
experiment_name='cross_validation',
model_name='fold_' + str(fold_num),
output_directory=os.path.join(
temp_dir_name, 'results'))
# score on hold out fold
eval_batch_size = model_definition['training']['eval_batch_size']
batch_size = model_definition['training']['batch_size']
preds = model.predict(
preprocessed_data[2],
eval_batch_size if eval_batch_size != 0 else batch_size)
# augment the training statistics with scoring metric fron
# the hold out fold
train_stats['fold_metric'] = {}
for metric_category in preds:
train_stats['fold_metric'][metric_category] = {}
for metric in preds[metric_category]:
train_stats['fold_metric'][metric_category][metric] = \
preds[metric_category][metric]
# collect training statistics for this fold
kfold_training_stats['fold_' + str(fold_num)] = train_stats
# consolidate raw fold metrics across all folds
raw_kfold_stats = {}
for fold_name in kfold_training_stats:
for category in kfold_training_stats[fold_name]['fold_metric']:
if category not in raw_kfold_stats:
raw_kfold_stats[category] = {}
category_stats = \
kfold_training_stats[fold_name]['fold_metric'][category]
for metric in category_stats:
if metric not in {'predictions', 'probabilities'}:
if metric not in raw_kfold_stats[category]:
raw_kfold_stats[category][metric] = []
raw_kfold_stats[category][metric] \
.append(category_stats[metric])
# calculate overall kfold statistics
overall_kfold_stats = {}
for category in raw_kfold_stats:
overall_kfold_stats[category] = {}
for metric in raw_kfold_stats[category]:
mean = np.mean(raw_kfold_stats[category][metric])
std = np.std(raw_kfold_stats[category][metric])
overall_kfold_stats[category][metric + '_mean'] = mean
overall_kfold_stats[category][metric + '_std'] = std
kfold_training_stats['overall'] = overall_kfold_stats
# save k-fold cv statistics
save_json(os.path.join(output_directory, 'kfold_training_statistics.json'),
kfold_training_stats)
# save k-fold split indices
if not skip_save_k_fold_split_indices:
save_json(os.path.join(output_directory, 'kfold_split_indices.json'),
kfold_split_indices)
logger.info('completed {:d}-fold cross validation'.format(k_fold))
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_progress_weights=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_progress_weights: Skips saving the weights at the end of
each epoch. If this is true, training cannot be resumed from the
exactly the state at the end of the previous epoch.
:type skip_save_progress_weights: 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_progress_weights=skip_save_progress_weights,
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():
# 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 save_test_statistics(test_stats, experiment_dir_name):
test_stats_fn = os.path.join(experiment_dir_name, 'test_statistics.json')
save_json(test_stats_fn, test_stats)
def preprocess_for_training(
model_definition,
dataset_type='generic',
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,
skip_save_processed_input=False,
preprocessing_params=default_preprocessing_parameters,
random_seed=default_random_seed):
# Check if hdf5 and json already exist
data_hdf5_fp = None
data_train_hdf5_fp = None
data_validation_hdf5_fp = None
data_test_hdf5_fp = None
train_set_metadata_json_fp = 'metadata.json'
if data_csv is not None:
data_hdf5_fp = os.path.splitext(data_csv)[0] + '.hdf5'
train_set_metadata_json_fp = os.path.splitext(data_csv)[0] + '.json'
if (os.path.isfile(data_hdf5_fp)
and os.path.isfile(train_set_metadata_json_fp)):
logging.info('Found hdf5 and json with the same filename '
'of the csv, using them instead')
data_csv = None
data_hdf5 = data_hdf5_fp
train_set_metadata_json = train_set_metadata_json_fp
if data_train_csv is not None:
data_train_hdf5_fp = os.path.splitext(data_train_csv)[0] + '.hdf5'
train_set_metadata_json_fp = os.path.splitext(
data_train_csv)[0] + '.json'
if (os.path.isfile(data_train_hdf5_fp)
and os.path.isfile(train_set_metadata_json_fp)):
logging.info('Found hdf5 and json with the same filename of '
'the train csv, using them instead')
data_train_csv = None
data_train_hdf5 = data_train_hdf5_fp
train_set_metadata_json = train_set_metadata_json_fp
if data_validation_csv is not None:
data_validation_hdf5_fp = os.path.splitext(
data_validation_csv)[0] + '.hdf5'
if os.path.isfile(data_validation_hdf5_fp):
logging.info('Found hdf5 with the same filename of '
'the validation csv, using it instead')
data_validation_csv = None
data_validation_hdf5 = data_validation_hdf5_fp
if data_test_csv is not None:
data_test_hdf5_fp = os.path.splitext(data_test_csv)[0] + '.hdf5'
if os.path.isfile(data_test_hdf5_fp):
logging.info('Found hdf5 with the same filename of '
'the validation csv, using it instead')
data_test_csv = None
data_test_hdf5 = data_test_hdf5_fp
model_definition['data_hdf5_fp'] = data_hdf5_fp
# Decide if to preprocess or just load
features = (model_definition['input_features'] +
model_definition['output_features'])
(concatenate_csv, concatenate_df, build_dataset,
build_dataset_df) = get_dataset_fun(dataset_type)
if data_df is not None:
# needs preprocessing
logging.info('Using full dataframe')
logging.info('Building dataset (it may take a while)')
data, train_set_metadata = build_dataset_df(data_df,
features,
preprocessing_params,
random_seed=random_seed)
if not skip_save_processed_input:
logging.info('Writing dataset')
data_utils.save_hdf5(data_hdf5_fp, data, train_set_metadata)
logging.info('Writing train set metadata with vocabulary')
data_utils.save_json(train_set_metadata_json_fp, train_set_metadata)
training_set, test_set, validation_set = split_dataset_tvt(
data, data['split'])
elif data_train_df is not None:
# needs preprocessing
logging.info('Using training dataframe')
logging.info('Building dataset (it may take a while)')
concatenated_df = concatenate_df(data_train_df, data_validation_df,
data_test_df)
data, train_set_metadata = build_dataset_df(concatenated_df,
features,
preprocessing_params,
random_seed=random_seed)
training_set, test_set, validation_set = split_dataset_tvt(
data, data['split'])
if not skip_save_processed_input:
logging.info('Writing dataset')
data_utils.save_hdf5(data_train_hdf5_fp, training_set,
train_set_metadata)
if validation_set is not None:
data_utils.save_hdf5(data_validation_hdf5_fp, validation_set,
train_set_metadata)
if test_set is not None:
data_utils.save_hdf5(data_test_hdf5_fp, test_set,
train_set_metadata)
logging.info('Writing train set metadata with vocabulary')
data_utils.save_json(train_set_metadata_json_fp, train_set_metadata)
elif data_csv is not None:
# Use data and ignore _train, _validation and _test.
# Also ignore data and train set metadata needs preprocessing
logging.info('Using full raw csv, no hdf5 and json file '
'with the same name have been found')
logging.info('Building dataset (it may take a while)')
data, train_set_metadata = build_dataset(data_csv,
features,
preprocessing_params,
random_seed=random_seed)
if not skip_save_processed_input:
logging.info('Writing dataset')
data_utils.save_hdf5(data_hdf5_fp, data, train_set_metadata)
logging.info('Writing train set metadata with vocabulary')
data_utils.save_json(train_set_metadata_json_fp,
train_set_metadata)
training_set, test_set, validation_set = split_dataset_tvt(
data, data['split'])
elif data_train_csv is not None:
# use data_train (including _validation and _test if they are present)
# and ignore data and train set metadata
# needs preprocessing
logging.info('Using training raw csv, no hdf5 and json '
'file with the same name have been found')
logging.info('Building dataset (it may take a while)')
concatenated_df = concatenate_csv(data_train_csv, data_validation_csv,
data_test_csv)
concatenated_df.csv = data_train_csv
data, train_set_metadata = build_dataset_df(concatenated_df,
features,
preprocessing_params,
random_seed=random_seed)
training_set, test_set, validation_set = split_dataset_tvt(
data, data['split'])
if not skip_save_processed_input:
logging.info('Writing dataset')
data_utils.save_hdf5(data_train_hdf5_fp, training_set,
train_set_metadata)
if validation_set is not None:
data_utils.save_hdf5(data_validation_hdf5_fp, validation_set,
train_set_metadata)
if test_set is not None:
data_utils.save_hdf5(data_test_hdf5_fp, test_set,
train_set_metadata)
logging.info('Writing train set metadata with vocabulary')
data_utils.save_json(train_set_metadata_json_fp,
train_set_metadata)
elif data_hdf5 is not None and train_set_metadata_json is not None:
# use data and train set metadata
# doesn't need preprocessing, just load
logging.info('Using full hdf5 and json')
training_set, test_set, validation_set = load_data(
data_hdf5,
model_definition['input_features'],
model_definition['output_features'],
shuffle_training=True)
train_set_metadata = load_metadata(train_set_metadata_json)
elif data_train_hdf5 is not None and train_set_metadata_json is not None:
# use data and train set metadata
# doesn't need preprocessing, just load
logging.info('Using hdf5 and json')
training_set = load_data(data_train_hdf5,
model_definition['input_features'],
model_definition['output_features'],
split_data=False)
train_set_metadata = load_metadata(train_set_metadata_json)
if data_validation_hdf5 is not None:
validation_set = load_data(data_validation_hdf5,
model_definition['input_features'],
model_definition['output_features'],
split_data=False)
else:
validation_set = None
if data_test_hdf5 is not None:
test_set = load_data(data_test_hdf5,
model_definition['input_features'],
model_definition['output_features'],
split_data=False)
else:
test_set = None
else:
raise RuntimeError('Insufficient input parameters')
replace_text_feature_level(model_definition,
[training_set, validation_set, test_set])
training_dataset = Dataset(training_set,
model_definition['input_features'],
model_definition['output_features'],
data_hdf5_fp)
validation_dataset = None
if validation_set is not None:
validation_dataset = Dataset(validation_set,
model_definition['input_features'],
model_definition['output_features'],
data_hdf5_fp)
test_dataset = None
if test_set is not None:
test_dataset = Dataset(test_set, model_definition['input_features'],
model_definition['output_features'],
data_hdf5_fp)
return (training_dataset, validation_dataset, test_dataset,
train_set_metadata)
def save(self, filepath):
save_json(filepath, self.__dict__)
def save_prediction_statistics(prediction_stats, experiment_dir_name):
test_stats_fn = os.path.join(experiment_dir_name,
'prediction_statistics.json')
save_json(test_stats_fn, prediction_stats)
def train(
self,
dataset=None,
training_set=None,
validation_set=None,
test_set=None,
training_set_metadata=None,
data_format=None,
experiment_name='api_experiment',
model_name='run',
model_resume_path=None,
skip_save_training_description=False,
skip_save_training_statistics=False,
skip_save_model=False,
skip_save_progress=False,
skip_save_log=False,
skip_save_processed_input=False,
output_directory='results',
random_seed=default_random_seed,
debug=False,
**kwargs
):
"""This function is used to perform a full training of the model on the
specified dataset.
# Inputs
:param dataset: (string, dict, DataFrame) source containing the entire dataset.
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 training_set: (string, dict, DataFrame) source containing training data.
:param validation_set: (string, dict, DataFrame) source containing validation data.
:param test_set: (string, dict, DataFrame) source containing test data.
:param training_set_metadata: (string, dict) metadata JSON file or loaded metadata.
Intermediate preprocess structure 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 data_format: (string) format to interpret data sources. Will be inferred
automatically if not specified.
:param experiment_name: (string) a name for the experiment, used for the save
directory
:param model_name: (string) a name for the model, used for the save
directory
:param model_resume_path: (string) path of a the model directory to
resume training of
:param skip_save_training_description: (bool, default: `False`) disables
saving the description JSON file.
:param skip_save_training_statistics: (bool, default: `False`) disables
saving training statistics JSON file.
: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 metric 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 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
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, DataFrame)) tuple containing:
- A dictionary of training statistics for each output feature containing
loss and metrics values for each epoch. The second return
- A Pandas DataFrame of preprocessed training data.
"""
# setup directories and file names
if model_resume_path is not None:
if os.path.exists(model_resume_path):
output_directory = model_resume_path
else:
if is_on_master():
logger.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():
output_directory = get_output_directory(
output_directory,
experiment_name,
model_name
)
else:
output_directory = None
# if we are skipping all saving,
# there is no need to create a directory that will remain empty
should_create_output_directory = not (
skip_save_training_description and
skip_save_training_statistics and
skip_save_model and
skip_save_progress and
skip_save_log and
skip_save_processed_input
)
description_fn = training_stats_fn = model_dir = None
if is_on_master():
if should_create_output_directory:
if not os.path.exists(output_directory):
os.makedirs(output_directory, exist_ok=True)
description_fn, training_stats_fn, model_dir = get_file_names(
output_directory)
# save description
if is_on_master():
description = get_experiment_description(
self.model_definition,
dataset=dataset,
training_set=training_set,
validation_set=validation_set,
test_set=test_set,
training_set_metadata=training_set_metadata,
data_format=data_format,
random_seed=random_seed
)
if not skip_save_training_description:
save_json(description_fn, description)
# print description
logger.info('Experiment name: {}'.format(experiment_name))
logger.info('Model name: {}'.format(model_name))
logger.info('Output directory: {}'.format(output_directory))
logger.info('\n')
for key, value in description.items():
logger.info('{}: {}'.format(key, pformat(value, indent=4)))
logger.info('\n')
# preprocess
preprocessed_data = preprocess_for_training(
self.model_definition,
dataset=dataset,
training_set=training_set,
validation_set=validation_set,
test_set=test_set,
training_set_metadata=training_set_metadata,
data_format=data_format,
skip_save_processed_input=skip_save_processed_input,
preprocessing_params=self.model_definition[PREPROCESSING],
random_seed=random_seed
)
(training_set,
validation_set,
test_set,
training_set_metadata) = preprocessed_data
self.training_set_metadata = training_set_metadata
if is_on_master():
logger.info('Training set: {0}'.format(training_set.size))
if validation_set is not None:
logger.info('Validation set: {0}'.format(validation_set.size))
if test_set is not None:
logger.info('Test set: {0}'.format(test_set.size))
if is_on_master():
if not skip_save_model:
# save train set metadata
os.makedirs(model_dir, exist_ok=True)
save_json(
os.path.join(
model_dir,
TRAIN_SET_METADATA_FILE_NAME
),
training_set_metadata
)
contrib_command("train_init", experiment_directory=output_directory,
experiment_name=experiment_name, model_name=model_name,
output_directory=output_directory,
resume=model_resume_path is not None)
# Build model if not provided
# if it was provided it means it was already loaded
if not self.model:
if is_on_master():
print_boxed('MODEL', print_fun=logger.debug)
# update model definition with metadata properties
update_model_definition_with_metadata(
self.model_definition,
training_set_metadata
)
self.model = LudwigModel.create_model(self.model_definition,
random_seed=random_seed)
# init trainer
trainer = Trainer(
**self.model_definition[TRAINING],
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,
random_seed=random_seed,
horoovd=self._horovod,
debug=debug
)
contrib_command("train_model", self.model, self.model_definition,
self.model_definition_fp)
# train model
if is_on_master():
print_boxed('TRAINING')
if not skip_save_model:
self.save_model_definition(model_dir)
train_stats = trainer.train(
self.model,
training_set,
validation_set=validation_set,
test_set=test_set,
save_path=model_dir,
)
train_trainset_stats, train_valiset_stats, train_testset_stats = train_stats
train_stats = {
TRAINING: train_trainset_stats,
VALIDATION: train_valiset_stats,
TEST: train_testset_stats
}
# save training statistics
if is_on_master():
if not skip_save_training_statistics:
save_json(training_stats_fn, train_stats)
# grab the results of the model with highest validation test performance
validation_field = trainer.validation_field
validation_metric = trainer.validation_metric
validation_field_result = train_valiset_stats[validation_field]
best_function = get_best_function(validation_metric)
# results of the model with highest validation test performance
if is_on_master() and validation_set is not None:
epoch_best_vali_metric, best_vali_metric = best_function(
enumerate(validation_field_result[validation_metric]),
key=lambda pair: pair[1]
)
logger.info(
'Best validation model epoch: {0}'.format(
epoch_best_vali_metric + 1)
)
logger.info(
'Best validation model {0} on validation set {1}: {2}'.format(
validation_metric, validation_field, best_vali_metric
))
if test_set is not None:
best_vali_metric_epoch_test_metric = train_testset_stats[
validation_field][validation_metric][
epoch_best_vali_metric]
logger.info(
'Best validation model {0} on test set {1}: {2}'.format(
validation_metric,
validation_field,
best_vali_metric_epoch_test_metric
)
)
logger.info(
'\nFinished: {0}_{1}'.format(experiment_name, model_name))
logger.info('Saved to: {0}'.format(output_directory))
contrib_command("train_save", output_directory)
self.training_set_metadata = training_set_metadata
if not skip_save_model:
# Load the best weights from saved checkpoint
self.load_weights(model_dir)
return train_stats, preprocessed_data, output_directory
type=yaml.safe_load,
default='{}',
help='the parameters for preprocessing the different features'
)
parser.add_argument(
'-rs',
'--random_seed',
type=int,
default=42,
help='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'
)
args = parser.parse_args()
data, train_set_metadata = build_dataset(
args.dataset_csv,
args.train_set_metadata_json,
args.features,
args.preprocessing_parameters,
args.random_seed
)
# write train set metadata, dataset
logger.info('Writing train set metadata with vocabulary')
data_utils.save_json(args.output_metadata_json, train_set_metadata)
logger.info('Writing dataset')
data_utils.save_hdf5(args.output_dataset_h5, data, train_set_metadata)
def _preprocess_csv_for_training(
features,
data_csv=None,
data_train_csv=None,
data_validation_csv=None,
data_test_csv=None,
train_set_metadata_json=None,
skip_save_processed_input=False,
preprocessing_params=default_preprocessing_parameters,
random_seed=default_random_seed
):
"""
Method to pre-process csv data
:param features: list of all features (input + output)
:param data_csv: path to the csv data
:param data_train_csv: training csv data
:param data_validation_csv: validation csv data
:param data_test_csv: test csv data
:param train_set_metadata_json: train set metadata json
:param skip_save_processed_input: if False, the pre-processed data is saved
as .hdf5 files in the same location as the csvs with the same names.
:param preprocessing_params: preprocessing parameters
:param random_seed: random seed
:return: training, test, validation datasets, training metadata
"""
train_set_metadata = None
if train_set_metadata_json is not None:
train_set_metadata = load_metadata(train_set_metadata_json)
if data_csv is not None:
# Use data and ignore _train, _validation and _test.
# Also ignore data and train set metadata needs preprocessing
logger.info(
'Using full raw csv, no hdf5 and json file '
'with the same name have been found'
)
logger.info('Building dataset (it may take a while)')
data, train_set_metadata = build_dataset(
data_csv,
features,
preprocessing_params,
train_set_metadata=train_set_metadata,
random_seed=random_seed
)
if not skip_save_processed_input:
logger.info('Writing dataset')
data_hdf5_fp = replace_file_extension(data_csv, 'hdf5')
data_utils.save_hdf5(data_hdf5_fp, data, train_set_metadata)
logger.info('Writing train set metadata with vocabulary')
train_set_metadata_json_fp = replace_file_extension(
data_csv,
'json'
)
data_utils.save_json(
train_set_metadata_json_fp, train_set_metadata)
training_set, test_set, validation_set = split_dataset_tvt(
data,
data['split']
)
elif data_train_csv is not None:
# use data_train (including _validation and _test if they are present)
# and ignore data and train set metadata
# needs preprocessing
logger.info(
'Using training raw csv, no hdf5 and json '
'file with the same name have been found'
)
logger.info('Building dataset (it may take a while)')
concatenated_df = concatenate_csv(
data_train_csv,
data_validation_csv,
data_test_csv
)
concatenated_df.csv = data_train_csv
data, train_set_metadata = build_dataset_df(
concatenated_df,
features,
preprocessing_params,
train_set_metadata=train_set_metadata,
random_seed=random_seed
)
training_set, test_set, validation_set = split_dataset_tvt(
data,
data['split']
)
if not skip_save_processed_input:
logger.info('Writing dataset')
data_train_hdf5_fp = replace_file_extension(data_train_csv, 'hdf5')
data_utils.save_hdf5(
data_train_hdf5_fp,
training_set,
train_set_metadata
)
if validation_set is not None:
data_validation_hdf5_fp = replace_file_extension(
data_validation_csv,
'hdf5'
)
data_utils.save_hdf5(
data_validation_hdf5_fp,
validation_set,
train_set_metadata
)
if test_set is not None:
data_test_hdf5_fp = replace_file_extension(data_test_csv,
'hdf5')
data_utils.save_hdf5(
data_test_hdf5_fp,
test_set,
train_set_metadata
)
logger.info('Writing train set metadata with vocabulary')
train_set_metadata_json_fp = replace_file_extension(data_train_csv,
'json')
data_utils.save_json(train_set_metadata_json_fp, train_set_metadata)
return training_set, test_set, validation_set, train_set_metadata
def save_evaluation_stats(test_stats, output_directory):
test_stats_fn = os.path.join(output_directory, "test_statistics.json")
save_json(test_stats_fn, test_stats)
