def experiment_cli(
config: dict,
config_file: str = None,
dataset: Union[str, dict, pd.DataFrame] = None,
training_set: Union[str, dict, pd.DataFrame] = None,
validation_set: Union[str, dict, pd.DataFrame] = None,
test_set: Union[str, dict, pd.DataFrame] = None,
training_set_metadata: Union[str, dict] = None,
data_format: str = None,
experiment_name: str = 'experiment',
model_name: str = 'run',
model_load_path: str = None,
model_resume_path: str = None,
eval_split: str = TEST,
skip_save_training_description: bool = False,
skip_save_training_statistics: bool = False,
skip_save_model: bool = False,
skip_save_progress: bool = False,
skip_save_log: bool = False,
skip_save_processed_input: bool = False,
skip_save_unprocessed_output: bool = False,
skip_save_predictions: bool = False,
skip_save_eval_stats: bool = False,
skip_collect_predictions: bool = False,
skip_collect_overall_stats: bool = False,
output_directory: str = 'results',
gpus: Union[str, int, List[int]] = None,
gpu_memory_limit: int = None,
allow_parallel_threads: bool = True,
use_horovod: bool = None,
random_seed: int = default_random_seed,
debug: bool = False,
logging_level: int = logging.INFO,
**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.
# Inputs
:param config: (dict) config which defines the different
parameters of the model, features, preprocessing and training.
:param config_file: (str, default: `None`) the filepath string
that specifies the config. It is a yaml file.
:param dataset: (Union[str, dict, pandas.DataFrame], default: `None`)
source containing the entire dataset to be used in the experiment.
If it has a split column, it will be used for splitting (0 for train,
1 for validation, 2 for test), otherwise the dataset will be
randomly split.
:param training_set: (Union[str, dict, pandas.DataFrame], default: `None`)
source containing training data.
:param validation_set: (Union[str, dict, pandas.DataFrame], default: `None`)
source containing validation data.
:param test_set: (Union[str, dict, pandas.DataFrame], default: `None`)
source containing test data.
:param training_set_metadata: (Union[str, dict], default: `None`)
metadata JSON file or loaded metadata. Intermediate preprocess
structure containing the mappings of the input
dataset created the first time an input file is used in the same
directory with the same name and a '.meta.json' extension.
:param data_format: (str, default: `None`) format to interpret data
sources. Will be inferred automatically if not specified. Valid
formats are `'auto'`, `'csv'`, `'excel'`, `'feather'`,
`'fwf'`, `'hdf5'` (cache file produced during previous training),
`'html'` (file containing a single HTML `<table>`), `'json'`, `'jsonl'`,
`'parquet'`, `'pickle'` (pickled Pandas DataFrame), `'sas'`, `'spss'`,
`'stata'`, `'tsv'`.
:param experiment_name: (str, default: `'experiment'`) name for
the experiment.
:param model_name: (str, default: `'run'`) name of the model that is
being used.
:param model_load_path: (str, default: `None`) if this is specified the
loaded model will be used as initialization
(useful for transfer learning).
:param model_resume_path: (str, default: `None`) resumes training of
the model from the path specified. The config is restored.
In addition to config, training statistics and loss for
epoch and the state of the optimizer are restored such that
training can be effectively continued from a previously interrupted
training process.
:param eval_split: (str, default: `test`) split on which
to perform evaluation. Valid values are `training`, `validation`
and `test`.
: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 improves, 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 and the returned model
will have the weights obtained at the end of training, instead of
the weights of the epoch with the best validation performance.
: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`) if input
dataset is provided it is preprocessed and cached by saving an HDF5
and JSON files to avoid running the preprocessing again. If this
parameter is `False`, the HDF5 and JSON file are not saved.
:param skip_save_unprocessed_output: (bool, default: `False`) by default
predictions and their probabilities are saved in both raw
unprocessed numpy files containing 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.
:param skip_save_predictions: (bool, default: `False`) skips saving test
predictions CSV files
:param skip_save_eval_stats: (bool, default: `False`) skips saving test
statistics JSON file
:param skip_collect_predictions: (bool, default: `False`) skips
collecting post-processed predictions during eval.
:param skip_collect_overall_stats: (bool, default: `False`) skips
collecting overall stats during eval.
:param output_directory: (str, default: `'results'`) the directory that
will contain the training statistics, TensorBoard logs, the saved
model and the training progress files.
:param gpus: (list, default: `None`) list of GPUs that are available
for training.
: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 use_horovod: (bool, default: `None`) flag for using horovod.
:param random_seed: (int: default: 42) random seed used for weights
initialization, splits and any other random function.
:param debug: (bool, default: `False) if `True` turns on `tfdbg` with
`inf_or_nan` checks.
:param logging_level: (int) Log level that will be sent to stderr.
# Return
:return: (Tuple[LudwigModel, dict, dict, tuple, str)) `(model, evaluation_statistics, training_statistics, preprocessed_data, output_directory)`
`model` LudwigModel instance
`evaluation_statistics` dictionary with evaluation performance
statistics on the test_set,
`training_statistics` is a dictionary of training statistics for
each output
feature containing loss and metrics values for each epoch,
`preprocessed_data` tuple containing preprocessed
`(training_set, validation_set, test_set)`, `output_directory`
filepath string to where results are stored.
"""
set_on_master(use_horovod)
config = check_which_config(config,
config_file)
if model_load_path:
model = LudwigModel.load(model_load_path)
else:
model = LudwigModel(
config=config,
logging_level=logging_level,
use_horovod=use_horovod,
gpus=gpus,
gpu_memory_limit=gpu_memory_limit,
allow_parallel_threads=allow_parallel_threads,
)
(
eval_stats,
train_stats,
preprocessed_data,
output_directory
) = model.experiment(
dataset=dataset,
training_set=training_set,
validation_set=validation_set,
test_set=test_set,
training_set_metadata=training_set_metadata,
data_format=data_format,
experiment_name=experiment_name,
model_name=model_name,
model_resume_path=model_resume_path,
eval_split=eval_split,
skip_save_training_description=skip_save_training_description,
skip_save_training_statistics=skip_save_training_statistics,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
skip_save_processed_input=skip_save_processed_input,
skip_save_unprocessed_output=skip_save_unprocessed_output,
skip_save_predictions=skip_save_predictions,
skip_save_eval_stats=skip_save_eval_stats,
skip_collect_predictions=skip_collect_predictions,
skip_collect_overall_stats=skip_collect_overall_stats,
output_directory=output_directory,
random_seed=random_seed,
debug=debug,
)
return model, eval_stats, train_stats, preprocessed_data, output_directory
def experiment_cli(model_definition,
model_definition_file=None,
dataset=None,
training_set=None,
validation_set=None,
test_set=None,
data_format=None,
training_set_metadata=None,
experiment_name='experiment',
model_name='run',
model_load_path=None,
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,
skip_save_unprocessed_output=False,
skip_save_predictions=False,
skip_save_eval_stats=False,
skip_collect_predictions=False,
skip_collect_overall_stats=False,
output_directory='results',
gpus=None,
gpu_memory_limit=None,
allow_parallel_threads=True,
use_horovod=None,
random_seed=default_random_seed,
debug=False,
logging_level=logging.INFO,
**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 dataset: 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.
:type dataset: Str, Dictionary, DataFrame
:param training_set: Source containing training data.
:type training_set: Str, Dictionary, DataFrame
:param validation_set: Source containing validation data.
:type validation_set: Str, Dictionary, DataFrame
:param test_set: Source containing test data.
:type test_set: Str, Dictionary, DataFrame
:param training_set_metadata: 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.
:type training_set_metadata: Str, Dictionary
:param data_format: Format to interpret data sources. Will be inferred
automatically if not specified.
:type data_format: 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 continuing a previously interrupted
training process.
:type model_resume_path: filepath (str)
:param skip_save_training_description: Disables saving
the description JSON file.
:type skip_save_training_description: Boolean
:param skip_save_training_statistics: Disables saving
training statistics JSON file.
:type skip_save_training_statistics: Boolean
: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 metric improves, 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
containing 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 skip_save_predictions: skips saving test predictions CSV files
:type skip_save_predictions: Boolean
:param skip_save_eval_stats: skips saving test statistics JSON file
:type skip_save_eval_stats: Boolean
:param skip_collect_predictions: skips collecting post-processed predictions during eval.
:type skip_collect_predictions: Boolean
:param skip_collect_overall_stats: skips collecting overall stats during eval.
:type skip_collect_overall_stats: Boolean
:param output_directory: The directory that will contain the training
statistics, the saved model and the training progress files.
:type output_directory: filepath (str)
:param gpus: List of GPUs that are available for training.
:type gpus: List
:param gpu_memory_limit: maximum memory in MB to allocate per GPU device.
:type gpu_memory_limit: Integer
:param allow_parallel_threads: allow TensorFlow to use multithreading parallelism
to improve performance at the cost of determinism.
:type allow_parallel_threads: Boolean
:param use_horovod: Flag for using horovod
:type use_horovod: Boolean
: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
:param logging_level: Log level to send to stderr.
:type logging_level: int
"""
set_on_master(use_horovod)
model_definition = check_which_model_definition(model_definition,
model_definition_file)
if model_load_path:
model = LudwigModel.load(model_load_path)
else:
model = LudwigModel(model_definition=model_definition,
logging_level=logging_level,
use_horovod=use_horovod,
gpus=gpus,
gpu_memory_limit=gpu_memory_limit,
allow_parallel_threads=allow_parallel_threads,
random_seed=random_seed)
(test_results, train_stats, preprocessed_data,
output_directory) = model.experiment(
dataset=dataset,
training_set=training_set,
validation_set=validation_set,
test_set=test_set,
training_set_metadata=training_set_metadata,
data_format=data_format,
experiment_name=experiment_name,
model_name=model_name,
model_resume_path=model_resume_path,
skip_save_training_description=skip_save_training_description,
skip_save_training_statistics=skip_save_training_statistics,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
skip_save_processed_input=skip_save_processed_input,
skip_save_unprocessed_output=skip_save_unprocessed_output,
skip_save_predictions=skip_save_predictions,
skip_save_eval_stats=skip_save_eval_stats,
skip_collect_predictions=skip_collect_predictions,
skip_collect_overall_stats=skip_collect_overall_stats,
output_directory=output_directory,
random_seed=random_seed,
debug=debug,
)
return model, test_results, train_stats, preprocessed_data, output_directory
def run_experiment(
config,
dataset=None,
training_set=None,
validation_set=None,
test_set=None,
training_set_metadata=None,
data_format=None,
experiment_name="hyperopt",
model_name="run",
# model_load_path=None,
# model_resume_path=None,
eval_split=VALIDATION,
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,
skip_save_unprocessed_output=False,
skip_save_predictions=False,
skip_save_eval_stats=False,
output_directory="results",
gpus=None,
gpu_memory_limit=None,
allow_parallel_threads=True,
backend=None,
callbacks=None,
random_seed=default_random_seed,
debug=False,
**kwargs):
# Collect training and validation losses and metrics
# & append it to `results`
model = LudwigModel(
config=config,
backend=backend,
gpus=gpus,
gpu_memory_limit=gpu_memory_limit,
allow_parallel_threads=allow_parallel_threads,
)
eval_stats, train_stats, _, _ = model.experiment(
dataset=dataset,
training_set=training_set,
validation_set=validation_set,
test_set=test_set,
training_set_metadata=training_set_metadata,
data_format=data_format,
experiment_name=experiment_name,
model_name=model_name,
# model_load_path=model_load_path,
# model_resume_path=model_resume_path,
eval_split=eval_split,
skip_save_training_description=skip_save_training_description,
skip_save_training_statistics=skip_save_training_statistics,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
skip_save_processed_input=skip_save_processed_input,
skip_save_unprocessed_output=skip_save_unprocessed_output,
skip_save_predictions=skip_save_predictions,
skip_save_eval_stats=skip_save_eval_stats,
output_directory=output_directory,
skip_collect_predictions=True,
skip_collect_overall_stats=False,
callbacks=callbacks,
random_seed=random_seed,
debug=debug,
)
return train_stats, eval_stats
def execute(
self,
config,
dataset=None,
training_set=None,
validation_set=None,
test_set=None,
training_set_metadata=None,
data_format=None,
experiment_name="hyperopt",
model_name="run",
# model_load_path=None,
# 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,
skip_save_unprocessed_output=False,
skip_save_predictions=False,
skip_save_eval_stats=False,
output_directory="results",
gpus=None,
gpu_memory_limit=None,
allow_parallel_threads=True,
backend=None,
random_seed=default_random_seed,
debug=False,
**kwargs):
hyperopt_results = []
trials = 0
while not self.hyperopt_sampler.finished():
sampled_parameters = self.hyperopt_sampler.sample_batch()
metric_scores = []
for i, parameters in enumerate(sampled_parameters):
modified_config = substitute_parameters(
copy.deepcopy(config), parameters)
trial_id = trials + i
model = LudwigModel(
config=modified_config,
backend=backend,
gpus=gpus,
gpu_memory_limit=gpu_memory_limit,
allow_parallel_threads=allow_parallel_threads,
)
eval_stats, train_stats, _, _ = model.experiment(
dataset=dataset,
training_set=training_set,
validation_set=validation_set,
test_set=test_set,
training_set_metadata=training_set_metadata,
data_format=data_format,
experiment_name=f'{experiment_name}_{trial_id}',
model_name=model_name,
# model_load_path=model_load_path,
# model_resume_path=model_resume_path,
eval_split=self.split,
skip_save_training_description=
skip_save_training_description,
skip_save_training_statistics=skip_save_training_statistics,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
skip_save_processed_input=skip_save_processed_input,
skip_save_unprocessed_output=skip_save_unprocessed_output,
skip_save_predictions=skip_save_predictions,
skip_save_eval_stats=skip_save_eval_stats,
output_directory=output_directory,
skip_collect_predictions=True,
skip_collect_overall_stats=False,
random_seed=random_seed,
debug=debug,
)
metric_score = self.get_metric_score(train_stats, eval_stats)
metric_scores.append(metric_score)
hyperopt_results.append({
"parameters": parameters,
"metric_score": metric_score,
"training_stats": train_stats,
"eval_stats": eval_stats,
})
trials += len(sampled_parameters)
self.hyperopt_sampler.update_batch(
zip(sampled_parameters, metric_scores))
hyperopt_results = self.sort_hyperopt_results(hyperopt_results)
return hyperopt_results
def run_api_commands(
input_features,
output_features,
data_csv,
output_dir,
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,
skip_save_unprocessed_output=False,
skip_save_predictions=False,
skip_save_eval_stats=False,
skip_collect_predictions=False,
skip_collect_overall_stats=False,
):
"""
Helper method to avoid code repetition in running an experiment
:param input_features: input schema
:param output_features: output schema
:param data_csv: path to data
:return: None
"""
config = {
'input_features': input_features,
'output_features': output_features,
'combiner': {
'type': 'concat',
'fc_size': 14
},
'training': {
'epochs': 2
}
}
model = LudwigModel(config)
# Training with csv
model.train(dataset=data_csv,
skip_save_training_description=skip_save_training_description,
skip_save_training_statistics=skip_save_training_statistics,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
skip_save_processed_input=skip_save_processed_input,
output_directory=output_dir)
model.predict(dataset=data_csv,
skip_save_unprocessed_output=skip_save_unprocessed_output,
skip_save_predictions=skip_save_predictions,
output_directory=output_dir)
model.evaluate(dataset=data_csv,
skip_save_unprocessed_output=skip_save_unprocessed_output,
skip_save_predictions=skip_save_predictions,
skip_save_eval_stats=skip_save_eval_stats,
collect_predictions=not skip_collect_predictions,
collect_overall_stats=not skip_collect_overall_stats,
output_directory=output_dir)
model.experiment(
dataset=data_csv,
skip_save_training_description=skip_save_training_description,
skip_save_training_statistics=skip_save_training_statistics,
skip_save_model=skip_save_model,
skip_save_progress=skip_save_progress,
skip_save_log=skip_save_log,
skip_save_processed_input=skip_save_processed_input,
skip_save_unprocessed_output=skip_save_unprocessed_output,
skip_save_predictions=skip_save_predictions,
skip_save_eval_stats=skip_save_eval_stats,
skip_collect_predictions=skip_collect_predictions,
skip_collect_overall_stats=skip_collect_overall_stats,
output_directory=output_dir)
with open(output_file, 'w+', newline='') as f_out:
csv_writer = csv.writer(f_out)
for row in processed_rows:
csv_writer.writerow(row)
if __name__ == "__main__":
if sys.argv[1] == "train":
if len(sys.argv) != 5:
print("Incorrect number of arguments. Please use format:\npython main.py train <path-to-input-csv-file> <path-to-output-csv-file> <ludwig-model-definition>")
preprocess_dataset(sys.argv[2], sys.argv[3], 1)
config = sys.argv[4]
model = LudwigModel(config)
train_stats = model.experiment(dataset=sys.argv[3], training_set=sys.argv[3], validation_set=sys.argv[3],
test_set=sys.argv[3], experiment_name='covid_inference', model_name='train')
print(train_stats)
elif sys.argv[1] == "evaluate":
if len(sys.argv) != 6:
print("Incorrect number of arguments. Please use format:\npython main.py evaluate <path-to-trained-model> <path-to-input-csv-file> <path-to-output-csv-file> <ludwig-model-definition>")
preprocess_dataset(sys.argv[3], sys.argv[4], 0)
config = sys.argv[5]
model = LudwigModel.load(sys.argv[2])
train_stats = model.evaluate(dataset=sys.argv[4], skip_save_predictions=False,
skip_save_eval_stats=False, collect_predictions=True, collect_overall_stats=True)
print(train_stats)
else:
print(
