def main(frac: float = default_sample_frac):
"""
Creates a sample of the solar wind data and saves it.
# Parameters
frac: `float`
should be between 0.0 and 1.0 and represent the proportion
of the dataset to include on the sample dataset.
"""
logging.info(f'making sample of {frac}%')
logging.info('reading config file')
config = load_data.read_config_file('./config/config.yml')
directories = config['directories']
interim_path = Path(directories['interim'])
# reading gt data
logging.info('reading training data')
solar_wind = load_data.read_feather(interim_path / 'solar_wind.feather')
logging.info('splitting dataset')
_, valid_idx = load_data.split_train_data(solar_wind,
test_frac=frac,
eval_mode=True)
sample_data = solar_wind.loc[valid_idx, :]
sample_data.reset_index(drop=True, inplace=True)
logging.info('saving file..')
sample_data.to_feather(interim_path / 'sample_solar_wind.feather')
def main():
"""
This function will save the solar wind data
as a Feather file
"""
# read the main config file
config = load_data.read_config_file('./config/config.yml')
# get the path to the CSV File
directories = config['directories']
raw_path = Path(directories['raw'])
interim_path = Path(directories['interim'])
interim_path.mkdir(exist_ok=True, parents=True)
logging.info('reading solar wind data..')
# reading CSV file
solar_wind = load_data.read_csv(raw_path / 'solar_wind.csv')
logging.info('saving to feather..')
# saving as feather file
solar_wind.to_feather(interim_path / 'solar_wind.feather')
def main(use_sample: bool = False, n_jobs: int = 1):
"""
This function will apply all the steps in order to create
a dataset ready to train models.
The following steps:
- read the data
- compute the solar wind features
- compute sattelite positions features
- take the log of smoothed_ssn values
- create the target for the actual time t and t + 1 hour
- merge all dataset into a single one
- save the dataset for future modeling
# Params
use_sample: `bool`, optional(defualt=False)
Whether or not to use the sample dataset
n_jobs: `in`, optional(defualt=1)
The number of jobs to run in parallel
"""
logging.info(f'use_sample={use_sample}, n_jobs={n_jobs}')
logging.info('reading config file')
config = load_data.read_config_file('./config/config.yml')
# directories
directories = config['directories']
raw_path = Path(directories['raw'])
interim_path = Path(directories['interim'])
processed_path = Path(directories['processed'])
processed_path.mkdir(exist_ok=True, parents=True)
# reading gt data
solar_wind_file = ('sample_solar_wind.feather'
if use_sample else 'solar_wind.feather')
logging.info('reading training data')
dst_labels = load_data.read_csv(raw_path / 'dst_labels.csv')
solar_wind = load_data.read_feather(interim_path / solar_wind_file)
sunspots = load_data.read_csv(raw_path / 'sunspots.csv')
stl_pos = load_data.read_csv(raw_path / 'satellite_positions.csv')
logging.info('preprocessing solar wing')
# preprocessing solar wind
# setting timedelta as index
solar_wind.set_index('timedelta', inplace=True)
# preprocessing solar wind time series
solar_wind = solar_wind_preprocessing(solar_wind)
logging.info('computing features')
start = time.time()
# computing solar wind features
data = split_into_period(solar_wind,
features=default.init_features,
n_jobs=n_jobs)
elapsed_time = (time.time() - start) / 60
logging.info(f'elapsed time {elapsed_time:.4f}')
logging.info('merging other datasets')
# create target
target = create_target(dst_labels)
# preprocessing sattelite positions
stl_pos = stl_preprocessing(stl_pos)
# taking the log of smoothed_ssn values
sunspots['smoothed_ssn'] = np.log(sunspots['smoothed_ssn'])
# merging dataframes to the main dataframe
data = merge_daily(data, stl_pos)
data = merge_daily(data, sunspots)
# merging target dataframe to the main dataframe
data = data.merge(target, how='left', on=['period', 'timedelta'])
# droping last values where there is not available data
data.dropna(subset=['t0', 't1'], inplace=True)
# reset index
data.reset_index(inplace=True, drop=True)
logging.info('saving')
output_filename = 'fe' if not use_sample else 'fe_sample'
# saving to feather format
data.to_feather(processed_path / f'{output_filename}.feather')
def main(experiment_path: str,
eval_mode: bool = True,
use_sample: bool = False,
test_frac: float = 0.2,
message: str = None,
fi_threshold: float = None):
"""
A function to train or validate an Experiment
# Parameters
experiment_path: `str`
A path to the folder's experiment config file.
the config file must be named config.yml and
it must contain the following keys:
model: `str`
the path to the model config file
pipeline: `str`
the path to the pipeline config file
optimizer: `Dict[str, Any]`
the parameters for the Adam optimizer
epochs: `int`, optinal(default=10)
the number of epochs to train the model
use_sigmoid: `bool`, optional(defualt=False)
Whether or not to use as the final activation
function of the model
eval_mode: `bool`, optional (default=True)
if False, the model will be train using all the data available,
otherwise, the trained model will be use for inference.
use_sample: `bool`, optional (default=False)
if True, we will use only a sample from the dataset.
to use it, before execute the make_sample.py file
to create this sample dataset.
test_frac: `float`, optional (default=0.2)
if eval_mode is True, the size of the valid dataset
will be the {test_frac}% of the main dataset.
message: `str`, optional (default=None)
we use mlflow to keep track of all parameters and errors of each experiment,
this parameter will register any string you pass into the experiment record in mlflow.
fi_threshold `float`, optional (default=None)
if already exists a feature importance file, this value will be use for filtering
the features that has greater importance values than {fi_threshold}.
"""
# getting experiment name
experiment = os.path.basename(experiment_path)
logging.info(f'running {experiment}')
logging.info(f'eval_mode={eval_mode}, use_sample={use_sample}')
logging.info('reading config file')
# creating experiment path and loading experiment config file
experiment_path = Path(experiment_path)
config = load_data.read_config_file('./config/config.yml')
experiment_config = load_data.read_config_file(experiment_path /
'config.yml')
# reading experiment's model and pipeline config file
pipeline_config = load_data.read_config_file(experiment_config['pipeline'])
model_config = load_data.read_config_file(experiment_config['model'])
directories = config['directories']
# getting the data path
processed_path = Path(directories['processed'])
# creating a prediction folder to save prediction after training
prediction_path = experiment_path / 'prediction'
prediction_path.mkdir(exist_ok=True, parents=True)
# creating a model path to save models after training
model_path = experiment_path / 'models'
# reading preprocessed data
filename = ('fe' if not use_sample else 'fe_sample')
logging.info('reading training data')
data = load_data.read_feather(processed_path / f'{filename}.feather')
logging.info('splitting dataset')
train_idx, valid_idx = load_data.split_train_data(data,
test_frac=test_frac,
eval_mode=eval_mode)
train_data = data.loc[train_idx, :]
valid_data = data.loc[valid_idx, :]
train_data.reset_index(drop=True, inplace=True)
valid_data.reset_index(drop=True, inplace=True)
# importing pipeline
logging.info('building pipeline')
pipeline = build_pipeline(pipeline_config)
logging.info(f'{pipeline}')
# fit pipeline
logging.info('training pipeline')
pipeline.fit(train_data)
# transform both training and valid dataset
logging.info('transforming datasets')
train_data = pipeline.transform(train_data)
valid_data = pipeline.transform(valid_data)
# getting the bottom and upper limit of the target
# in the case we want to use the sigmoid function
# as the final activation function of our model
use_sigmoid = experiment_config.pop('use_sigmoid', False)
y_limit = ((train_data['t0'].agg(
('max', 'min')) * 1.2).to_list() if use_sigmoid else None)
# loading the features to train our model
# if exists a feature importance file
# we can use it to train our model only with revelant features
features = load_data.get_features(train_data,
experiment_path=experiment_path,
fi_threshold=fi_threshold,
ignore_features=default.ignore_features)
in_features = len(features)
logging.info(f'modeling using {len(features)} features')
logging.info(f'{features[:30]}')
# creating datasets
train_ds = Dataset.from_dataframe(train_data,
features=features,
target=target_name,
device=device)
valid_ds = Dataset.from_dataframe(valid_data,
features=features,
target=target_name,
device=device)
# creating dataloaders
train_dl = DataLoader(dataset=train_ds,
batch_size=batch_size,
shuffle=True)
not_shuffle_train_dl = DataLoader(dataset=train_ds,
batch_size=batch_size,
shuffle=False)
valid_dl = DataLoader(dataset=valid_ds,
batch_size=batch_size,
shuffle=False)
# creating databunch
bunch = DataBunch(train_dl, valid_dl)
# importing the model instance
model_instance = model_library[model_config['instance']]
# init the model
model = model_instance(in_features=in_features,
out_features=len(target_name),
y_limit=y_limit,
**model_config['parameters']).to(device=device)
# init optimizer
optimizer = optim.Adam(model.parameters(),
**experiment_config['optimizer'])
# creating learner instance
logging.info('creating learner instance')
cbs = [
Recoder,
MetricRecorderCallBack(metrics.torch_rmse), ModelCheckpointCallBack,
ProgressBarCallBack
]
learner = Learner(model, optimizer, bunch, callbacks=cbs)
logging.info('training model')
# importing epochs, default is 10
epochs = experiment_config.pop('epochs', 10)
# train the model
learner.fit(epochs, seed=2020)
# avg the last 5 epochs weights
top_models = np.arange(epochs)[-5:]
learner.modelcheckpoint.load_averaged_model(top_models)
logging.info('prediction h0 and h1 models')
# predicting
valid_output = predict_dl(learner.model, valid_dl)
train_output = predict_dl(learner.model, not_shuffle_train_dl)
valid_data[['yhat_t0', 'yhat_t1']] = valid_output['prediction'].numpy()
train_data[['yhat_t0', 'yhat_t1']] = train_output['prediction'].numpy()
# computing metrics
train_error = compute_metrics(train_data, suffix='_train')
valid_error = compute_metrics(valid_data, suffix='_valid')
train_error_period = compute_metrics_per_period(train_data,
suffix='_train')
valid_error_period = compute_metrics_per_period(valid_data,
suffix='_valid')
logging.info('errors')
logging.info(f'{train_error}')
logging.info(f'{valid_error}')
logging.info('period errors')
logging.info(f'{train_error_period}')
logging.info(f'{valid_error_period}')
if eval_mode:
with mlflow.start_run(run_name=experiment):
# saving predictions
train_prediction = train_data.loc[:, default.keep_columns]
train_prediction.to_csv(prediction_path / 'train.csv', index=False)
# saving training progress
learner.metrics_table.to_csv(experiment_path / 'trn_progress.csv',
index=False)
# saving errors
train_error_period.to_csv(experiment_path / 'train_erros.csv',
index=False)
valid_error_period.to_csv(experiment_path / 'valid_erros.csv',
index=False)
# valid_prediction = valid_data.loc[:, default.keep_columns]
valid_data.to_csv(prediction_path / 'valid.csv', index=False)
# saving feature importances if there is aviable
fi = permutation_importance(model=learner.model,
data=valid_data,
features=features,
target=target_name,
score_func=metrics.rmse)
if fi_threshold is None:
fi.to_csv(experiment_path / 'fi_h0.csv', index=False)
fi.to_csv(experiment_path / 'fi_h1.csv', index=False)
# saving to mlflow
# saving metrics
mlflow.log_metrics(train_error)
mlflow.log_metrics(valid_error)
# saving model parameters
mlflow.log_params(model_config['parameters'])
mlflow.log_params(experiment_config['optimizer'])
mlflow.log_params({
'epochs': epochs,
'use_sigmoid': use_sigmoid,
'fi_threshold': fi_threshold,
'in_features': in_features
})
tags = {
'use_sample': use_sample,
'model_instance': model_config['instance'],
'experiment': experiment
}
if message is not None:
tags['message'] = message
mlflow.set_tags(tags)
else:
# creating model path
test_error = calculate_error_on_test(train_data)
test_error = pd.DataFrame([test_error])
test_error.to_csv(experiment_path / 'check_test_error.csv',
index=False)
model_path.mkdir(exist_ok=True, parents=True)
joblib.dump(learner.model, model_path / 'model_h0.pkl')
joblib.dump(pipeline, model_path / 'pipeline.pkl')
joblib.dump(features, model_path / 'features.pkl')
def main(experiment_path: str,
eval_mode: bool = True,
use_sample: bool = False,
test_frac: float = 0.2,
message: str = None,
fi_threshold: float = None):
"""
A function to train or validate an Experiment
# Parameters
experiment_path: `str`
A path to the folder's experiment config file.
the config file must be named config.yml and
it must contain the following keys:
model: `str`
the path to the model config file
pipeline: `str`
the path to the pipeline config file
eval_mode: `bool`, optional (default=True)
if False, the model will be train using all the data available,
otherwise, the trained model will be use for inference.
use_sample: `bool`, optional (default=False)
if True, we will use only a sample from the dataset.
to use it, before execute the make_sample.py file
to create this sample dataset.
test_frac: `float`, optional (default=0.2)
if eval_mode is True, the size of the valid dataset
will be the {test_frac}% of the main dataset.
message: `str`, optional (default=None)
we use mlflow to keep track of all parameters and errors of each experiment,
this parameter will register any string you pass into the experiment record in mlflow.
fi_threshold `float`, optional (default=None)
if already exists a feature importance file, this value will be use for filtering
the features that has greater importance values than {fi_threshold}.
"""
# getting experiment name
experiment = os.path.basename(experiment_path)
logging.info(f'running {experiment}')
logging.info(f'eval_mode={eval_mode}, use_sample={use_sample}')
logging.info('reading config file')
# creating experiment path and loading experiment config file
experiment_path = Path(experiment_path)
config = load_data.read_config_file('./config/config.yml')
experiment_config = load_data.read_config_file(experiment_path /
'config.yml')
# reading experiment's model and pipeline config file
pipeline_config = load_data.read_config_file(experiment_config['pipeline'])
model_config = load_data.read_config_file(experiment_config['model'])
directories = config['directories']
# getting the data path
processed_path = Path(directories['processed'])
# creating a prediction folder to save prediction after training
prediction_path = experiment_path / 'prediction'
prediction_path.mkdir(exist_ok=True, parents=True)
# creating a model path to save models after training
model_path = experiment_path / 'models'
# reading preprocessed data
filename = ('fe' if not use_sample else 'fe_sample')
logging.info('reading training data')
data = load_data.read_feather(processed_path / f'{filename}.feather')
logging.info('splitting dataset')
train_idx, valid_idx = load_data.split_train_data(data,
test_frac=test_frac,
eval_mode=eval_mode)
train_data = data.loc[train_idx, :]
valid_data = data.loc[valid_idx, :]
train_data.reset_index(drop=True, inplace=True)
valid_data.reset_index(drop=True, inplace=True)
# importing pipeline
logging.info('building pipeline')
pipeline = build_pipeline(pipeline_config)
logging.info(f'{pipeline}')
# fit pipeline
logging.info('training pipeline')
pipeline.fit(train_data)
# transform both training and valid dataset
logging.info('transforming datasets')
train_data = pipeline.transform(train_data)
valid_data = pipeline.transform(valid_data)
# loading the features to train our model
# if exists a feature importance file
# we can use it to train our model only with revelant features
features = load_data.get_features(train_data,
experiment_path=experiment_path,
fi_threshold=fi_threshold,
ignore_features=default.ignore_features)
in_features = len(features)
logging.info(f'modeling using {len(features)} features')
logging.info(f'{features[:30]}')
# importing model instance
model_instance = model_library[model_config['instance']]
logging.info('training horizon 0 model')
# training model for horizon 0
model_h0 = model_instance(**model_config['parameters'])
model_h0.fit(train_data.loc[:, features], train_data.loc[:, 't0'])
logging.info('training horizon 1 model')
# training model for horizon 1
model_h1 = model_instance(**model_config['parameters'])
model_h1.fit(train_data.loc[:, features], train_data.loc[:, 't1'])
logging.info('prediction h0 and h1 models')
# predicting
train_data['yhat_t0'] = model_h0.predict(train_data.loc[:, features])
train_data['yhat_t1'] = model_h1.predict(train_data.loc[:, features])
valid_data['yhat_t0'] = model_h0.predict(valid_data.loc[:, features])
valid_data['yhat_t1'] = model_h1.predict(valid_data.loc[:, features])
# compute errors
train_error = compute_metrics(train_data, suffix='_train')
valid_error = compute_metrics(valid_data, suffix='_valid')
train_error_period = compute_metrics_per_period(train_data,
suffix='_train')
valid_error_period = compute_metrics_per_period(valid_data,
suffix='_valid')
logging.info('errors')
logging.info(f'{train_error}')
logging.info(f'{valid_error}')
logging.info('period errors')
logging.info(f'{train_error_period}')
logging.info(f'{valid_error_period}')
if eval_mode:
with mlflow.start_run(run_name=experiment):
# saving predictions
train_prediction = train_data.loc[:, default.keep_columns]
train_prediction.to_csv(prediction_path / 'train.csv', index=False)
# saving errors
train_error_period.to_csv(experiment_path / 'train_erros.csv',
index=False)
valid_error_period.to_csv(experiment_path / 'valid_erros.csv',
index=False)
# valid_prediction = valid_data.loc[:, default.keep_columns]
valid_data.to_csv(prediction_path / 'valid.csv', index=False)
# saving feature importances if there is aviable
fi_h0 = feature_importances(model_h0, features)
fi_h1 = feature_importances(model_h1, features)
if (fi_h0
is not None) and (fi_h1
is not None) and (fi_threshold is None):
fi_h0.to_csv(experiment_path / 'fi_h0.csv', index=False)
fi_h1.to_csv(experiment_path / 'fi_h1.csv', index=False)
# saving to mlflow
# saving metrics
mlflow.log_metrics(train_error)
mlflow.log_metrics(valid_error)
mlflow.log_params({
'fi_threshold': fi_threshold,
'in_features': in_features
})
# saving model parameters
mlflow.log_params(model_config['parameters'])
tags = {
'use_sample': use_sample,
'model_instance': model_config['instance'],
'experiment': experiment
}
if message is not None:
tags['message'] = message
mlflow.set_tags(tags)
else:
# creating model path
test_error = calculate_error_on_test(train_data)
test_error = pd.DataFrame([test_error])
test_error.to_csv(experiment_path / 'check_test_error.csv',
index=False)
model_path.mkdir(exist_ok=True, parents=True)
joblib.dump(model_h0, model_path / 'model_h0.pkl')
joblib.dump(model_h1, model_path / 'model_h1.pkl')
joblib.dump(pipeline, model_path / 'pipeline.pkl')
joblib.dump(features, model_path / 'features.pkl')