def train_aoi(output_path, data_fpath, labels_fpath):
r"""
CommandLine:
python -m ibeis_cnn.train --test-train_aoi
Example:
>>> # DISABLE_DOCTEST
>>> from ibeis_cnn.train import * # NOQA
>>> result = train_aoi()
>>> print(result)
"""
era_size = 256
batch_size = 16
max_epochs = era_size * 16
hyperparams = ut.argparse_dict({
'era_size': era_size,
'learning_rate': .01,
'rate_schedule': 0.75,
'momentum': .9,
'weight_decay': 0.0001,
'augment_on': True,
'augment_on_validate': True,
'whiten_on': False,
'max_epochs': max_epochs,
'stopping_patience': max_epochs,
'class_weight': None,
})
ut.colorprint('[netrun] Ensuring Dataset', 'yellow')
dataset = ingest_data.get_numpy_dataset2('aoi', data_fpath, labels_fpath,
output_path)
X_train, y_train = dataset.subset('train')
X_valid, y_valid = dataset.subset('valid')
print('dataset.training_dpath = %r' % (dataset.training_dpath, ))
input_shape = (
batch_size,
dataset.data_shape[0] + 4,
)
ut.colorprint('[netrun] Architecture Specification', 'yellow')
model = AoIModel(input_shape=input_shape,
training_dpath=dataset.training_dpath,
**hyperparams)
ut.colorprint('[netrun] Initialize architecture', 'yellow')
model.output_dims = 1
model.input_shape = (
None,
dataset.data_shape[0] + 4,
)
model.batch_size = batch_size
model.init_arch()
ut.colorprint('[netrun] * Initializing new weights', 'lightgray')
if model.has_saved_state():
model.load_model_state()
ut.colorprint('[netrun] Training Requested', 'yellow')
# parse training arguments
config = ut.argparse_dict(
dict(
era_size=era_size,
max_epochs=max_epochs,
show_confusion=False,
))
model.monitor_config.update(**config)
print('\n[netrun] Model Info')
model.print_layer_info()
ut.colorprint('[netrun] Begin training', 'yellow')
model.fit(X_train, y_train, X_valid=X_valid, y_valid=y_valid)
model_path = model.save_model_state()
return model_path
def train_classifier(output_path, data_fpath, labels_fpath):
r"""
CommandLine:
python -m ibeis_cnn.train --test-train_classifier
Example:
>>> # DISABLE_DOCTEST
>>> from ibeis_cnn.train import * # NOQA
>>> result = train_classifier()
>>> print(result)
"""
era_size = 16
max_epochs = 256
hyperparams = ut.argparse_dict(
{
'era_size' : era_size,
'batch_size' : 128,
'learning_rate' : .01,
'rate_schedule' : 0.75,
'momentum' : .9,
'weight_decay' : 0.0001,
'augment_on' : True,
'whiten_on' : True,
'max_epochs' : max_epochs,
}
)
ut.colorprint('[netrun] Ensuring Dataset', 'yellow')
dataset = ingest_data.get_numpy_dataset2('classifier', data_fpath, labels_fpath, output_path)
X_train, y_train = dataset.subset('train')
X_valid, y_valid = dataset.subset('valid')
print('dataset.training_dpath = %r' % (dataset.training_dpath,))
ut.colorprint('[netrun] Architecture Specification', 'yellow')
model = ClassifierModel(
data_shape=dataset.data_shape,
training_dpath=dataset.training_dpath,
**hyperparams)
ut.colorprint('[netrun] Init encoder and convert labels', 'yellow')
if hasattr(model, 'init_encoder'):
model.init_encoder(y_train)
ut.colorprint('[netrun] Initialize archchitecture', 'yellow')
model.init_arch()
ut.colorprint('[netrun] * Initializing new weights', 'lightgray')
if model.has_saved_state():
model.load_model_state()
# else:
# model.reinit_weights()
# ut.colorprint('[netrun] Need to initialize training state', 'yellow')
# X_train, y_train = dataset.subset('train')
# model.ensure_data_params(X_train, y_train)
ut.colorprint('[netrun] Training Requested', 'yellow')
# parse training arguments
config = ut.argparse_dict(dict(
monitor=True,
monitor_updates=True,
show_confusion=True,
era_size=era_size,
max_epochs=max_epochs,
))
model.monitor_config.update(**config)
if getattr(model, 'encoder', None) is not None:
class_list = list(model.encoder.classes_)
y_train = np.array([class_list.index(_) for _ in y_train ])
y_valid = np.array([class_list.index(_) for _ in y_valid ])
print('\n[netrun] Model Info')
model.print_layer_info()
ut.colorprint('[netrun] Begin training', 'yellow')
model.fit(X_train, y_train, X_valid=X_valid, y_valid=y_valid)
model_path = model.save_model_state()
return model_path