def pick_model_weights(model: keras.models.Model,
dataset_name,
path='../output/Models/Weights'):
model_names = []
for path in glob.glob('{}/{}_*.h5'.format(path, dataset_name)):
model_names.append(path)
print('Please enter you model number form list below:')
for i, path in enumerate(model_names):
print('{}. {}'.format(i + 1, path))
model_number = int(input('?')) - 1
model.load_weights(model_names[model_number])
return model
def train(model: keras.models.Model,
optimizer: dict,
save_path: str,
train_dir: str,
valid_dir: str,
batch_size: int = 32,
epochs: int = 10,
samples_per_epoch=1000,
pretrained=None,
augment: bool = True,
weight_mode=None,
verbose=0,
**kwargs):
""" Trains the model with the given configurations. """
shape = model.input_shape[1:3]
optimizer_cpy = optimizer.copy()
shared_gen_args = {
'rescale': 1. / 255, # to preserve the rgb palette
}
train_gen_args = {}
if augment:
train_gen_args = {
"fill_mode": 'reflect',
'horizontal_flip': True,
'vertical_flip': True,
'width_shift_range': .15,
'height_shift_range': .15,
'shear_range': .5,
'rotation_range': 45,
'zoom_range': .2,
}
gen = IDG(**{**shared_gen_args, **train_gen_args})
gen = gen.flow_from_directory(train_dir,
target_size=shape,
batch_size=batch_size,
seed=SEED)
val_count = len(
glob(os.path.join(valid_dir, '**', '*.jpg'), recursive=True))
valid_gen = IDG(**shared_gen_args)
optim = getattr(keras.optimizers, optimizer['name'])
if optimizer.pop('name') != 'sgd':
optimizer.pop('nesterov')
schedule = optimizer.pop('schedule')
if schedule == 'decay' and 'lr' in optimizer.keys():
initial_lr = optimizer.pop('lr')
else:
initial_lr = 0.01
optim = optim(**optimizer)
callbacks = [
utils.checkpoint(save_path),
utils.csv_logger(save_path),
]
if pretrained is not None:
if not os.path.exists(pretrained):
raise FileNotFoundError()
model.load_weights(pretrained, by_name=False)
if verbose == 1:
print("Loaded weights from {}".format(pretrained))
if optimizer_cpy['name'] == 'sgd':
if schedule == 'decay':
callbacks.append(utils.step_decay(epochs, initial_lr=initial_lr))
elif schedule == 'big_drop':
callbacks.append(utils.constant_schedule())
model.compile(optim,
loss='categorical_crossentropy',
metrics=['accuracy', top3_acc])
create_xml_description(save=os.path.join(save_path, 'model_config.xml'),
title=model.name,
epochs=epochs,
batch_size=batch_size,
samples_per_epoch=samples_per_epoch,
augmentations=augment,
schedule=schedule,
optimizer=optimizer_cpy,
**kwargs)
if weight_mode:
class_weights = [[key, value] for key, value in weight_mode.items()]
filen = os.path.join(save_path, 'class_weights.npy')
np.save(filen, class_weights)
h = None # has to be initialized here, so we can reference it later
try:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
h = model.fit_generator(
gen,
steps_per_epoch=samples_per_epoch / batch_size,
epochs=epochs,
validation_data=valid_gen.flow_from_directory(
valid_dir,
target_size=shape,
batch_size=batch_size,
seed=SEED),
validation_steps=val_count / batch_size,
callbacks=callbacks,
class_weight=weight_mode,
verbose=2)
except KeyboardInterrupt:
save_results(verbose=1, save_path=save_path, model=model, hist=h)
return
save_results(verbose=1, save_path=save_path, model=model, hist=h)