def main():
config_path = args.conf
initial_weights = args.weights
with open(config_path) as config_buffer:
config = json.loads(config_buffer.read())
train_set, valid_set, classes = data.create_training_instances(config['train']['train_folder'],
None,
config['train']['cache_name'],
config['model']['labels'])
num_classes = len(classes)
print('Readed {} classes: {}'.format(num_classes, classes))
train_generator = gen.BatchGenerator(
instances=train_set,
labels=classes,
batch_size=config['train']['batch_size'],
input_sz=config['model']['infer_shape'],
shuffle=True,
norm=data.normalize
)
valid_generator = gen.BatchGenerator(
instances=valid_set,
labels=classes,
batch_size=config['train']['batch_size'],
input_sz=config['model']['infer_shape'],
norm=data.normalize,
infer=True
)
early_stop = EarlyStopping(
monitor='val_loss',
min_delta=0,
patience=20,
mode='min',
verbose=1
)
reduce_on_plateau = ReduceLROnPlateau(
monitor='val_loss',
factor=0.5,
patience=5,
verbose=1,
mode='min',
min_delta=0.01,
cooldown=0,
min_lr=0
)
net_input_shape = (config['model']['infer_shape'][0],
config['model']['infer_shape'][1],
3)
train_model = models.create(
base_name=config['model']['base'],
num_classes=num_classes,
input_shape=net_input_shape)
if initial_weights:
train_model.load_weights(initial_weights)
print(train_model.summary())
# plot_model(train_model, to_file='images/MobileNetv2.png', show_shapes=True)
optimizer = Adam(lr=config['train']['learning_rate'], clipnorm=0.001)
train_model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])
checkpoint_name = utils.get_checkpoint_name(config)
utils.makedirs_4_file(checkpoint_name)
static_chk_name = utils.get_static_checkpoint_name(config)
utils.makedirs_4_file(static_chk_name)
checkpoint_vloss = cbs.CustomModelCheckpoint(
model_to_save=train_model,
filepath=checkpoint_name,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min',
period=1
)
neptune_mon = cbs.NeptuneMonitor(
monitoring=['loss', 'val_loss', 'accuracy', 'val_accuracy'],
neptune=neptune
)
chk_static = ModelCheckpoint(
filepath=static_chk_name,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min',
period=1
)
callbacks = [early_stop, reduce_on_plateau, checkpoint_vloss, neptune_mon, chk_static]
### NEPTUNE ###
sources_to_upload = [
'models.py',
'config.json'
]
params = {
'infer_size': "H{}xW{}".format(*config['model']['infer_shape']),
'classes': config['model']['labels'],
}
neptune.create_experiment(
name=utils.get_neptune_name(config),
upload_stdout=False,
upload_source_files=sources_to_upload,
params=params
)
### NEPTUNE ###
hist = train_model.fit_generator(
generator=train_generator,
steps_per_epoch=len(train_generator) * config['train']['train_times'],
validation_data=valid_generator,
validation_steps=len(valid_generator) * config['valid']['valid_times'],
epochs=config['train']['nb_epochs'],
verbose=2 if config['train']['debug'] else 1,
callbacks=callbacks,
workers=multiprocessing.cpu_count(),
max_queue_size=100
)
neptune.send_artifact(static_chk_name)
neptune.send_artifact('config.json')
files_path, labels = get_path_labels(data_path)
for i in range(len(labels)):
print('{} : {}'.format(files_path[i], labels[i]))
# パスとラベルから学習、テスト、評価データを作成
X_train, X_test, y_train, y_test = train_test_split(files_path,
labels,
train_size=0.8)
X_train, X_val, y_train, y_val = train_test_split(X_train,
y_train,
train_size=0.8)
# ジェネレータを作成
train_batch_generator = generator.BatchGenerator(X_train, y_train,
batch_size, sample_rate,
sample_length, threshold)
test_batch_generator = generator.BatchGenerator(X_val, y_val, batch_size,
sample_rate, sample_length,
threshold)
# モデル構築
dnn = CNN()
model = dnn.build_model()
# 学習
fit_history = model.fit_generator(
train_batch_generator,
epochs=epoch,
verbose=1,
steps_per_epoch=train_batch_generator.batches_per_epoch,
def main():
config_path = args.conf
initial_weights = args.weights
with open(config_path) as config_buffer:
config = json.loads(config_buffer.read())
train_set, valid_set = data.create_training_instances(config['train']['train_folder'],
config['train']['train_masks'],
config['valid']['valid_folder'],
config['valid']['valid_masks'],
config['train']['cache_name'])
train_generator = gen.BatchGenerator(
instances = train_set,
batch_size = config['train']['batch_size'],
input_sz = config['model']['input_shape'],
shuffle = True,
jitter = 0.3,
norm = data.normalize,
downsample = 2
)
valid_generator = gen.BatchGenerator(
instances = valid_set,
batch_size = config['train']['batch_size'],
input_sz = config['model']['input_shape'],
norm = data.normalize,
infer = True,
downsample = 2
)
early_stop = EarlyStopping(
monitor = 'val_loss',
min_delta = 0,
patience = 100,
mode = 'min',
verbose = 1
)
reduce_on_plateau = ReduceLROnPlateau(
monitor = 'loss',
factor = 0.5,
patience = 5,
verbose = 1,
mode = 'min',
min_delta= 0.01,
cooldown = 0,
min_lr = 0
)
# Swapped as net input -> [H x W x C]
net_input_shape = (config['model']['input_shape'][1],
config['model']['input_shape'][0],
3)
train_model = models.create(
base = config['model']['base'],
input_shape = net_input_shape)
if initial_weights:
train_model.load_weights(initial_weights)
model_render_file = 'images/{}.png'.format(config['model']['base'])
if not os.path.isdir(os.path.dirname(model_render_file)):
os.makedirs(os.path.dirname(model_render_file))
plot_model(train_model, to_file=model_render_file, show_shapes=True)
# print_summary(train_model)
optimizer = Adam(lr=config['train']['learning_rate'], clipnorm=0.001)
# optimizer = SGD(lr=config['train']['learning_rate'], clipnorm=0.001)
train_model.compile(loss=models.result_loss, optimizer=optimizer,
metrics=[models.iou_loss, models.dice_coef_loss, models.pixelwise_crossentropy])
chk_name = config['train']['saved_weights_name']
chk_root, chk_ext = os.path.splitext(chk_name)
checkpoint_vloss = CustomModelCheckpoint(
model_to_save = train_model,
filepath = chk_root+'_ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}'+chk_ext,
monitor = 'val_loss',
verbose = 1,
save_best_only = True,
mode = 'min',
period = 1
)
if chk_name:
if not os.path.isdir(os.path.dirname(chk_name)):
os.makedirs(os.path.dirname(chk_name))
callbacks = [early_stop, reduce_on_plateau, checkpoint_vloss]
hist = train_model.fit_generator(
generator=train_generator,
steps_per_epoch=len(train_generator) * config['train']['train_times'],
validation_data=valid_generator,
validation_steps=len(valid_generator) * config['valid']['valid_times'],
epochs=config['train']['nb_epochs'],
verbose=2 if config['train']['debug'] else 1,
callbacks=callbacks,
workers=os.cpu_count(),
max_queue_size=100
)
if not os.path.exists('model'):
os.makedirs('model')
df = pd.DataFrame.from_dict(hist.history)
df.to_csv('model/hist.csv', encoding='utf-8', index=False)