save_best_only=True,
save_weights_only=True,
mode='min',
period=1)
csv_logger = CSVLogger(TRAINING_LOG, append=True)
tb = TensorBoard(log_dir=LOGS_DIR,
histogram_freq=0,
write_graph=True,
write_images=False)
callbacks_list = [lrate, checkpoint, csv_logger, tb]
# sgd optimizer with lr multipliers
multisgd = MultiSGD(lr=base_lr,
momentum=momentum,
decay=0.0,
nesterov=False,
lr_mult=lr_mult)
# start training
model.compile(loss=losses,
loss_weights=loss_weights,
optimizer=multisgd,
metrics=["accuracy"])
model.fit_generator(train_di,
steps_per_epoch=train_samples // batch_size,
epochs=max_iter,
callbacks=callbacks_list,
validation_data=val_di,
validation_steps=val_samples // batch_size,
data_shape=(3, 368, 368),
label_shape=(57, 46, 46),
split_point=38,
batch_size=batch_size,
shuffle=True)
x, y1, y2 = di.next()
print("x : ", x.shape)
print("y1 : ", y1.shape)
print("y2 : ", y2.shape)
# as suggested in: https://github.com/fchollet/keras/issues/5920
last_layer_variables = list()
for layer in model.layers:
#print(layer.weights)
if layer.name in ['prediction']:
last_layer_variables.extend(layer.weights)
multisgd = MultiSGD(lr=base_lr,
momentum=momentum,
decay=weight_decay,
nesterov=False,
exception_vars=last_layer_variables,
multiplier=0.1)
model.compile(loss='mean_squared_error',
optimizer=multisgd,
metrics=['accuracy'])
Input(shape=(image_shape[0] // 8, image_shape[1] // 8, 38)),
Input(shape=(image_shape[0] // 8, image_shape[1] // 8, 19)),
Input(shape=(image_shape[0] // 8, image_shape[1] // 8, 38)),
Input(shape=(image_shape[0] // 8, image_shape[1] // 8, 19)),
Input(shape=(image_shape[0] // 8, image_shape[1] // 8, 38)),
Input(shape=(image_shape[0] // 8, image_shape[1] // 8, 19))
]
params = model.trainable_weights
loss = total_eucl_loss(model.output,
labels,
len(labels),
batch_size=batch_size)
# opt = keras.optimizers.SGD(lr=1e-5, momentum=0.9, decay=1e-6, nesterov=True).get_updates(params, [], loss)
opt = MultiSGD(lr=1e-5,
momentum=0.9,
decay=1e-6,
nesterov=True,
lr_mult=optimizer_lr_mult(model)).get_updates(
params, [], loss)
train_model = K.function(model.input + labels + [K.learning_phase()],
[loss], opt)
# define generaters
t_generator = Generator("images/train_dataset_2014.npy")
v_generator = Generator("images/valid_dataset_2014.npy")
# train for model
for epoch in range(epochs):
t_total_coss = 0
t_steps = t_generator.samples_length // batch_size
for step in range(t_steps):