def warm_up(model, data_iterator, log_file):
learning_rate_schedules = optimizers.schedules.PolynomialDecay(
initial_learning_rate=c.minimum_learning_rate,
decay_steps=c.warm_iterations,
end_learning_rate=c.initial_learning_rate)
optimizer = optimizers.SGD(learning_rate=learning_rate_schedules,
momentum=0.9,
nesterov=True)
sum_ce = 0
sum_correct_num = 0
for i in tqdm(range(c.iterations_per_epoch)):
images, labels = data_iterator.next()
ce, prediction = train_step(model, images, labels, optimizer)
correct_num = correct_num_batch(labels, prediction)
sum_ce += ce * c.batch_size
sum_correct_num += correct_num
print('ce: {:.4f}, accuracy: {:.4f}, l2 loss: {:.4f}'.format(
ce, correct_num / c.batch_size, l2_loss(model)))
log_file.write(
'warm up: cross entropy loss: {:.4f}, l2 loss: {:.4f}, accuracy: {:.4f}\n'
.format(sum_ce / c.train_num, l2_loss(model),
sum_correct_num / c.train_num))
def train_step(model, images, labels, optimizer):
with tf.GradientTape() as tape:
prediction = model(images, training=True)
ce = cross_entropy_batch(labels, prediction, label_smoothing=c.label_smoothing)
l2 = l2_loss(model)
loss = ce + l2
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
return ce, prediction
def train(model, data_iterator, optimizer, log_file):
sum_ce = 0
sum_correct_num = 0
for i in tqdm(range(c.iterations_per_epoch)):
images, labels = data_iterator.next()
ce, prediction = train_step(model, images, labels, optimizer)
correct_num = correct_num_batch(labels, prediction)
sum_ce += ce * c.batch_size
sum_correct_num += correct_num
print('ce: {:.4f}, accuracy: {:.4f}, l2 loss: {:.4f}'.format(
ce, correct_num / c.batch_size, l2_loss(model)))
log_file.write(
'train: cross entropy loss: {:.4f}, l2 loss: {:.4f}, accuracy: {:.4f}\n'
.format(sum_ce / c.train_num, l2_loss(model),
sum_correct_num / c.train_num))
def test(model, log_file):
data_iterator = test_10_crop_iterator()
sum_ce = 0
sum_correct_num = 0
for i in tqdm(range(c.test_num)):
images, labels = data_iterator.next()
ce, prediction = test_step(model, images, labels)
sum_ce += ce
if np.argmax(prediction) == np.argmax(labels):
sum_correct_num += 1
# print('ce: {:.4f}, accuracy: {:.4f}'.format(ce, sum_correct_num / (i + 1)))
log_file.write(
'test: cross entropy loss: {:.4f}, l2 loss: {:.4f}, accuracy: {:.4f}\n'
.format(sum_ce / c.test_num, l2_loss(model),
sum_correct_num / c.test_num))
def test(model, log_file):
data_iterator = test_iterator()
sum_ce = 0
sum_correct_num = 0
for i in tqdm(range(c.test_iterations)):
images, labels = data_iterator.next()
ce, prediction = test_step(model, images, labels)
correct_num = correct_num_batch(labels, prediction)
sum_ce += ce * c.batch_size
sum_correct_num += correct_num
print('ce: {:.4f}, accuracy: {:.4f}'.format(ce, correct_num /
c.batch_size))
log_file.write(
'test: cross entropy loss: {:.4f}, l2 loss: {:.4f}, accuracy: {:.4f}\n'
.format(sum_ce / c.test_num, l2_loss(model),
sum_correct_num / c.test_num))
ce, correct_num / c.batch_size, l2_loss(model)))
log_file.write(
'warm up: cross entropy loss: {:.4f}, l2 loss: {:.4f}, accuracy: {:.4f}\n'
.format(sum_ce / c.train_num, l2_loss(model),
sum_correct_num / c.train_num))
if __name__ == '__main__':
# gpu config
physical_devices = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(device=physical_devices[0],
enable=True)
# load data
train_data_iterator = train_iterator()
# get model
model = ResNet(50)
# show
model.build(input_shape=(None, ) + c.input_shape)
model.summary()
print('initial l2 loss:{:.4f}'.format(l2_loss(model)))
with open(c.log_file, 'a') as f:
warm_up(model, train_data_iterator, f)
test(model, f)
model.save_weights(c.save_weight_file, save_format='h5')
if __name__ == '__main__':
# gpu config
physical_devices = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(device=physical_devices[0], enable=True)
# load data
train_data_iterator = train_iterator()
# get model
# model = ResNet(50)
model = ResNet_v2(50)
# show
model.build(input_shape=(None,) + c.input_shape)
model.summary()
print('initial l2 loss:{:.4f}'.format(l2_loss(model)))
# load pretrain
if c.load_weight_file is not None:
model.load_weights(c.load_weight_file)
print('pretrain weight l2 loss:{:.4f}'.format(l2_loss(model)))
# train
with open(c.log_file, 'a') as f:
warm_up(model, train_data_iterator, f)
test(model, f)
model.save_weights(c.save_weight_file, save_format='h5')
learning_rate_schedules = tf.keras.experimental.CosineDecay(initial_learning_rate=c.initial_learning_rate,
decay_steps=c.epoch_num * c.iterations_per_epoch,
alpha=c.minimum_learning_rate)