def load_model():
text_model = DenseNet(classes=n_classes,
input_shape=image_shape,
depth=40,
growth_rate=12,
bottleneck=True,
reduction=0.5,
dropout_rate=0.0,
weight_decay=1e-4)
text_model.load_weights(text_model_weight)
return text_model
base_model = DenseNet(classes=n_classes,
input_shape=image_shape,
depth=args.depth,
growth_rate=args.growth_rate,
bottleneck=args.bc_mode,
reduction=args.reduction,
dropout_rate=1.0 - args.keep_prob,
weight_decay=args.weight_decay)
if args.train:
batch_size *= n_gpus
if os.path.exists("saves/%s.weight" % model_identifier):
print("Loading model...")
base_model.load_weights("saves/%s.weight" % model_identifier,
by_name=True)
if n_gpus > 1:
model = multi_gpu_model(base_model, n_gpus)
else:
model = base_model
def loss_func(y_true, y_pred):
l2_loss = tf.add_n(
[tf.nn.l2_loss(var) for var in model.trainable_weights])
return categorical_crossentropy(y_true, y_pred) + l2_loss * 1e-4
optimizer = SGD(lr=initial_learning_rate,
clipvalue=0.5,
momentum=0.9,
decay=1e-4,
# Start Tensorflow session
with tf.Session() as sess:
# Initialize all variables
ckpt = tf.train.get_checkpoint_state(checkpoint_path)
if ckpt and ckpt.model_checkpoint_path:
pass
else:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# Load the pretrained weights into the non-trainable layer
if pretrain and mode == 'alexnet':
model.load_initial_weights(sess)
elif pretrain and mode == 'densenet':
model_op.load_weights()
elif pretrain and mode == 'resnet':
raise ValueError("Invalid pretrain for resnet")
if ckpt and ckpt.model_checkpoint_path:
print('Training from saved model ...')
saver.restore(sess, ckpt.model_checkpoint_path)
graph = tf.get_default_graph()
# Add the model graph to TensorBoard
writer.add_graph(sess.graph)
print("{} Start training...".format(datetime.now()))
print("{} Open Tensorboard at --logdir {}".format(datetime.now(),
filewriter_path))