def get_trained_model(args):
""" Returns a model with loaded weights. """
model = get_frontend(input_width, input_height)
if has_context_module:
model = add_context(model)
model = add_softmax(model)
def load_tf_weights():
""" Load pretrained weights converted from Caffe to TF. """
# 'latin1' enables loading .npy files created with python2
weights_data = np.load(args.weights_path, encoding='latin1').item()
for layer in model.layers:
if layer.name in weights_data.keys():
layer_weights = weights_data[layer.name]
layer.set_weights(
(layer_weights['weights'], layer_weights['biases']))
def load_keras_weights():
""" Load a Keras checkpoint. """
model.load_weights(args.weights_path)
if args.weights_path.endswith('.npy'):
load_tf_weights()
elif args.weights_path.endswith('.h5'):
load_keras_weights()
else:
raise Exception("Unknown weights format.")
return model
def train(train_list_fname='benchmark_RELEASE/dataset/train.txt',
val_list_fname='benchmark_RELEASE/dataset/val.txt',
img_root='benchmark_RELEASE/dataset/img',
mask_root='benchmark_RELEASE/dataset/pngs',
weights_path='converted/dilation8_pascal_voc.npy',
batch_size=2,
learning_rate=0.0001):
# Create image generators for the training and validation sets. Validation has
# no data augmentation.
transformer_train = RandomTransformer(horizontal_flip=True,
vertical_flip=True)
datagen_train = SegmentationDataGenerator(transformer_train)
transformer_val = RandomTransformer(horizontal_flip=False,
vertical_flip=False)
datagen_val = SegmentationDataGenerator(transformer_val)
train_desc = 'lr{:.0e}-bs{:03d}'.format(learning_rate, batch_size)
checkpoints_folder = 'trained/' + train_desc
try:
os.makedirs(checkpoints_folder)
except OSError:
shutil.rmtree(checkpoints_folder, ignore_errors=True)
os.makedirs(checkpoints_folder)
model_checkpoint = callbacks.ModelCheckpoint(
checkpoints_folder + '/ep{epoch:02d}-vl{val_loss:.4f}.hdf5',
monitor='loss')
tensorboard_cback = callbacks.TensorBoard(
log_dir='{}/tboard'.format(checkpoints_folder),
histogram_freq=0,
write_graph=False,
write_images=False)
csv_log_cback = callbacks.CSVLogger(
'{}/history.log'.format(checkpoints_folder))
reduce_lr_cback = callbacks.ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
patience=5,
verbose=1,
min_lr=0.05 * learning_rate)
model = add_softmax(get_frontend(500, 500))
load_weights(model, weights_path)
model.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizers.SGD(lr=learning_rate, momentum=0.9),
metrics=['accuracy'])
# Build absolute image paths
def build_abs_paths(basenames):
img_fnames = [os.path.join(img_root, f) + '.jpg' for f in basenames]
mask_fnames = [os.path.join(mask_root, f) + '.png' for f in basenames]
return img_fnames, mask_fnames
train_basenames = [l.strip() for l in open(train_list_fname).readlines()]
val_basenames = [l.strip() for l in open(val_list_fname).readlines()][:500]
train_img_fnames, train_mask_fnames = build_abs_paths(train_basenames)
val_img_fnames, val_mask_fnames = build_abs_paths(val_basenames)
skipped_report_cback = callbacks.LambdaCallback(
on_epoch_end=lambda a, b: open(
'{}/skipped.txt'.format(checkpoints_folder), 'a').write(
'{}\n'.format(datagen_train.skipped_count)))
print("Inside build_abs_paths")
model.fit_generator(
datagen_train.flow_from_list(train_img_fnames,
train_mask_fnames,
shuffle=True,
batch_size=batch_size,
img_target_size=(500, 500),
mask_target_size=(16, 16)),
samples_per_epoch=len(train_basenames),
nb_epoch=20,
validation_data=datagen_val.flow_from_list(val_img_fnames,
val_mask_fnames,
batch_size=8,
img_target_size=(500, 500),
mask_target_size=(16, 16)),
nb_val_samples=len(val_basenames),
callbacks=[
model_checkpoint,
tensorboard_cback,
csv_log_cback,
reduce_lr_cback,
skipped_report_cback,
])
def train():
global train_list_fname
global val_list_fname
global img_root
global mask_root
global weights_path
global batch_size
global learning_rate
global modeltype
train_data_gen = SegmentationDataGenerator(
RandomTransformer(horizontal_flip=True, vertical_flip=True))
val_data_gen = SegmentationDataGenerator(
RandomTransformer(horizontal_flip=True, vertical_flip=True))
trained_log = '{}-lr{:.0e}-bs{:03d}'.format(
time.strftime("%Y-%m-%d %H:%M"), learning_rate, batch_size)
checkpoints_folder = 'trained_log/' + trained_log
try:
os.makedirs(checkpoints_folder)
except OSError:
shutil.rmtree(checkpoints_folder, ignore_errors=True)
os.makedirs(checkpoints_folder)
model_checkpoint = callbacks.ModelCheckpoint(
checkpoints_folder + '/ep{epoch:02d}-vl{val_loss:.4f}.hdf5',
monitor='loss')
model_tensorboard = callbacks.TensorBoard(
log_dir='{}/tboard'.format(checkpoints_folder),
histogram_freq=0,
write_graph=False,
write_images=False)
model_csvlogger = callbacks.CSVLogger(
'{}/history.log'.format(checkpoints_folder))
model_reducelr = callbacks.ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
patience=5,
verbose=1,
min_lr=0.05 * learning_rate)
model = add_softmax(dilated_frontend(500, 500))
#load_weights(model, weights_path)
model.compile(optimizer=optimizers.SGD(lr=learning_rate, momentum=0.9),
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
train_basenames = [l.strip() for l in open(train_list_fname).readlines()]
val_basenames = [l.strip() for l in open(val_list_fname).readlines()][:500]
train_img_fnames, train_mask_fnames = build_abs_paths(train_basenames)
val_img_fnames, val_mask_fnames = build_abs_paths(val_basenames)
model_skipped = callbacks.LambdaCallback(on_epoch_end=lambda a, b: open(
'{}/skipped.txt'.format(checkpoints_folder), 'a').write('{}\n'.format(
train_data_gen.skipped_count)))
model.fit_generator(
train_data_gen.flow_from_list(train_img_fnames,
train_mask_fnames,
shuffle=True,
batch_size=batch_size,
img_target_size=(500, 500),
mask_target_size=(16, 16)),
steps_per_epoch=(len(train_basenames) / batch_size),
epochs=50,
validation_data=val_data_gen.flow_from_list(val_img_fnames,
val_mask_fnames,
batch_size=8,
img_target_size=(500, 500),
mask_target_size=(16, 16)),
validation_steps=(len(val_basenames) / 8),
callbacks=[
model_checkpoint, model_tensorboard, model_csvlogger,
model_reducelr, model_skipped
])