def ___conv4_block(input, k=1, dropout=0.0):
init = input
channel_axis = 1 if K.image_dim_ordering() == 'th' else -1
# Check if input number of filters is same as 64 * k, else
# create convolution2d for this input
if K.image_dim_ordering() == 'th':
if init._keras_shape[1] != 64 * k:
init = Conv2D(64 * k, (1, 1), activation='linear',
padding='same')(init)
else:
if init._keras_shape[-1] != 64 * k:
init = Conv2D(64 * k, (1, 1), activation='linear',
padding='same')(init)
x = Conv2D(64 * k, (3, 3), padding='same')(input)
x = BatchNormalization(axis=channel_axis)(x)
x = Activation('relu')(x)
if dropout > 0.0:
x = Dropout(dropout)(x)
x = Conv2D(64 * k, (3, 3), padding='same')(x)
x = BatchNormalization(axis=channel_axis)(x)
x = Activation('relu')(x)
m = add([init, x])
return m
def generator_unet_upsampling(img_shape,
disc_img_shape,
model_name="generator_unet_upsampling"):
filters_num = 64
axis_num = -1
channels_num = img_shape[-1]
min_s = min(img_shape[:-1])
unet_input = Input(shape=img_shape, name="unet_input")
conv_num = int(np.floor(np.log(min_s) / np.log(2)))
list_filters_num = [filters_num * min(8, (2**i)) for i in range(conv_num)]
# Encoder
first_conv = Conv2D(list_filters_num[0], (3, 3),
strides=(2, 2),
name='unet_conv2D_1',
padding='same')(unet_input)
list_encoder = [first_conv]
for i, f in enumerate(list_filters_num[1:]):
name = 'unet_conv2D_' + str(i + 2)
conv = conv_block_unet(list_encoder[-1], f, name, axis_num)
list_encoder.append(conv)
# prepare decoder filters
list_filters_num = list_filters_num[:
-2][::
-1] # list_filters_num -> [512, 512, 512, 256, 128, 64]
if len(list_filters_num) < conv_num - 1:
list_filters_num.append(filters_num)
# Decoder
first_up_conv = up_conv_block_unet(list_encoder[-1],
list_encoder[-2],
list_filters_num[0],
"unet_upconv2D_1",
axis_num,
dropout=True)
list_decoder = [first_up_conv]
for i, f in enumerate(list_filters_num[1:]):
name = "unet_upconv2D_" + str(i + 2)
if i < 2:
d = True
else:
d = False
up_conv = up_conv_block_unet(list_decoder[-1],
list_encoder[-(i + 3)],
f,
name,
axis_num,
dropout=d)
list_decoder.append(up_conv)
x = Activation('relu')(list_decoder[-1])
x = UpSampling2D(size=(2, 2))(x)
x = Conv2D(disc_img_shape[-1], (3, 3), name="last_conv", padding='same')(x)
x = Activation('tanh')(x)
generator_unet = Model(inputs=[unet_input], outputs=[x])
return generator_unet
def up_conv_block_unet(x, x2, f, name, bn_axis, bn=True, dropout=False):
x = Activation('relu')(x)
x = UpSampling2D(size=(2, 2))(x)
x = Conv2D(f, (3, 3), name=name, padding='same')(x)
if bn: x = BatchNormalization(axis=bn_axis)(x)
if dropout: x = Dropout(0.5)(x)
x = Concatenate(axis=bn_axis)([x, x2])
return x
def __conv1_block(input):
x = Conv2D(16, (3, 3), padding='same')(input)
channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
x = BatchNormalization(axis=channel_axis)(x)
x = Activation('relu')(x)
return x
def conv_block(input, growth_rate, dropout_rate=None, weight_decay=1e-4):
x = BatchNormalization(axis=-1, epsilon=1.1e-5)(input)
x = Activation('relu')(x)
x = Conv2D(growth_rate, (3, 3),
kernel_initializer='he_normal',
padding='same')(x)
if (dropout_rate):
x = Dropout(dropout_rate)(x)
return x
def layers(input_shape):
return [
ZeroPadding2D(padding=3,
input_shape=input_shape,
data_format='channels_first'), # <1>
Conv2D(48, (7, 7), data_format='channels_first'),
Activation('relu'),
ZeroPadding2D(padding=2, data_format='channels_first'), # <2>
Conv2D(32, (5, 5), data_format='channels_first'),
Activation('relu'),
ZeroPadding2D(padding=2, data_format='channels_first'),
Conv2D(32, (5, 5), data_format='channels_first'),
Activation('relu'),
ZeroPadding2D(padding=2, data_format='channels_first'),
Conv2D(32, (5, 5), data_format='channels_first'),
Activation('relu'),
Flatten(),
Dense(512),
Activation('relu'),
]
def dense_cnn(input, nclass):
_dropout_rate = 0.2
_weight_decay = 1e-4
_nb_filter = 64
# conv 64 5*5 s=2
x = Conv2D(_nb_filter, (5, 5),
strides=(2, 2),
kernel_initializer='he_normal',
padding='same',
use_bias=False,
kernel_regularizer=l2(_weight_decay))(input)
# 64 + 8 * 8 = 128
x, _nb_filter = dense_block(x, 8, _nb_filter, 8, None, _weight_decay)
# 128
x, _nb_filter = transition_block(x, 128, _dropout_rate, 2, _weight_decay)
# 128 + 8 * 8 = 192
x, _nb_filter = dense_block(x, 8, _nb_filter, 8, None, _weight_decay)
# 192 -> 128
x, _nb_filter = transition_block(x, 128, _dropout_rate, 2, _weight_decay)
# 128 + 8 * 8 = 192
x, _nb_filter = dense_block(x, 8, _nb_filter, 8, None, _weight_decay)
x = BatchNormalization(axis=-1, epsilon=1.1e-5)(x)
x = Activation('relu')(x)
x = Permute((2, 1, 3), name='permute')(x)
x = keras.TimeDistributed(Flatten(), name='flatten')(x)
y_pred = Dense(nclass, name='out', activation='softmax')(x)
# basemodel = Model(inputs=input, outputs=y_pred)
# basemodel.summary()
return y_pred
def transition_block(input,
nb_filter,
dropout_rate=None,
pooltype=1,
weight_decay=1e-4):
x = BatchNormalization(axis=-1, epsilon=1.1e-5)(input)
x = Activation('relu')(x)
x = Conv2D(nb_filter, (1, 1),
kernel_initializer='he_normal',
padding='same',
use_bias=False,
kernel_regularizer=l2(weight_decay))(x)
if (dropout_rate):
x = Dropout(dropout_rate)(x)
if (pooltype == 2):
x = AveragePooling2D((2, 2), strides=(2, 2))(x)
elif (pooltype == 1):
x = ZeroPadding2D(padding=(0, 1))(x)
x = AveragePooling2D((2, 2), strides=(2, 1))(x)
elif (pooltype == 3):
x = AveragePooling2D((2, 2), strides=(2, 1))(x)
return x, nb_filter
def _main_(args):
config_path = args.conf
with open(config_path) as config_buffer:
config = json.loads(config_buffer.read())
if config['backup']['create_backup']:
config = create_backup(config)
keras.backend.tensorflow_backend.set_session(get_session())
#path for the training and validation dataset
datasetTrainPath = os.path.join(args.folder, "train")
datasetValPath = os.path.join(args.folder, "val")
for folder in [datasetTrainPath, datasetValPath]:
if not os.path.isdir(folder):
raise Exception("{} doesn't exist!".format(folder))
classesTrain = next(os.walk(datasetTrainPath))[1]
classesVal = next(os.walk(datasetValPath))[1]
if not classesVal == classesTrain:
raise Exception(
"The training and validation classes must be the same!")
else:
folders = classesTrain
#training configuration
epochs = config['train']['nb_epochs']
batchSize = config['train']['batch_size']
width = config['model']['input_size_w']
height = config['model']['input_size_h']
depth = 3 if config['model']['gray_mode'] == False else 1
#config keras generators
if len(
folders
) == 2: #if just have 2 classes, the model will have a binary output
classes = 1
else:
classes = len(folders)
#count all samples
imagesTrainPaths = []
imagesValPaths = []
for folder in folders:
imagesTrainPaths += list(
list_images(os.path.join(datasetTrainPath, folder)))
imagesValPaths += list(
list_images(os.path.join(datasetValPath, folder)))
generator_config = {
'IMAGE_H': height,
'IMAGE_W': width,
'IMAGE_C': depth,
'BATCH_SIZE': batchSize
}
#callbacks
model_name = config['train']['saved_weights_name']
checkPointSaverBest = ModelCheckpoint(model_name,
monitor='val_acc',
verbose=1,
save_best_only=True,
save_weights_only=False,
mode='auto',
period=1)
ckp_model_name = os.path.splitext(model_name)[1] + "_ckp.h5"
checkPointSaver = ModelCheckpoint(ckp_model_name,
verbose=1,
save_best_only=False,
save_weights_only=False,
period=10)
tb = TensorBoard(log_dir=config['train']['tensorboard_log_dir'],
histogram_freq=0,
batch_size=batchSize,
write_graph=True,
write_grads=False,
write_images=False,
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None)
#create the classification model
# make the feature extractor layers
if depth == 1:
input_size = (height, width, 1)
input_image = Input(shape=input_size)
else:
input_size = (height, width, 3)
input_image = Input(shape=input_size)
feature_extractor = import_feature_extractor(config['model']['backend'],
input_size)
train_generator = BatchGenerator(imagesTrainPaths,
generator_config,
norm=feature_extractor.normalize,
jitter=True)
val_generator = BatchGenerator(imagesValPaths,
generator_config,
norm=feature_extractor.normalize,
jitter=False)
features = feature_extractor.extract(input_image)
# make the model head
output = Conv2D(classes, (1, 1), padding="same")(features)
output = BatchNormalization()(output)
output = LeakyReLU(alpha=0.1)(output)
output = GlobalAveragePooling2D()(output)
output = Activation("sigmoid")(output) if classes == 1 else Activation(
"softmax")(output)
if config['train']['pretrained_weights'] != "":
model = load_model(config['model']['pretrained_weights'])
else:
model = Model(input_image, output)
opt = Adam()
model.compile(loss="binary_crossentropy"
if classes == 1 else "categorical_crossentropy",
optimizer=opt,
metrics=["accuracy"])
model.summary()
model.fit_generator(train_generator,
steps_per_epoch=len(imagesTrainPaths) // batchSize,
epochs=epochs,
validation_data=val_generator,
validation_steps=len(imagesValPaths) // batchSize,
callbacks=[checkPointSaverBest, checkPointSaver, tb],
workers=12,
max_queue_size=40)