def model():
inputLayer = Input(shape=(256, 256, 3))
DSRCNN = UpSampling2D(size=2)(inputLayer)
inputImage = DSRCNN
DSRCNN = Conv2D(9, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(16, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(32, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(64, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(64, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(64, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(32, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(16, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(9, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("relu")(DSRCNN)
DSRCNN = Conv2D(3, (3, 3),
strides=(1, 1),
padding="same",
init='glorot_uniform')(DSRCNN)
DSRCNN = Activation("linear")(DSRCNN)
residualImage = DSRCNN
outputImage = add([inputImage, residualImage])
DSRCNN = Model(inputLayer, outputImage)
print DSRCNN.summary()
return DSRCNN
def residual_TCN_LSTM(n_nodes,
conv_len,
n_classes,
n_feat,
max_len,
loss='categorical_crossentropy',
online=False,
optimizer="rmsprop",
depth=3,
return_param_str=False):
n_layers = len(n_nodes)
inputs = Input(shape=(max_len, n_feat))
model = inputs
prev = Conv1D(n_nodes[0], conv_len, padding='same')(model)
# encoder
for i in range(n_layers):
for j in range(depth):
# convolution over the temporal dimension
current = encoder_identify_block(prev, n_nodes[i], conv_len)
# residual connection within residual block
if j != 0:
model = add([prev, current])
model = Activation('relu')(model)
prev = current
if i < (n_layers - 1):
model = MaxPooling1D(2)(model)
# decoder
# for i in range(n_layers):
#
# for j in range(depth):
#
# current = decoder_identify_block(model, n_nodes[-i - 1])
# model = add([prev, current])
# prev = current
# model = UpSampling1D(2)(model)
# Output FC layer
model = TimeDistributed(Dense(n_classes, activation="softmax"))(model)
model = Model(inputs=inputs, outputs=model)
model.compile(loss=loss,
optimizer=optimizer,
sample_weight_mode="temporal",
metrics=['accuracy'])
model.summary()
if return_param_str:
param_str = "ED-TCN_C{}_L{}".format(conv_len, n_layers)
if online:
param_str += "_online"
return model, param_str
else:
return model
def model_train(img_size,
batch_size,
epochs,
optimizer,
learning_rate,
train_list,
validation_list,
style=2):
print('Style {}.'.format(style))
if style == 1:
input_img = Input(shape=img_size)
#model = Sequential()
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal',
input_shape=img_size)(input_img)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(1, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
res_img = model
output_img = merge.Add()([res_img, input_img])
model = Model(input_img, output_img)
#model.load_weights('vdsr_model_edges.h5')
adam = Adam(lr=0.000005)
#sgd = SGD(lr=1e-3, momentum=0.9, decay=1e-4, nesterov=False)
sgd = SGD(lr=0.01, momentum=0.9, decay=0.001, nesterov=False)
#model.compile(sgd, loss='mse', metrics=[PSNR, "accuracy"])
model.compile(adam,
loss='mse',
metrics=[ssim, ssim_metric, PSNR, "accuracy"])
model.summary()
else:
input_img = Input(shape=img_size)
model = Conv2D(64, (3, 3),
padding='valid',
kernel_initializer='he_normal')(input_img)
model_0 = Activation('relu')(model)
total_conv = 22 # should be even number
total_conv -= 2 # subtract first and last
residual_block_num = 5 # should be even number
for _ in range(residual_block_num): # residual block
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model_0)
model = Activation('relu')(model)
for _ in range(int(total_conv / residual_block_num) - 1):
model = Conv2D(64, (3, 3),
padding='same',
kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model_0 = add([model, model_0])
model = Conv2D(1, (3, 3),
padding='valid',
kernel_initializer='he_normal')(model)
res_img = model
input_img1 = crop(1, 2, -2)(input_img)
input_img1 = crop(2, 2, -2)(input_img1)
print(input_img.shape)
print(input_img1.shape)
output_img = merge.Add()([res_img, input_img1])
# output_img = res_img
model = Model(input_img, output_img)
# model.load_weights('./vdsr_model_edges.h5')
# adam = Adam(lr=learning_rate)
adam = Adadelta()
# sgd = SGD(lr=1e-7, momentum=0.9, decay=1e-2, nesterov=False)
sgd = SGD(lr=learning_rate,
momentum=0.9,
decay=1e-4,
nesterov=False,
clipnorm=1)
if optimizer == 0:
model.compile(adam, loss='mse', metrics=[ssim, ssim_metric, PSNR])
else:
model.compile(sgd, loss='mse', metrics=[ssim, ssim_metric, PSNR])
model.summary()
mycallback = MyCallback(model)
timestamp = time.strftime("%m%d-%H%M", time.localtime(time.time()))
csv_logger = callbacks.CSVLogger(
'data/callbacks/training_{}.log'.format(timestamp))
filepath = "./checkpoints/weights-improvement-{epoch:03d}-{PSNR:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor=PSNR, verbose=1, mode='max')
callbacks_list = [mycallback, checkpoint, csv_logger]
# print('Loading training data.')
# x = load_images(DATA_PATH)
# print('Loading data label.')
# y = load_images(LABEL_PATH)
# print('Loading validation data.')
# val = load_images(VAL_PATH)
# print('Loading validation label.')
# val_label = load_images(VAL_LABEL_PATH)
# print(x.shape)
# print(y.shape)
# print(val.shape)
# print(val_label.shape)
with open('./model/vdsr_architecture.json', 'w') as f:
f.write(model.to_json())
# datagen = ImageDataGenerator(rotation_range=45,
# zoom_range=0.15,
# horizontal_flip=True,
# vertical_flip=True)
# history = model.fit_generator(datagen.flow(x, y, batch_size=batch_size),
# steps_per_epoch=len(x) // batch_size,
# validation_data=(val, val_label),
# validation_steps=len(val) // batch_size,
# epochs=epochs,
# callbacks=callbacks_list,
# verbose=1,
# shuffle=True,
# workers=256,
# use_multiprocessing=True)
history = model.fit_generator(
image_gen(train_list, batch_size=batch_size),
steps_per_epoch=384400 * (len(train_list)) // batch_size,
# steps_per_epoch=4612800//batch_size,
validation_data=image_gen(validation_list, batch_size=batch_size),
validation_steps=384400 * (len(validation_list)) // batch_size,
epochs=epochs,
workers=1024,
callbacks=callbacks_list,
verbose=1)
print("Done training!!!")
print("Saving the final model ...")
model.save('vdsr_model.h5') # creates a HDF5 file
del model # deletes the existing model
# plt.plot(history.history['accuracy'])
# plt.plot(history.history['val_accuracy'])
# plt.title('Model accuracy')
# plt.ylabel('Accuracy')
# plt.xlabel('Epoch')
# plt.legend(['Train', 'validation'], loc='upper left')
# # plt.show()
# plt.savefig('accuracy.png')
# Plot training & validation loss values
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('Model loss')
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.legend(['Train', 'validation'], loc='upper left')
# plt.show()
plt.savefig('loss.png')
plt.plot(history.history['PSNR'])
plt.plot(history.history['val_PSNR'])
plt.title('Model PSNR')
plt.ylabel('PSNR')
plt.xlabel('Epoch')
plt.legend(['Train', 'validation'], loc='upper left')
# plt.show()
plt.savefig('PSNR.png')
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(1, (3, 3), padding='same', kernel_initializer='he_normal')(model)
res_img = model
output_img = add([res_img, input_img])
model = Model(input_img, output_img)
# model.load_weights('./checkpoints/weights-improvement-20-26.93.hdf5')
adam = Adam(lr=0.001, decay=1e-5, clipvalue=0.1, epsilon=1e-8)
sgd = SGD(lr=1e-2, momentum=0.9, decay=1e-4, nesterov=False)
model.compile(adam, loss='mse', metrics=[PSNR, "accuracy"])
model.summary()
# 每个epoch保存一次模型
filepath = "./checkpoints2/vdsr-{epoch:02d}-{PSNR:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor=PSNR, verbose=1, mode='max')
callbacks_list = [checkpoint]
# 记录每次训练的loss，PSNR值
train_log = CSVLogger(filename="train.log")
model.fit(x=train_input, y=train_label, batch_size=BATCH_SIZE, epochs=EPOCHS, callbacks=callbacks_list, shuffle=True)
print("Done training!!!")
def ED_TCN(n_nodes,
conv_len,
n_classes,
n_feat,
max_len,
loss='categorical_crossentropy',
online=False,
optimizer="rmsprop",
activation='norm_relu',
attention=True,
return_param_str=False):
n_layers = len(n_nodes)
# inputs = Input(shape=(max_len, n_feat))
inputs = Input(shape=(None, n_feat))
# attention layer, apply weightings to input
if attention:
model = attention_block(inputs, 100)
else:
model = inputs
# ---- Encoder ----
for i in range(n_layers):
# Pad beginning of sequence to prevent usage of future data
if online: model = ZeroPadding1D((conv_len // 2, 0))(model)
# convolution over the temporal dimension
model = Conv1D(n_nodes[i], conv_len, padding='same')(model)
if online: model = Cropping1D((0, conv_len // 2))(model)
model = SpatialDropout1D(0.3)(model)
if activation == 'norm_relu':
model = Activation('relu')(model)
model = Lambda(channel_normalization,
name="encoder_norm_{}".format(i))(model)
elif activation == 'wavenet':
model = WaveNet_activation(model)
else:
model = Activation(activation)(model)
# hidden features layer when in the last interation
model = MaxPooling1D(2)(model)
# ---- Decoder ----
for i in range(n_layers):
model = UpSampling1D(2)(model)
if online: model = ZeroPadding1D((conv_len // 2, 0))(model)
model = Conv1D(n_nodes[-i - 1], conv_len, padding='same')(model)
if online: model = Cropping1D((0, conv_len // 2))(model)
model = SpatialDropout1D(0.3)(model)
if activation == 'norm_relu':
model = Activation('relu')(model)
model = Lambda(channel_normalization,
name="decoder_norm_{}".format(i))(model)
elif activation == 'wavenet':
model = WaveNet_activation(model)
else:
model = Activation(activation)(model)
# Output FC layer
model = TimeDistributed(Dense(n_classes, activation="softmax"))(model)
model = Model(inputs=inputs, outputs=model)
model.compile(loss=loss,
optimizer=optimizer,
sample_weight_mode="temporal",
metrics=['accuracy'])
model.summary()
if return_param_str:
param_str = "ED-TCN_C{}_L{}".format(conv_len, n_layers)
if online:
param_str += "_online"
return model, param_str
else:
return model
def model_train(img_size, batch_size, epochs, optimizer, learning_rate, train_list, validation_list, style=2):
print('Style {}.'.format(style))
if style == 1:
input_img = Input(shape=img_size)
#model = Sequential()
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal', input_shape=img_size)(input_img)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(1, (3, 3), padding='same', kernel_initializer='he_normal')(model)
res_img = model
output_img = merge.Add()([res_img, input_img])
model = Model(input_img, output_img)
#model.load_weights('vdsr_model_edges.h5')
adam = Adam(lr=0.000005)
sgd = SGD(lr=0.01, momentum=0.9, decay=0.001, nesterov=False)
model.compile(adam, loss='mse', metrics=[ssim, ssim_metric, PSNR, "accuracy"])
model.summary()
else:
input_img = Input(shape=img_size)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal', use_bias=False)(input_img)
model = BatchNormalization()(model)
model_0 = Activation('relu')(model)
total_conv = 22 # should be even number
total_conv -= 2 # subtract first and last
residual_block_num = 5 # should be even number
for _ in range(residual_block_num): # residual block
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal', use_bias=False)(model_0)
model = BatchNormalization()(model)
model = Activation('relu')(model)
print(_)
for _ in range(int(total_conv/residual_block_num)-1):
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal', use_bias=False)(model)
model = BatchNormalization()(model)
model = Activation('relu')(model)
model_0 = add([model, model_0])
print(_)
model = Conv2DTranspose(64, (5, 5), padding='valid', kernel_initializer='he_normal', use_bias=False)(model)
model = BatchNormalization()(model)
model = LeakyReLU()(model)
model = Conv2D(1, (5, 5), padding='valid', kernel_initializer='he_normal')(model)
res_img = model
#input_img1 = crop(1,22,-22)(input_img)
#input_img1 = crop(2,22,-22)(input_img1)
print(input_img.shape)
output_img = merge.Add()([res_img, input_img])
# output_img = res_img
model = Model(input_img, output_img)
# model.load_weights('./vdsr_model_edges.h5')
# adam = Adam(lr=learning_rate)
adam = Adadelta()
sgd = SGD(lr=learning_rate, momentum=0.9, decay=1e-4, nesterov=False, clipnorm=1)
if optimizer == 0:
model.compile(adam, loss='mse', metrics=[ssim, ssim_metric, PSNR])
else:
model.compile(sgd, loss='mse', metrics=[ssim, ssim_metric, PSNR])
model.summary()
mycallback = MyCallback(model)
timestamp = time.strftime("%m%d-%H%M", time.localtime(time.time()))
csv_logger = callbacks.CSVLogger('data/callbacks/deconv/training_{}.log'.format(timestamp))
filepath="./checkpoints/deconv/weights-improvement-{epoch:03d}-{PSNR:.2f}-{ssim:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor=PSNR, verbose=1, mode='max')
callbacks_list = [mycallback, checkpoint, csv_logger]
with open('./model/deconv/vdsr_architecture.json', 'w') as f:
f.write(model.to_json())
history = model.fit_generator(image_gen(train_list, batch_size=batch_size),
steps_per_epoch=(409600//8)*len(train_list) // batch_size,
validation_data=image_gen(validation_list,batch_size=batch_size),
validation_steps=(409600//8)*len(validation_list) // batch_size,
epochs=epochs,
workers=1024,
callbacks=callbacks_list,
verbose=1)
print("Done training!!!")
print("Saving the final model ...")
model.save('vdsr_model.h5') # creates a HDF5 file
del model # deletes the existing model
# Plot training & validation loss values
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('Model loss')
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.legend(['Train', 'validation'], loc='upper left')
# plt.show()
plt.savefig('loss.png')
plt.plot(history.history['PSNR'])
plt.plot(history.history['val_PSNR'])
plt.title('Model PSNR')
plt.ylabel('PSNR')
plt.xlabel('Epoch')
plt.legend(['Train', 'validation'], loc='upper left')
# plt.show()
plt.savefig('PSNR.png')
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(1, (3, 3), padding='same', kernel_initializer='he_normal')(model)
res_img = model
output_img = add([res_img, input_img])
model = Model(input_img, output_img)
# model.load_weights('./checkpoints/weights-improvement-20-26.93.hdf5')
adam = Adam(lr=0.00001)
sgd = SGD(lr=1e-5, momentum=0.9, decay=1e-4, nesterov=False)
model.compile(adam, loss='mse', metrics=[PSNR, "accuracy"])
model.summary()
filepath="./checkpoints/weights-improvement-{epoch:02d}-{PSNR:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor=PSNR, verbose=1, mode='max')
callbacks_list = [checkpoint]
model.fit_generator(image_gen(train_list), steps_per_epoch=len(train_list) // BATCH_SIZE, \
validation_data=image_gen(test_list), validation_steps=len(train_list) // BATCH_SIZE, \
epochs=EPOCHS, workers=8, callbacks=callbacks_list)
print("Done training!!!")
print("Saving the final model ...")
model.save('vdsr_model.h5') # creates a HDF5 file
del model # deletes the existing model
def ResCat_aux(n_classes, input_shape, dropout=0.5, aux_ind=0):
filter_numbers = [32, 64, 128]
init_filters = 32
merge_layers = []
outputs = []
input = Input(shape=input_shape)
conv1 = Conv2D(filters=init_filters,
kernel_size=(7, 7),
strides=(2, 2),
padding='valid',
kernel_initializer='he_normal',
kernel_regularizer=l2(1.e-4))(input)
norm = BatchNormalization(axis=CHANNEL_AXIS)(conv1)
relu = Activation("relu")(norm)
dr = Dropout(dropout)(relu)
model = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), padding="same")(dr)
for fn in filter_numbers:
res1 = bn_relu_conv(model, fn, (3, 3), dropout=dropout)
res2 = bn_relu_conv(res1, fn, (3, 3), strides=2, dropout=dropout)
input_shape = K.int_shape(model)
residual_shape = K.int_shape(res2)
stride_width = int(
round(input_shape[ROW_AXIS] / residual_shape[ROW_AXIS]))
stride_height = int(
round(input_shape[COL_AXIS] / residual_shape[COL_AXIS]))
equal_channels = input_shape[CHANNEL_AXIS] == residual_shape[
CHANNEL_AXIS]
# 1 X 1 conv if shape is different. Else identity.
if stride_width > 1 or stride_height > 1 or not equal_channels:
model = Conv2D(filters=residual_shape[CHANNEL_AXIS],
kernel_size=(1, 1),
strides=(stride_width, stride_height),
padding="valid",
kernel_initializer="he_normal",
kernel_regularizer=l2(0.0001))(model)
model = add([res2, model])
merge_layers.append(model)
# Last activation
norm = BatchNormalization(axis=CHANNEL_AXIS)(model)
model = Activation("relu")(norm)
# Classifier block
block_shape = K.int_shape(model)
model = AveragePooling2D(pool_size=(block_shape[ROW_AXIS],
block_shape[COL_AXIS]),
strides=(1, 1))(model)
flatten = Flatten()(model)
dense = Dense(units=n_classes,
kernel_initializer="he_normal",
activation="sigmoid",
name='main')(flatten)
outputs.append(dense)
## AUX, does obj exists?
aux = merge_layers[aux_ind]
pool = AveragePooling2D(pool_size=(3, 3), strides=(2, 2))(aux)
flatten = Flatten()(pool)
fc = Dense(units=filter_numbers[-1],
kernel_initializer="he_normal",
activation="relu")(flatten)
dense = Dense(units=2,
kernel_initializer="he_normal",
activation="softmax",
name='aux')(fc)
outputs.append(dense)
model = Model(inputs=input, outputs=outputs)
model.summary()
return model
