I'll cut to the chase, so let's get right to the code!
from import Sequential, Model from import Dense, Flatten, Conv2D, MaxPool2D, Dropout, Conv2DTranspose, Lambda, Input, Reshape, Add, Multiply from import Adam def deconv(x): height = x.get_shape()[1].value width = x.get_shape()[2].value new_height = height*2 new_width = width*2 x_resized = .resize_images(x, [new_height, new_width], .NEAREST_NEIGHBOR) return x_resized def Generator(scope='generator'): imgs_noise = Input(shape=inputs_shape) x = Conv2D(filters=32, kernel_size=(9,9), strides=(1,1), padding='same', activation='relu')(imgs_noise) x = Conv2D(filters=64, kernel_size=(3,3), strides=(2,2), padding='same', activation='relu')(x) x = Conv2D(filters=128, kernel_size=(3,3), strides=(2,2), padding='same', activation='relu')(x) x1 = Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu')(x) x1 = Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu')(x1) x2 = Add()([x1, x]) x3 = Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu')(x2) x3 = Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu')(x3) x4 = Add()([x3, x2]) x5 = Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu')(x4) x5 = Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu')(x5) x6 = Add()([x5, x4]) x = MaxPool2D(pool_size=(2,2))(x6) x = Lambda(deconv)(x) x = Conv2D(filters=64, kernel_size=(3, 3), strides=(1,1), padding='same',activation='relu')(x) x = Lambda(deconv)(x) x = Conv2D(filters=32, kernel_size=(3, 3), strides=(1,1), padding='same',activation='relu')(x) x = Lambda(deconv)(x) x = Conv2D(filters=3, kernel_size=(3, 3), strides=(1, 1), padding='same',activation='tanh')(x) x = Lambda(lambda x: x+1)(x) y = Lambda(lambda x: x*127.5)(x) model = Model(inputs=imgs_noise, outputs=y) () return model my_generator = Generator() my_generator.compile(loss='binary_crossentropy', optimizer=Adam(0.7, decay=1e-3), metrics=['accuracy'])
Additional knowledge:Containing Lambda custom layer keras model, save the problems encountered and solutions
First, many applications, keras contains layers can not meet the requirements, you need to customize the layer through Lambda to achieve some layer, in this case, can only save the model weights, can not be used to save the model. Save will report
TypeError: can't pickle _thread.RLock objects
Second, the solution, in order to facilitate the subsequent deployment, can be converted to tensorflow PB for deployment.
from import load_model
import tensorflow as tf
import os, sys
from keras import backend as K
from import graph_util, graph_io
def h5_to_pb(h5_weight_path, output_dir, out_prefix="output_", log_tensorboard=True):
if not (output_dir):
(output_dir)
h5_model = build_model()
h5_model.load_weights(h5_weight_path)
out_nodes = []
for i in range(len(h5_model.outputs)):
out_nodes.append(out_prefix + str(i + 1))
(h5_model.output[i], out_prefix + str(i + 1))
model_name = ((h5_weight_path)[-1])[0] + '.pb'
sess = K.get_session()
init_graph = .as_graph_def()
main_graph = graph_util.convert_variables_to_constants(sess, init_graph, out_nodes)
graph_io.write_graph(main_graph, output_dir, name=model_name, as_text=False)
if log_tensorboard:
from import import_pb_to_tensorboard
import_pb_to_tensorboard.import_to_tensorboard((output_dir, model_name), output_dir)
def build_model():
inputs = Input(shape=(784,), name='input_img')
x = Dense(64, activation='relu')(inputs)
x = Dense(64, activation='relu')(x)
y = Dense(10, activation='softmax')(x)
h5_model = Model(inputs=inputs, outputs=y)
return h5_model
if __name__ == '__main__':
if len() == 3:
# usage: python3 h5_to_pb.py h5_weight_path output_dir
h5_to_pb(h5_weight_path=[1], output_dir=[2])
Above this Keras using Lambda layer details is all that I have shared with you, I hope to give you a reference, and I hope you support me more.