变分自编码器和条件变分自编码器python实现源码+说明.zip

import tensorflow as tf from tensorflow.keras.layers import Input, Dense, Layer, Embedding from tensorflow.keras.models import Model from tensorflow.keras.losses import binary_crossentropy from tensorflow.keras.optimizers import Adam from tensorflow.keras.utils import plot_model import numpy as np import os param_model_path = "models/cvae/" param_result_path = "result/cvae/" if not os.path.exists(param_model_path): os.makedirs(param_model_path) if not os.path.exists(param_result_path): os.makedirs(param_result_path) param_img_width, param_img_heigth = 28, 28 param_intermediate_size = 512 param_latent_size = 20 param_batch_size = 512 param_epochs = 100 class Sampling(Layer): def __init__(self, **kwargs): super(Sampling, self).__init__(**kwargs) def call(self, inputs, **kwargs): mu, logvar = inputs epsilon = tf.random.normal(shape=tf.shape(mu)) return mu + tf.exp(0.5 * logvar) * epsilon class VAELOSS(Layer): def __init__(self, **kwargs): super(VAELOSS, self).__init__(**kwargs) def call(self, inputs, **kwargs): output, realimg, mu, logvar = inputs bceloss = binary_crossentropy(realimg, output) klloss = -0.5 * (logvar + 1 - mu ** 2 - tf.exp(logvar)) klloss = tf.reduce_sum(klloss, axis=-1) loss = tf.reduce_mean(param_img_heigth * param_img_width * bceloss + klloss) self.add_loss(loss) return output class CustomEmbedding(Layer): def __init__(self, outputsize, **kwargs): super(CustomEmbedding, self).__init__(**kwargs) self.outputsize = outputsize def build(self, input_shape): self.embedding = Embedding(10, self.outputsize) super(CustomEmbedding, self).build(input_shape) def call(self, inputs, **kwargs): x = self.embedding(inputs) return tf.squeeze(x, axis=1) class USER: def build_model(self, summary=False, plot=False): realimg = Input(shape=[param_img_heigth * param_img_width, ], name="realimg") label = Input(shape=[1, ], name="label", dtype=tf.int32) label_embed = CustomEmbedding(param_img_heigth * param_img_width, name="labelembeddingen")(label) x = Dense(param_intermediate_size, activation="relu", name="interdense")(realimg + label_embed) mu = Dense(param_latent_size, name="mu")(x) logvar = Dense(param_latent_size, name="logvar")(x) z = Sampling(name="sampling")(inputs=(mu, logvar)) encoder = Model(inputs=[realimg, label], outputs=[mu, logvar, z], name="encoder") latentimg = Input(shape=[param_latent_size, ], name="latentimg") label_embed = CustomEmbedding(param_latent_size, name="labelembeddingde")(label) x = Dense(param_intermediate_size, activation="relu", name="invinterdense")(latentimg + label_embed) outputimg = Dense(param_img_width * param_img_heigth, activation="sigmoid", name="invimg")(x) decoder = Model(inputs=[latentimg, label], outputs=outputimg, name="decoder") output1 = decoder((encoder([realimg, label])[2], label)) output = VAELOSS(name="vaeloss")(inputs=(output1, realimg, mu, logvar)) vae = Model(inputs=[realimg, label], outputs=output, name="vae") if summary: encoder.summary() decoder.summary() vae.summary() if plot: plot_model(encoder, param_model_path + "encoder.png", show_shapes=True) plot_model(decoder, param_model_path + "decoder.png", show_shapes=True) plot_model(vae, param_model_path + "vae.png", show_shapes=True) return encoder, decoder, vae def train(self): (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data() x_train = np.reshape(x_train, [-1, param_img_width * param_img_heigth]) x_test = np.reshape(x_test, [-1, param_img_width * param_img_heigth]) x_train = x_train.astype('float32') / 255 x_test = x_test.astype('float32') / 255 encoder, decoder, vae = self.build_model(summary=True, plot=True) vae.compile(optimizer=Adam(0.01)) vae.fit([x_train, y_train], epochs=param_epochs, batch_size=param_batch_size, validation_data=[x_test, y_test]) encoder.save_weights(param_model_path + "encoder.h5") decoder.save_weights(param_model_path + "decoder.h5") vae.save_weights(param_model_path + "vae.h5") def predict(self): _, _, vae = self.build_model() vae.load_weights(param_model_path + "vae.h5") (_, _), (x_test, y_test) = tf.keras.datasets.mnist.load_data() x_test = np.reshape(x_test[:11], [-1, param_img_width * param_img_heigth]) x_test = x_test.astype('float32') / 255 y_test = y_test[:11] y_predict = vae.predict([x_test, y_test]) realimages = np.reshape(x_test * 255, [-1, param_img_width, param_img_heigth, 1]) fakeimages = np.reshape(y_predict * 255, [-1, param_img_width, param_img_heigth, 1]) for i in range(11): realimg = tf.image.encode_jpeg(realimages[i], quality=100) fakeimg = tf.image.encode_jpeg(fakeimages[i], quality=100) with tf.io.gfile.GFile(param_result_path + 'realimg' + str(i) + '_' + str(y_test[i]) + '.jpg', 'wb') as file: file.write(realimg.numpy()) with tf.io.gfile.GFile(param_result_path + 'fakeimg' + str(i) + '_' + str(y_test[i]) + '.jpg', 'wb') as file: file.write(fakeimg.numpy()) def test(self): _, decoder, _ = self.build_model() decoder.load_weights(param_model_path + "decoder.h5") latentimgs = tf.random.normal([11, param_latent_size]) labels = tf.random.uniform([11], 0, 10, dtype=tf.int32) fakeimgs = decoder.predict([latentimgs, labels]) fakeimgs = np.reshape(fakeimgs * 255, [-1, param_img_width, param_img_heigth, 1]) for i in range(11): fakeimg = tf.image.encode_jpeg(fakeimgs[i], quality=100) with tf.io.gfile.GFile(param_result_path + 'testfakeimg' + str(i) + '_' + str(labels[i].numpy()) + '.jpg', 'wb') as file: file.write(fakeimg.numpy()) if __name__ == "__main__": user = USER() user.train() user.predict() user.test()