BiGAN的复现
Posted nanhaijindiao
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了BiGAN的复现相关的知识,希望对你有一定的参考价值。
数据集是10000个样本,前8000个训练集,后面的用来测试。每个样本是1*144(重构成12*12的矩阵),将原始BiGAN有编码器、判别器和生成器,将里面的全连接层全部替换成了卷积。
from __future__ import print_function, division from keras.datasets import mnist from keras.layers import Input, Dense, Reshape, Flatten, Dropout, multiply, GaussianNoise from keras.layers import BatchNormalization, Activation, Embedding, ZeroPadding2D from keras.layers import MaxPooling2D, concatenate from keras.layers.advanced_activations import LeakyReLU from keras.layers.convolutional import UpSampling2D, Conv2D from keras.models import Sequential, Model from keras.optimizers import Adam from keras import losses from keras.utils import to_categorical from pandas import read_csv import keras.backend as K import pandas as pd import matplotlib.pyplot as plt import numpy as np class BIGAN(): def __init__(self): self.img_rows = 12 self.img_cols = 12 self.channels = 1 self.img_shape = (self.img_rows, self.img_cols, self.channels) self.latent_dim = 100 optimizer = Adam(0.0002, 0.5) # Build and compile the discriminator self.discriminator = self.build_discriminator() self.discriminator.compile(loss=[‘binary_crossentropy‘], optimizer=optimizer, metrics=[‘accuracy‘]) # Build the generator self.generator = self.build_generator() # Build the encoder self.encoder = self.build_encoder() # The part of the bigan that trains the discriminator and encoder self.discriminator.trainable = False # Generate image from sampled noise z = Input(shape=(self.latent_dim, )) img_ = self.generator(z) # Encode image img = Input(shape=self.img_shape) z_ = self.encoder(img) # Latent -> img is fake, and img -> latent is valid fake = self.discriminator([z, img_]) valid = self.discriminator([z_, img]) # Set up and compile the combined model # Trains generator to fool the discriminator self.bigan_generator = Model([z, img], [fake, valid]) self.bigan_generator.compile(loss=[‘binary_crossentropy‘, ‘binary_crossentropy‘], optimizer=optimizer) def build_encoder(self): model = Sequential() model.add(Conv2D(16, kernel_size=3, strides=2, input_shape=self.img_shape, padding="same")) model.add(LeakyReLU(alpha=0.2)) model.add(Dropout(0.25)) model.add(Conv2D(32, kernel_size=3, strides=2, padding="same")) model.add(ZeroPadding2D(padding=((0,1),(0,1)))) model.add(BatchNormalization(momentum=0.8)) model.add(LeakyReLU(alpha=0.2)) model.add(Dropout(0.25)) model.add(Conv2D(64, kernel_size=3, strides=2, padding="same")) model.add(BatchNormalization(momentum=0.8)) model.add(LeakyReLU(alpha=0.2)) model.add(Dropout(0.25)) model.add(Conv2D(128, kernel_size=3, strides=1, padding="same")) model.add(BatchNormalization(momentum=0.8)) model.add(LeakyReLU(alpha=0.2)) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(self.latent_dim)) model.summary() img = Input(shape=self.img_shape) z = model(img) return Model(img, z) def build_generator(self): model = Sequential() model.add(Dense(64 * 3 * 3, activation="relu", input_dim=self.latent_dim)) model.add(Reshape((3, 3, 64))) model.add(UpSampling2D()) model.add(Conv2D(64, kernel_size=3, padding="same")) model.add(BatchNormalization(momentum=0.8)) model.add(Activation("relu")) model.add(UpSampling2D()) model.add(Conv2D(32, kernel_size=3, padding="same")) model.add(BatchNormalization(momentum=0.8)) model.add(Activation("relu")) model.add(Conv2D(self.channels, kernel_size=3, padding="same")) model.add(Activation("tanh")) model.summary() z = Input(shape=(self.latent_dim,)) gen_img = model(z) return Model(z, gen_img) def build_discriminator(self): z = Input(shape=(self.latent_dim, )) img = Input(shape=self.img_shape) d_in = concatenate([z, Flatten()(img)]) model = Dense(14*14, activation="relu")(d_in) model = Reshape((14, 14, 1))(model) model = Conv2D(16, kernel_size=3, strides=2,padding="same")(model) model = LeakyReLU(alpha=0.2)(model) model = Dropout(0.25)(model) model = Conv2D(32, kernel_size=3, strides=2, padding="same")(model) model = ZeroPadding2D(padding=((0,1),(0,1)))(model) model = BatchNormalization(momentum=0.8)(model) model = LeakyReLU(alpha=0.2)(model) model = Dropout(0.25)(model) model = Conv2D(64, kernel_size=3, strides=2, padding="same")(model) model = BatchNormalization(momentum=0.8)(model) model = LeakyReLU(alpha=0.2)(model) model = Dropout(0.25)(model) model = Conv2D(128, kernel_size=3, strides=1, padding="same")(model) model = BatchNormalization(momentum=0.8)(model) model = LeakyReLU(alpha=0.2)(model) model = Dropout(0.25)(model) model = Flatten()(model) validity = Dense(1, activation="sigmoid")(model) return Model([z, img], validity) def train(self, epochs, batch_size=128, sample_interval=50): # Load the dataset dataset = read_csv(‘GANData.csv‘) values = dataset.values XY= values n_train_hours1 =8000 n_train_hours2 = n_train_hours1+1 x_train=XY[:n_train_hours1,:] x_test =XY[n_train_hours2:, :] X_train = x_train.reshape(-1,12,12,1) X_test = x_test.reshape(-1,12,12,1) # Adversarial ground truths valid = np.ones((batch_size, 1)) fake = np.zeros((batch_size, 1)) FHZ=np.zeros((epochs, 3)) for epoch in range(epochs): # --------------------- # Train Discriminator # --------------------- # Sample noise and generate img z = np.random.normal(size=(batch_size, self.latent_dim)) imgs_ = self.generator.predict(z) # Select a random batch of images and encode idx = np.random.randint(0, X_train.shape[0], batch_size) imgs = X_train[idx] z_ = self.encoder.predict(imgs) # Train the discriminator (img -> z is valid, z -> img is fake) d_loss_real = self.discriminator.train_on_batch([z_, imgs], valid) d_loss_fake = self.discriminator.train_on_batch([z, imgs_], fake) d_loss = 0.5 * np.add(d_loss_real, d_loss_fake) # --------------------- # Train Generator # --------------------- # Train the generator (z -> img is valid and img -> z is is invalid) g_loss = self.bigan_generator.train_on_batch([z, imgs], [valid, fake]) # Plot the progress print ("%d [D loss: %f, acc: %.2f%%] [G loss: %f]" % (epoch, d_loss[0], 100*d_loss[1], g_loss[0])) FHZ[epoch,0]=d_loss[0] FHZ[epoch,1]=d_loss[1] FHZ[epoch,2]=g_loss[0] # If at save interval => save generated image samples if epoch % sample_interval == 0: self.sample_interval(epoch) return FHZ def sample_interval(self, epoch): r, c = 5, 5 z = np.random.normal(size=(25, self.latent_dim)) gen_imgs = self.generator.predict(z) gen_imgs = 0.5 * gen_imgs + 0.5 decoded_imgs = gen_imgs.reshape((gen_imgs.shape[0], -1)) print(‘decoded_imgs.shape:‘,decoded_imgs.shape) data=decoded_imgs data_df = pd.DataFrame(data) # create and writer pd.DataFrame to excel writer = pd.ExcelWriter(‘Result.xlsx‘) data_df.to_excel(writer,‘page_1‘,float_format=‘%.5f‘) # float_format 控制精度 writer.save() # Rescale images 0 - 1 #fig.savefig("images/mnist_%d.png" % epoch) if __name__ == ‘__main__‘: bigan = BIGAN() d_loss=bigan.train(epochs=10, batch_size=32, sample_interval=9) import numpy numpy.savetxt("d_lossnum.csv", d_loss, delimiter=‘,‘)
以上是关于BiGAN的复现的主要内容,如果未能解决你的问题,请参考以下文章
论文复现使用PaddleDetection复现OrientedRepPoints的复现笔记