WGAN-GP與WGAN及GAN的比較
阿新 • • 發佈：2019-01-08
以下為MNIST測試，主要為了構建模型，只跑了，少量epoch,效果如下:
WGAN 2個epoch
wgan-gp 6個epoch
gan 10個epoch
有時間可以多跑幾輪，這裡就不展示了。
程式碼如下
以下測試自己的資料集from datetime import datetime import os import matplotlib.pyplot as plt import numpy as np import tensorflow as tf from six.moves import xrange from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets("MNIST_data",one_hot=True) data= mnist.train.images#(55000,784) def Save_genImages(gen, epoch): r,c = 10,10 fig,axs = plt.subplots(r,c) cnt = 0 print(gen.shape) for i in range(r): for j in range(c): axs[i,j].imshow(gen[cnt][:,:],cmap='Greys_r') axs[i,j].axis('off') cnt += 1 if not os.path.exists('gen_mnist'): os.makedirs('gen_mnist') fig.savefig('gen_mnist/%d.jpg' % epoch) plt.close() def Save_lossValue(epoch,iters,d_loss,g_loss): with open('loss2.txt','a') as f: f.write("第%d個epoch,第%d個batch , d_loss: %.8f, g_loss: %.8f"%(epoch, iters, d_loss, g_loss)+'\n') def plot_loss(loss): fig,ax = plt.subplots(figsize=(20,7)) losses = np.array(loss) plt.plot(losses.T[0], label="Discriminator Loss") plt.plot(losses.T[1], label="Generator Loss") plt.title("Training Losses") plt.legend() plt.savefig('loss2.jpg') plt.show() #定義Relu啟用函式 def Relu(name, tensor): return tf.nn.relu(tensor,name) #定義LeakyRelu啟用函式 def LeakyRelu(x, alpha=0.2): return tf.maximum(x, alpha * x) #定義全連線層 def Fully_connected(name, value, output_shape): with tf.variable_scope(name, reuse=None) as scope: shape = value.get_shape().as_list() w = tf.get_variable('w', [shape[1], output_shape], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=0.01)) b = tf.get_variable('b', [output_shape], dtype=tf.float32, initializer=tf.constant_initializer(0.0)) return tf.matmul(value, w) + b #定義一維卷積 def Conv1d(name, tensor, ksize, out_dim, stride, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w',[ksize,tensor.get_shape()[-1],out_dim],dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv1d(tensor,w,stride,padding=padding) b = tf.get_variable('b',[out_dim],'float32',initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) #定義二維卷積 def Conv2d(name, tensor, filter_size1 ,filter_size2, out_dim, stride1, stride2, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w',[filter_size1, filter_size2, tensor.get_shape()[-1], out_dim], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv2d(tensor, w, [1, stride1, stride2, 1], padding=padding) b = tf.get_variable('b',[out_dim], 'float32', initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) #定義二維反捲積 def Deconv2d(name, tensor, filter_size1, filter_size2, outshape, stride1, stride2, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w', [filter_size1, filter_size2, outshape[-1], tensor.get_shape()[-1]], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv2d_transpose(tensor, w, outshape, strides=[1,stride1, stride2, 1], padding=padding) b = tf.get_variable('b', [outshape[-1]],'float32', initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) def Get_inputs(real_size,noise_size): real_img = tf.placeholder(tf.float32, [None, real_size], name='real_img') noise_img = tf.placeholder(tf.float32, [None, noise_size], name='noise_img') return real_img, noise_img def Discriminator(img, reuse=False, name='discriminator'): with tf.variable_scope(name, reuse=reuse): output = Fully_connected('df1',img,2048) output = LeakyRelu(output) output = Fully_connected('df2',output,1024) output = LeakyRelu(output) output = Fully_connected('df3',output,512) output = LeakyRelu(output) output = Fully_connected('df4',output,256) output = LeakyRelu(output) output = Fully_connected('df5',output,1) return output def Generator(noise_img, reuse=False, name='generator'): with tf.variable_scope(name,reuse=reuse): output = Fully_connected('gf1',noise_img,2048) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf2',output,1024) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf3',output,512) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf4',output,256) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf5',output,784) output = tf.nn.tanh(output) return output mode = 'wgan' # gan, wgan, wgan-gp noise = 'normal' # normal0_1, normal, uniform batch_size = 100 epochs = 10 n_sample = 100 lamda = 10 img_size = 784 noise_size = 100 tf.reset_default_graph() real_img, noise_img = Get_inputs(img_size,noise_size)#feed於此 real_data = real_img fake_data = Generator(noise_img) disc_real = Discriminator(real_data,reuse=False) disc_fake = Discriminator(fake_data,reuse=True) #生成器和判別器中的tensor train_vars = tf.trainable_variables() g_vars = [var for var in train_vars if var.name.startswith("generator")] d_vars = [var for var in train_vars if var.name.startswith("discriminator")] #普通的GAN if mode == 'gan': gen_cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,labels=tf.ones_like(disc_fake))) #生成器loss disc_cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,labels=tf.zeros_like(disc_fake))) disc_cost += tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_real,labels=tf.ones_like(disc_real))) disc_cost /= 2. #判別器loss #優化器 gen_train_op = tf.train.AdamOptimizer(learning_rate=2e-4, beta1=0.5).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.AdamOptimizer(learning_rate=2e-4,beta1=0.5).minimize(disc_cost,var_list=d_vars) clip_disc_weights = None #wgan elif mode == 'wgan': gen_cost = -tf.reduce_mean(disc_fake) #生成器loss disc_cost = tf.reduce_mean(disc_fake) - tf.reduce_mean(disc_real) #判別器loss #優化器 gen_train_op = tf.train.RMSPropOptimizer(learning_rate=5e-5).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.RMSPropOptimizer(learning_rate=5e-5).minimize(disc_cost,var_list=d_vars) clip_ops = [] #將判別器權重截斷到[-0.01,0.01] for var in train_vars: if var.name.startswith("discriminator"): clip_bounds = [-0.01, 0.01] clip_ops.append(tf.assign(var,tf.clip_by_value(var,clip_bounds[0],clip_bounds[1]))) clip_disc_weights = tf.group(*clip_ops) elif mode == 'wgan-gp': gen_cost = -tf.reduce_mean(disc_fake) #生成器loss disc_cost = tf.reduce_mean(disc_fake) - tf.reduce_mean(disc_real) #判別器loss #梯度懲罰 alpha = tf.random_uniform(shape=[batch_size,1],minval=0.,maxval=1.) interpolates = alpha*fake_data + (1-alpha)*real_data gradients = tf.gradients(Discriminator(interpolates,reuse=True),[interpolates])[0] slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients),reduction_indices=[1])) gradient_penalty = tf.reduce_mean((slopes-1.)**2) disc_cost += lamda * gradient_penalty clip_disc_weights = None #優化器 gen_train_op = tf.train.AdamOptimizer(learning_rate=1e-4,beta1=0.5,beta2=0.9).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.AdamOptimizer(learning_rate=1e-4,beta1=0.5,beta2=0.9).minimize(disc_cost,var_list=d_vars) saver = tf.train.Saver() def Train(): losses = [] with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for e in range(epochs): for i in xrange(len(data)//batch_size): batch_images = data[i*batch_size:(i+1)*batch_size] batch_images = batch_images.reshape(batch_size,784) #batch = mnist.train.next_batch(batch_size) #batch_images = batch[0].reshape((batch_size,784)) if noise != 'normal0_1' : batch_images = batch_images*2 -1 if noise == 'uniform': batch_noise = np.random.uniform(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal': batch_noise = np.random.normal(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal0_1': batch_noise = np.random.normal(0, 1, size=(batch_size, noise_size)) if mode == 'gan': #普通的gan，判別器，生成器各訓練一次 disc_iters = 1 else: #wgan和wgan-gp，判別器訓練多次，生成器訓練一次 disc_iters = 6 for x in range(0, disc_iters): _,d_loss = sess.run([disc_train_op,disc_cost],feed_dict={real_data:batch_images,noise_img:batch_noise}) if clip_disc_weights is not None: _ = sess.run(clip_disc_weights) _,g_loss = sess.run([gen_train_op,gen_cost],feed_dict={noise_img:batch_noise}) Save_lossValue(e,i,d_loss,g_loss) print("第%d個epoch,第%d個batch , d_loss: %.8f, g_loss: %.8f"%(e, i, d_loss, g_loss)) losses.append((d_loss,g_loss)) if noise == 'uniform': sample_noise = np.random.uniform(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal': sample_noise = np.random.normal(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal0_1': sample_noise = np.random.normal(0, 1, size=(batch_size, noise_size)) gen_samples = sess.run(Generator(noise_img,reuse=True),feed_dict={noise_img:sample_noise}) print(gen_samples.shape) saver.save(sess,'checkpoints/test2.ckpt') if e % 1 == 0: gen = gen_samples.reshape(100,28,28) if noise != 'normal0_1': gen = (gen+1)/2 Save_genImages(gen, e) plot_loss(losses) def Test(): saver = tf.train.Saver(var_list=g_vars) with tf.Session() as sess: saver.restore(sess,tf.train.latest_checkpoint("checkpoints")) # saver.restore(sess,'checkppoints/test2.ckpt') if noise == 'uniform': sample_noise = np.random.uniform(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal': sample_noise = np.random.normal(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal0_1': sample_noise = np.random.normal(0, 1, size=(batch_size, noise_size)) gen_samples = sess.run(Generator(noise_img,reuse=True),feed_dict={noise_img:sample_noise}) if noise != 'normal0_1': gen_images = (gen_samples+1)/2 if __name__ == '__main__': Train() #Test()
from datetime import datetime import os import matplotlib.pyplot as plt import numpy as np import tensorflow as tf from six.moves import xrange data = np.load('final37.npy') data = data[:,:,0:60] #顯示原始資料影象 def Show_images(data,show_nums,save=False): index = 0 for n in range(show_nums): show_images = data[index:index+100] show_images = show_images.reshape(100,3,60,1) r,c = 10,10 fig,axs = plt.subplots(r,c) cnt = 0 for i in range(r): for j in range(c): xy = show_images[cnt] for k in range(len(xy)): x = xy[k][0:30] y = xy[k][30:60] if k == 0 : axs[i,j].plot(x,y,color='blue',linewidth=2) if k == 1: axs[i,j].plot(x,y,color='red',linewidth=2) if k == 2: axs[i,j].plot(x,y,color='green',linewidth=2) axs[i,j].axis('off') cnt += 1 index += 100 if save: if not os.path.exists('This_epoch'): os.makedirs('This_epoch') fig.savefig('This_epoch/%d.jpg' % n) plt.close() else: plt.show() def Save_genImages(gen, epoch): r,c = 10,10 fig,axs = plt.subplots(r,c) cnt = 0 for i in range(r): for j in range(c): xy = gen[cnt] for k in range(len(xy)): x = xy[k][0:30] y = xy[k][30:60] if k == 0: axs[i,j].plot(x,y,color='blue') if k == 1: axs[i,j].plot(x,y,color='red') if k == 2: axs[i,j].plot(x,y,color='green') axs[i,j].axis('off') cnt += 1 if not os.path.exists('gen_img1'): os.makedirs('gen_img1') fig.savefig('gen_img1/%d.jpg' % epoch) plt.close() def Save_lossValue(epoch,iters,d_loss,g_loss): with open('losst.txt','a') as f: f.write("第%d個epoch,第%d個batch , d_loss: %.8f, g_loss: %.8f"%(epoch, iters, d_loss, g_loss)+'\n') def plot_loss(loss): fig,ax = plt.subplots(figsize=(20,7)) losses = np.array(loss) plt.plot(losses.T[0], label="Discriminator Loss") plt.plot(losses.T[1], label="Generator Loss") plt.title("Training Losses") plt.legend() plt.savefig('loss.jpg') plt.show() #定義Relu啟用函式 def Relu(name, tensor): return tf.nn.relu(tensor,name) #定義LeakyRelu啟用函式 def LeakyRelu(name, x, leak=0.2): return tf.maximum(x,leak*x,name=name) #定義全連線層 def Fully_connected(name, value, output_shape): with tf.variable_scope(name,reuse=None) as scope: shape = value.get_shape().as_list() w = tf.get_variable('w',[shape[1],output_shape],dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=0.01)) b = tf.get_variable('b',[output_shape],dtype=tf.float32,initializer=tf.constant_initializer(0.0)) return tf.matmul(value,w) + b #定義一維卷積 def Conv1d(name, tensor, ksize, out_dim, stride, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w',[ksize,tensor.get_shape()[-1],out_dim],dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv1d(tensor,w,stride,padding=padding) b = tf.get_variable('b',[out_dim],'float32',initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) #定義二維卷積 def Conv2d(name, tensor, filter_size1 ,filter_size2, out_dim, stride1, stride2, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w',[filter_size1, filter_size2, tensor.get_shape()[-1], out_dim], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv2d(tensor, w, [1, stride1, stride2, 1], padding=padding) b = tf.get_variable('b',[out_dim], 'float32', initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) #定義二維反捲積 def Deconv2d(name, tensor, filter_size1, filter_size2, outshape, stride1, stride2, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w', [filter_size1, filter_size2, outshape[-1], tensor.get_shape()[-1]], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv2d_transpose(tensor, w, outshape, strides=[1,stride1, stride2, 1], padding=padding) b = tf.get_variable('b', [outshape[-1]],'float32', initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) def Get_inputs(real_size,noise_size): real_img = tf.placeholder(tf.float32, [None, real_size], name='real_img') noise_img = tf.placeholder(tf.float32, [None, noise_size], name='noise_img') return real_img, noise_img def Generator(noise_img, reuse=False, alpha=0.01): with tf.variable_scope('generator',reuse=reuse): # print(noise_img.shape) output = tf.layers.dense(noise_img,128) # print(output.shape) output = tf.maximum(alpha * output,output) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.layers.dropout(output, rate=0.25) output = tf.layers.dense(output,512) output = tf.maximum(alpha * output,output) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.layers.dropout(output,rate=0.25) output = tf.layers.dense(output,180) output = tf.tanh(output) return output def Discriminator(img,reuse=False,alpha=0.01): with tf.variable_scope("discriminator", reuse=reuse): print(img.shape) output = tf.layers.dense(img,512) output = tf.maximum(alpha * output, output) output = tf.layers.dense(output,128) output = tf.maximum(alpha * output, output) output = tf.layers.dense(output,1) return output mode = 'gan' #gan, wgan, wgan-gp batch_size = 100 epochs = 1 n_sample = 100 learning_rate = 0.0002 lamda = 10 img_size = 180 noise_size = 100 tf.reset_default_graph() real_img, noise_img = Get_inputs(img_size,noise_size)#feed於此 real_data = real_img fake_data = Generator(noise_img) disc_real = Discriminator(real_data,reuse=False) disc_fake = Discriminator(fake_data,reuse=True) #生成器和判別器中的tensor train_vars = tf.trainable_variables() g_vars = [var for var in train_vars if var.name.startswith("generator")] d_vars = [var for var in train_vars if var.name.startswith("discriminator")] #普通的GAN if mode == 'gan': gen_cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,labels=tf.ones_like(disc_fake))) #生成器loss disc_cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,labels=tf.zeros_like(disc_fake))) disc_cost += tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_real,labels=tf.ones_like(disc_real))) disc_cost /= 2. #判別器loss #優化器 gen_train_op = tf.train.AdamOptimizer(learning_rate=2e-4, beta1=0.5).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.AdamOptimizer(learning_rate=2e-4,beta1=0.5).minimize(disc_cost,var_list=d_vars) clip_disc_weights = None #wgan elif mode == 'wgan': gen_cost = -tf.reduce_mean(disc_fake) #生成器loss disc_cost = tf.reduce_mean(disc_fake) - tf.reduce_mean(disc_real) #判別器loss #優化器 gen_train_op = tf.train.RMSPropOptimizer(learning_rate=5e-5).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.RMSPropOptimizer(learning_rate=5e-5).minimize(disc_cost,var_list=d_vars) clip_ops = [] #將判別器權重截斷到[-0.01,0.01] for var in train_vars: if var.name.startswith("discriminator"): clip_bounds = [-0.01, 0.01] clip_ops.append(tf.assign(var,tf.clip_by_value(var,clip_bounds[0],clip_bounds[1]))) clip_disc_weights = tf.group(*clip_ops) elif mode == 'wgan-gp': gen_cost = -tf.reduce_mean(disc_fake) #生成器loss disc_cost = tf.reduce_mean(disc_fake) - tf.reduce_mean(disc_real) #判別器loss #梯度懲罰 alpha = tf.random_uniform(shape=[batch_size,1],minval=0.,maxval=1.) interpolates = alpha*fake_data + (1-alpha)*real_data gradients = tf.gradients(Discriminator(interpolates,reuse=True),[interpolates])[0] slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients),reduction_indices=[1])) gradient_penalty = tf.reduce_mean((slopes-1.)**2) disc_cost += lamda * gradient_penalty clip_disc_weights = None #優化器 gen_train_op = tf.train.AdamOptimizer(learning_rate=1e-4,beta1=0.5,beta2=0.9).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.AdamOptimizer(learning_rate=1e-4,beta1=0.5,beta2=0.9).minimize(disc_cost,var_list=d_vars) saver = tf.train.Saver() def Train(): losses = [] with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for e in range(epochs): for i in xrange(len(data)//batch_size): batch_images = data[i*batch_size:(i+1)*batch_size] batch_images = batch_images.reshape(batch_size,180) batch_images = batch_images*2 -1 batch_noise = np.random.uniform(-1,1,size=(batch_size,noise_size)) if mode == 'gan': #普通的gan，判別器，生成器各訓練一次 disc_iters = 2 else: #wgan和wgan-gp，判別器訓練多次，生成器訓練一次 disc_iters = 2 for x in range(0,disc_iters): _,d_loss = sess.run([disc_train_op,disc_cost],feed_dict={real_data:batch_images,noise_img:batch_noise}) if clip_disc_weights is not None: _ = sess.run(clip_disc_weights) _,g_loss = sess.run([gen_train_op,gen_cost],feed_dict={noise_img:batch_noise}) Save_lossValue(e,i,d_loss,g_loss) print("第%d個epoch,第%d個batch , d_loss: %.8f, g_loss: %.8f"%(e, i, d_loss, g_loss)) losses.append((d_loss,g_loss)) sample_noise = np.random.uniform(-1,1,size=(100,100)) gen_samples = sess.run(Generator(noise_img,reuse=True),feed_dict={noise_img:sample_noise}) print(gen_samples.shape) saver.save(sess,'checkpoints/test.ckpt') if e % 1 == 0: gen = gen_samples.reshape(100,3,60,1) Save_genImages(gen, e) plot_loss(losses) def Test(): saver = tf.train.Saver(var_list=g_vars) with tf.Session() as sess: saver.restore(sess,tf.train.latest_checkpoint("checkpoints")) # saver.restore(sess,'checkppoints/b.ckpt') sample_noise = np.random.uniform(-1, 1, size=(10000,noise_size)) gen_samples = sess.run(Generator(noise_img,reuse=True),feed_dict={noise_img:sample_noise}) gen_images = (gen_samples+1)/2 show_num = len(gen_images)//100 Show_images(gen_images,show_num,save=True) if __name__ == '__main__': Train() #Test()
from datetime import datetime import os import matplotlib.pyplot as plt import numpy as np import tensorflow as tf from six.moves import xrange data = np.load('data/final37.npy') data = data[:,:,0:60] #顯示原始資料影象 def Show_images(data,show_nums,save=False): index = 0 for n in range(show_nums): show_images = data[index:index+100] show_images = show_images.reshape(100,3,60,1) r,c = 10,10 fig,axs = plt.subplots(r,c) cnt = 0 for i in range(r): for j in range(c): xy = show_images[cnt] for k in range(len(xy)): x = xy[k][0:30] y = xy[k][30:60] if k == 0 : axs[i,j].plot(x,y,color='blue',linewidth=2) if k == 1: axs[i,j].plot(x,y,color='red',linewidth=2) if k == 2: axs[i,j].plot(x,y,color='green',linewidth=2) axs[i,j].axis('off') cnt += 1 index += 100 if save: if not os.path.exists('This_epoch2'): os.makedirs('This_epoch2') fig.savefig('This_epoch2/%d.jpg' % n) plt.close() else: plt.show() def Save_genImages(gen, epoch): r,c = 10,10 fig,axs = plt.subplots(r,c) cnt = 0 for i in range(r): for j in range(c): xy = gen[cnt] for k in range(len(xy)): x = xy[k][0:30] y = xy[k][30:60] if k == 0: axs[i,j].plot(x,y,color='blue') if k == 1: axs[i,j].plot(x,y,color='red') if k == 2: axs[i,j].plot(x,y,color='green') axs[i,j].axis('off') cnt += 1 if not os.path.exists('gen_img2'): os.makedirs('gen_img2') fig.savefig('gen_img2/%d.jpg' % epoch) plt.close() def Save_lossValue(epoch,iters,d_loss,g_loss): with open('loss2.txt','a') as f: f.write("第%d個epoch,第%d個batch , d_loss: %.8f, g_loss: %.8f"%(epoch, iters, d_loss, g_loss)+'\n') def plot_loss(loss): fig,ax = plt.subplots(figsize=(20,7)) losses = np.array(loss) plt.plot(losses.T[0], label="Discriminator Loss") plt.plot(losses.T[1], label="Generator Loss") plt.title("Training Losses") plt.legend() plt.savefig('loss2.jpg') plt.show() #定義Relu啟用函式 def Relu(name, tensor): return tf.nn.relu(tensor,name) #定義LeakyRelu啟用函式 def LeakyRelu(x, alpha=0.2): return tf.maximum(x, alpha * x) #定義全連線層 def Fully_connected(name, value, output_shape): with tf.variable_scope(name, reuse=None) as scope: shape = value.get_shape().as_list() w = tf.get_variable('w', [shape[1], output_shape], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=0.01)) b = tf.get_variable('b', [output_shape], dtype=tf.float32, initializer=tf.constant_initializer(0.0)) return tf.matmul(value, w) + b #定義一維卷積 def Conv1d(name, tensor, ksize, out_dim, stride, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w',[ksize,tensor.get_shape()[-1],out_dim],dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv1d(tensor,w,stride,padding=padding) b = tf.get_variable('b',[out_dim],'float32',initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) #定義二維卷積 def Conv2d(name, tensor, filter_size1 ,filter_size2, out_dim, stride1, stride2, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w',[filter_size1, filter_size2, tensor.get_shape()[-1], out_dim], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv2d(tensor, w, [1, stride1, stride2, 1], padding=padding) b = tf.get_variable('b',[out_dim], 'float32', initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) #定義二維反捲積 def Deconv2d(name, tensor, filter_size1, filter_size2, outshape, stride1, stride2, padding, stddev=0.01): with tf.variable_scope(name): w = tf.get_variable('w', [filter_size1, filter_size2, outshape[-1], tensor.get_shape()[-1]], dtype=tf.float32, initializer=tf.random_normal_initializer(stddev=stddev)) var = tf.nn.conv2d_transpose(tensor, w, outshape, strides=[1,stride1, stride2, 1], padding=padding) b = tf.get_variable('b', [outshape[-1]],'float32', initializer=tf.constant_initializer(0.01)) return tf.nn.bias_add(var,b) def Get_inputs(real_size,noise_size): real_img = tf.placeholder(tf.float32, [None, real_size], name='real_img') noise_img = tf.placeholder(tf.float32, [None, noise_size], name='noise_img') return real_img, noise_img def Get_noise(noise,batch_size): if noise == 'uniform': batch_noise = np.random.uniform(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal': batch_noise = np.random.normal(-1, 1, size=(batch_size, noise_size)) elif noise == 'normal0_1': batch_noise = np.random.normal(0, 1, size=(batch_size, noise_size)) return batch_noise def Discriminator(img, reuse=False, name='discriminator'): with tf.variable_scope(name, reuse=reuse): output = Fully_connected('df1',img,2048) output = LeakyRelu(output) output = Fully_connected('df2',output,1024) output = LeakyRelu(output) output = Fully_connected('df3',output,512) output = LeakyRelu(output) output = Fully_connected('df4',output,256) output = LeakyRelu(output) output = Fully_connected('df5',output,1) return output def Generator(noise_img, reuse=False, name='generator'): with tf.variable_scope(name,reuse=reuse): output = Fully_connected('gf1',noise_img,2048) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf2',output,1024) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf3',output,512) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf4',output,256) output = tf.layers.batch_normalization(output,momentum=0.8,training=True) output = tf.nn.relu(output) output = Fully_connected('gf5',output,180) output = tf.nn.tanh(output) return output mode = 'wgan-gp' # gan, wgan, wgan-gp noise = 'uniform' # normal0_1, normal, uniform batch_size = 100 epochs = 100 n_sample = 100 lamda = 10 img_size = 180 noise_size = 100 tf.reset_default_graph() real_img, noise_img = Get_inputs(img_size,noise_size)#feed於此 real_data = real_img fake_data = Generator(noise_img) disc_real = Discriminator(real_data,reuse=False) disc_fake = Discriminator(fake_data,reuse=True) #生成器和判別器中的tensor train_vars = tf.trainable_variables() g_vars = [var for var in train_vars if var.name.startswith("generator")] d_vars = [var for var in train_vars if var.name.startswith("discriminator")] #普通的GAN if mode == 'gan': gen_cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,labels=tf.ones_like(disc_fake))) #生成器loss disc_cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,labels=tf.zeros_like(disc_fake))) disc_cost += tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_real,labels=tf.ones_like(disc_real))) disc_cost /= 2. #判別器loss #優化器 gen_train_op = tf.train.AdamOptimizer(learning_rate=2e-4, beta1=0.5).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.AdamOptimizer(learning_rate=2e-4,beta1=0.5).minimize(disc_cost,var_list=d_vars) clip_disc_weights = None #wgan elif mode == 'wgan': gen_cost = -tf.reduce_mean(disc_fake) #生成器loss disc_cost = tf.reduce_mean(disc_fake) - tf.reduce_mean(disc_real) #判別器loss #優化器 gen_train_op = tf.train.RMSPropOptimizer(learning_rate=5e-5).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.RMSPropOptimizer(learning_rate=5e-5).minimize(disc_cost,var_list=d_vars) clip_ops = [] #將判別器權重截斷到[-0.01,0.01] for var in train_vars: if var.name.startswith("discriminator"): clip_bounds = [-0.01, 0.01] clip_ops.append(tf.assign(var,tf.clip_by_value(var,clip_bounds[0],clip_bounds[1]))) clip_disc_weights = tf.group(*clip_ops) elif mode == 'wgan-gp': gen_cost = -tf.reduce_mean(disc_fake) #生成器loss disc_cost = tf.reduce_mean(disc_fake) - tf.reduce_mean(disc_real) #判別器loss #梯度懲罰 alpha = tf.random_uniform(shape=[batch_size,1],minval=0.,maxval=1.) interpolates = alpha*fake_data + (1-alpha)*real_data gradients = tf.gradients(Discriminator(interpolates,reuse=True),[interpolates])[0] slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients),reduction_indices=[1])) gradient_penalty = tf.reduce_mean((slopes-1.)**2) disc_cost += lamda * gradient_penalty clip_disc_weights = None #優化器 gen_train_op = tf.train.AdamOptimizer(learning_rate=1e-4,beta1=0.5,beta2=0.9).minimize(gen_cost,var_list=g_vars) disc_train_op = tf.train.AdamOptimizer(learning_rate=1e-4,beta1=0.5,beta2=0.9).minimize(disc_cost,var_list=d_vars) saver = tf.train.Saver() def Train(): losses = [] with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for e in range(epochs): for i in xrange(len(data)//batch_size): batch_images = data[i*batch_size:(i+1)*batch_size] batch_images = batch_images.reshape(batch_size,180) if noise != 'normal0_1' : batch_images = batch_images*2 -1 batch_noise = Get_noise(noise,100) if mode == 'gan': #普通的gan，判別器，生成器各訓練一次 disc_iters = 1 else: #wgan和wgan-gp，判別器訓練多次，生成器訓練一次 disc_iters = 6 for x in range(0, disc_iters): _,d_loss = sess.run([disc_train_op,disc_cost],feed_dict={real_data:batch_images,noise_img:batch_noise}) if clip_disc_weights is not None: _ = sess.run(clip_disc_weights) _,g_loss = sess.run([gen_train_op,gen_cost],feed_dict={noise_img:batch_noise}) Save_lossValue(e,i,d_loss,g_loss) print("第%d個epoch,第%d個batch , d_loss: %.8f, g_loss: %.8f"%(e, i, d_loss, g_loss)) losses.append((d_loss,g_loss)) sample_noise = Get_noise(noise,100) gen_samples = sess.run(Generator(noise_img,reuse=True),feed_dict={noise_img:sample_noise}) saver.save(sess,'checkpoints/test2.ckpt') if e % 1 == 0: gen = gen_samples.reshape(100,3,60,1) if noise != 'normal0_1': gen = (gen+1)/2 Save_genImages(gen, e) plot_loss(losses) def Test(): saver = tf.train.Saver(var_list=g_vars) with tf.Session() as sess: saver.restore(sess,tf.train.latest_checkpoint("checkpoints")) # saver.restore(sess,'checkppoints/test2.ckpt') sample_noise = Get_noise(noise,10000) gen_samples = sess.run(Generator(noise_img,reuse=True),feed_dict={noise_img:sample_noise}) if noise != 'normal0_1': gen_images = (gen_samples+1)/2 show_num = len(gen_images)//100 Show_images(gen_images,show_num,save=True) if __name__ == '__main__': Train() #Test()