gan.py

import numpy as np
import tensorflow as tf
from keras.models import Model
from keras.layers import Input
from keras.layers.core import Reshape, Dense, Dropout, Flatten
from keras.layers.advanced_activations import LeakyReLU
from keras.losses import MSE
from keras.optimizers import SGD
from keras import backend as K
from keras import initializers
from visdom import Visdom
​
viz = Visdom()
np.random.seed(1000)
sess = tf.Session()
K.set_session(sess)
​
randomDim = 100
​
# given two different random seed
seed_true = np.random.uniform(-1, 1, size=100*randomDim).reshape(100, randomDim)
seed_test = np.random.uniform(-3, 3, size=100*randomDim).reshape(100, randomDim)
seed_var = K.variable(seed_test)
​
​
seed = Input(shape=(randomDim,), tensor=seed_var)
_ = Dense(256, input_dim=randomDim, kernel_initializer=initializers.RandomNormal(stddev=0.02))(seed)
_ = LeakyReLU(0.2)(_)
_ = Dense(512)(_)
_ = LeakyReLU(0.2)(_)
_ = Dense(1024)(_)
_ = LeakyReLU(0.2)(_)
image = Dense(784, activation='tanh')(_)
generator = Model(seed, image)
generator.load_weights("./models/gan_generator_epoch_200.h5")
​
D_true = sess.run(generator.output , {seed: seed_true})
viz.images(D_true.reshape(100, 1, 28, 28), nrow=10, win='image_true', opts={'caption': 'image_true'})
​
noise = np.random.normal(0, 0.5, size=100*784).reshape(100, 784)
D_noise = D_true + noise
viz.images(D_noise.reshape(100, 1, 28, 28), nrow=10, win='image_noise', opts={'caption': 'image_noise'})
​
D_test = sess.run(generator.output, {seed: seed_test})
viz.images(D_test.reshape(100, 1, 28, 28), nrow=10, win='image_test', opts={'caption': 'image_test'})
viz.images(D_test.reshape(100, 1, 28, 28), nrow=10, win='image_recover', opts={'caption': 'image_recover'})
​
opt = SGD()
loss_func = K.mean(MSE(D_noise, generator.output), axis=None)
dx = opt.get_gradients(loss_func, generator.input)
old = seed_var
new = K.clip(seed_var - 0.3 * dx[0], -2., 2.)
train = K.function([], [loss_func], [(old, new)])
​
step = 0
loss_value = [np.inf]
​
while loss_value[0] > 0.05:
    loss_value = train([])
    if step % 1000 == 0:
        
        seed_test = K.eval(seed_var)
        D_test = sess.run(generator.output , {seed: seed_test})
        viz.images(D_test.reshape(100, 1, 28, 28), nrow=10, win='image_recover', opts={'caption': 'image_recover'})
        update = 'append' if viz.win_exists('loss') else None
        viz.line(X=np.array([step]), Y=np.array(loss_value), win='loss', update=update)
​
    step += 1