In [1]:
import os
os.environ['KERAS_BACKEND']='tensorflow' # 也可以使用 tensorflow
#os.environ['THEANO_FLAGS']='floatX=float32,device=cuda,exception_verbosity=high'
os.environ['THEANO_FLAGS']='floatX=float32,device=cuda,optimizer=fast_compile'
modifed from https://github.com/martinarjovsky/WassersteinGAN
In [2]:
import keras.backend as K
K.set_image_data_format('channels_first')
from keras.models import Sequential, Model
from keras.layers import Conv2D, ZeroPadding2D, BatchNormalization, Input
from keras.layers import Conv2DTranspose, Reshape, Activation, Cropping2D, Flatten
from keras.layers.advanced_activations import LeakyReLU
from keras.activations import relu
from keras.initializers import RandomNormal
conv_init = RandomNormal(0, 0.02)
gamma_init = RandomNormal(1., 0.02)
In [3]:
def DCGAN_D(isize, nz, nc, ndf, n_extra_layers=0):
assert isize%2==0
_ = inputs = Input(shape=(nc, isize, isize))
_ = Conv2D(filters=ndf, kernel_size=4, strides=2, use_bias=False,
padding = "same",
kernel_initializer = conv_init,
name = 'initial.conv.{0}-{1}'.format(nc, ndf)
) (_)
_ = LeakyReLU(alpha=0.2, name = 'initial.relu.{0}'.format(ndf))(_)
csize, cndf = isize// 2, ndf
while csize > 5:
assert csize%2==0
in_feat = cndf
out_feat = cndf*2
_ = Conv2D(filters=out_feat, kernel_size=4, strides=2, use_bias=False,
padding = "same",
kernel_initializer = conv_init,
name = 'pyramid.{0}-{1}.conv'.format(in_feat, out_feat)
) (_)
if 0: # toggle batchnormalization
_ = BatchNormalization(name = 'pyramid.{0}.batchnorm'.format(out_feat),
momentum=0.9, axis=1, epsilon=1.01e-5,
gamma_initializer = gamma_init,
)(_, training=1)
_ = LeakyReLU(alpha=0.2, name = 'pyramid.{0}.relu'.format(out_feat))(_)
csize, cndf = (csize+1)//2, cndf*2
_ = Conv2D(filters=1, kernel_size=csize, strides=1, use_bias=False,
kernel_initializer = conv_init,
name = 'final.{0}-{1}.conv'.format(cndf, 1)
) (_)
outputs = Flatten()(_)
return Model(inputs=inputs, outputs=outputs)
In [4]:
def DCGAN_G(isize, nz, nc, ngf, n_extra_layers=0):
cngf= ngf//2
tisize = isize
while tisize > 5:
cngf = cngf * 2
assert tisize%2==0
tisize = tisize // 2
_ = inputs = Input(shape=(nz,))
_ = Reshape((nz, 1,1))(_)
_ = Conv2DTranspose(filters=cngf, kernel_size=tisize, strides=1, use_bias=False,
kernel_initializer = conv_init,
name = 'initial.{0}-{1}.convt'.format(nz, cngf))(_)
_ = BatchNormalization(gamma_initializer = gamma_init, momentum=0.9, axis=1, epsilon=1.01e-5,
name = 'initial.{0}.batchnorm'.format(cngf))(_, training=1)
_ = Activation("relu", name = 'initial.{0}.relu'.format(cngf))(_)
csize, cndf = tisize, cngf
while csize < isize//2:
in_feat = cngf
out_feat = cngf//2
_ = Conv2DTranspose(filters=out_feat, kernel_size=4, strides=2, use_bias=False,
kernel_initializer = conv_init, padding="same",
name = 'pyramid.{0}-{1}.convt'.format(in_feat, out_feat)
) (_)
_ = BatchNormalization(gamma_initializer = gamma_init,
momentum=0.9, axis=1, epsilon=1.01e-5,
name = 'pyramid.{0}.batchnorm'.format(out_feat))(_, training=1)
_ = Activation("relu", name = 'pyramid.{0}.relu'.format(out_feat))(_)
csize, cngf = csize*2, cngf//2
_ = Conv2DTranspose(filters=nc, kernel_size=4, strides=2, use_bias=False,
kernel_initializer = conv_init, padding="same",
name = 'final.{0}-{1}.convt'.format(cngf, nc)
)(_)
outputs = Activation("tanh", name = 'final.{0}.tanh'.format(nc))(_)
return Model(inputs=inputs, outputs=outputs)
Parameters
In [5]:
nc = 3
nz = 100
ngf = 64
ndf = 64
n_extra_layers = 0
Diters = 5
λ = 10
imageSize = 32
batchSize = 64
lrD = 1e-4
lrG = 1e-4
print models
In [6]:
netD = DCGAN_D(imageSize, nz, nc, ndf, n_extra_layers)
netD.summary()
In [7]:
netG = DCGAN_G(imageSize, nz, nc, ngf, n_extra_layers)
netG.summary()
In [8]:
from keras.optimizers import RMSprop, SGD, Adam
compute Wasserstein loss and gradient penalty
In [9]:
netD_real_input = Input(shape=(nc, imageSize, imageSize))
noisev = Input(shape=(nz,))
netD_fake_input = netG(noisev)
ϵ_input = K.placeholder(shape=(None,1,1,1))
netD_mixed_input = Input(shape=(nc, imageSize, imageSize),
tensor=ϵ_input * netD_real_input + (1-ϵ_input) * netD_fake_input)
loss_real = K.mean(netD(netD_real_input))
loss_fake = K.mean(netD(netD_fake_input))
grad_mixed = K.gradients(netD(netD_mixed_input), [netD_mixed_input])[0]
norm_grad_mixed = K.sqrt(K.sum(K.square(grad_mixed), axis=[1,2,3]))
grad_penalty = K.mean(K.square(norm_grad_mixed -1))
loss = loss_fake - loss_real + λ * grad_penalty
training_updates = Adam(lr=lrD, beta_1=0.0, beta_2=0.9).get_updates(netD.trainable_weights,[],loss)
netD_train = K.function([netD_real_input, noisev, ϵ_input],
[loss_real, loss_fake],
training_updates)
loss for netG
In [10]:
loss = -loss_fake
training_updates = Adam(lr=lrG, beta_1=0.0, beta_2=0.9).get_updates(netG.trainable_weights,[], loss)
netG_train = K.function([noisev], [loss], training_updates)
Download CIFAR10 if needed
In [11]:
from PIL import Image
import numpy as np
import tarfile
# Download dataset
url = "https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz"
import os
import urllib
from urllib.request import urlretrieve
def reporthook(a,b,c):
print("\rdownloading: %5.1f%%"%(a*b*100.0/c), end="")
tar_gz = "cifar-10-python.tar.gz"
if not os.path.isfile(tar_gz):
print('Downloading data from %s' % url)
urlretrieve(url, tar_gz, reporthook=reporthook)
import pickle
train_X=[]
train_y=[]
tar_gz = "cifar-10-python.tar.gz"
with tarfile.open(tar_gz) as tarf:
for i in range(1, 6):
dataset = "cifar-10-batches-py/data_batch_%d"%i
print("load",dataset)
with tarf.extractfile(dataset) as f:
result = pickle.load(f, encoding='latin1')
train_X.extend( result['data'].reshape(-1,3,32,32)/255*2-1)
train_y.extend(result['labels'])
train_X=np.float32(train_X)
train_y=np.int32(train_y)
dataset = "cifar-10-batches-py/test_batch"
print("load",dataset)
with tarf.extractfile(dataset) as f:
result = pickle.load(f, encoding='latin1')
test_X=np.float32(result['data'].reshape(-1,3,32,32)/255*2-1)
test_y=np.int32(result['labels'])
also using test_X
In [12]:
train_X = np.concatenate([train_X, test_X])
train_X = np.concatenate([train_X[:,:,:,::-1], train_X])
utility to show images
In [13]:
from IPython.display import display
def showX(X, rows=1):
assert X.shape[0]%rows == 0
int_X = ( (X+1)/2*255).clip(0,255).astype('uint8')
# N*3072 -> N*3*32*32 -> 32 * 32N * 3
int_X = np.moveaxis(int_X.reshape(-1,3,32,32), 1, 3)
int_X = int_X.reshape(rows, -1, 32, 32,3).swapaxes(1,2).reshape(rows*32,-1, 3)
display(Image.fromarray(int_X))
# 訓練資料， X 的前 20 筆
showX(train_X[:20])
print(train_y[:20])
name_array = np.array("airplane car bird cat deer dog frog horse boat truck".split(' '))
print(name_array[train_y[:20]])
In [14]:
fixed_noise = np.random.normal(size=(batchSize, nz)).astype('float32')
In [15]:
import time
t0 = time.time()
niter = 100
gen_iterations = 0
errG = 0
targetD = np.float32([2]*batchSize+[-2]*batchSize)[:, None]
targetG = np.ones(batchSize, dtype=np.float32)[:, None]
for epoch in range(niter):
i = 0
# 每個 epoch 洗牌一下
np.random.shuffle(train_X)
batches = train_X.shape[0]//batchSize
while i < batches:
if gen_iterations < 25 or gen_iterations % 500 == 0:
_Diters = 100
else:
_Diters = Diters
j = 0
while j < _Diters and i < batches:
j+=1
real_data = train_X[i*batchSize:(i+1)*batchSize]
i+=1
noise = np.random.normal(size=(batchSize, nz))
ϵ = np.random.uniform(size=(batchSize, 1, 1 ,1))
errD_real, errD_fake = netD_train([real_data, noise, ϵ])
errD = errD_real - errD_fake
if gen_iterations%500==0:
print('[%d/%d][%d/%d][%d] Loss_D: %f Loss_G: %f Loss_D_real: %f Loss_D_fake %f'
% (epoch, niter, i, batches, gen_iterations,errD, errG, errD_real, errD_fake), time.time()-t0)
fake = netG.predict(fixed_noise)
showX(fake, 4)
noise = np.random.normal(size=(batchSize, nz))
errG, = netG_train([noise])
gen_iterations+=1
In [ ]:
In [ ]: