Merge branch 'master' of https://github.com/Jittor/gan-jittor
This commit is contained in:
commit
110b0d7c12
|
|
@ -0,0 +1,212 @@
|
|||
import jittor as jt
|
||||
from jittor import init
|
||||
import argparse
|
||||
import os
|
||||
import numpy as np
|
||||
import math
|
||||
from jittor.dataset.mnist import MNIST
|
||||
import jittor.transform as transform
|
||||
from jittor import nn
|
||||
import cv2
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--n_epochs", type=int, default=200, help="number of epochs of training")
|
||||
parser.add_argument("--batch_size", type=int, default=64, help="size of the batches")
|
||||
parser.add_argument("--lr", type=float, default=0.0002, help="adam: learning rate")
|
||||
parser.add_argument("--b1", type=float, default=0.5, help="adam: decay of first order momentum of gradient")
|
||||
parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
|
||||
parser.add_argument("--n_cpu", type=int, default=8, help="number of cpu threads to use during batch generation")
|
||||
parser.add_argument("--latent_dim", type=int, default=62, help="dimensionality of the latent space")
|
||||
parser.add_argument("--img_size", type=int, default=32, help="size of each image dimension")
|
||||
parser.add_argument("--channels", type=int, default=1, help="number of image channels")
|
||||
parser.add_argument("--sample_interval", type=int, default=400, help="number of image channels")
|
||||
opt = parser.parse_args()
|
||||
|
||||
transforms = transform.Compose([
|
||||
transform.Resize(opt.img_size),
|
||||
transform.Gray(),
|
||||
transform.ImageNormalize(mean=[0.5],std=[0.5]),
|
||||
])
|
||||
dataloader = MNIST(train=True,transform=transforms).set_attrs(batch_size=opt.batch_size,shuffle=True)
|
||||
os.makedirs("images",exist_ok=True)
|
||||
jt.flags.use_cuda = 1
|
||||
def save_image(img, path, nrow=10):
|
||||
N,C,W,H = img.shape
|
||||
img2=img.reshape([-1,W*nrow*nrow,H])
|
||||
img=img2[:,:W*nrow,:]
|
||||
for i in range(1,nrow):
|
||||
img=np.concatenate([img,img2[:,W*nrow*i:W*nrow*(i+1),:]],axis=2)
|
||||
min_=img.min()
|
||||
max_=img.max()
|
||||
img=(img-min_)/(max_-min_)*255
|
||||
img=img.transpose((1,2,0))
|
||||
cv2.imwrite(path,img)
|
||||
|
||||
class BatchNorm1d(nn.Module):
|
||||
def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=None, is_train=True, sync=True):
|
||||
assert affine == None
|
||||
|
||||
self.sync = sync
|
||||
self.num_features = num_features
|
||||
self.is_train = is_train
|
||||
self.eps = eps
|
||||
self.momentum = momentum
|
||||
self.weight = init.constant((num_features,), "float32", 1.0)
|
||||
self.bias = init.constant((num_features,), "float32", 0.0)
|
||||
self.running_mean = init.constant((num_features,), "float32", 0.0).stop_grad()
|
||||
self.running_var = init.constant((num_features,), "float32", 1.0).stop_grad()
|
||||
|
||||
def execute(self, x):
|
||||
if self.is_train:
|
||||
xmean = jt.mean(x, dims=[0], keepdims=1)
|
||||
x2mean = jt.mean(x*x, dims=[0], keepdims=1)
|
||||
if self.sync and jt.mpi:
|
||||
xmean = xmean.mpi_all_reduce("mean")
|
||||
x2mean = x2mean.mpi_all_reduce("mean")
|
||||
|
||||
xvar = x2mean-xmean*xmean
|
||||
norm_x = (x-xmean)/jt.sqrt(xvar+self.eps)
|
||||
self.running_mean += (xmean.sum([0])-self.running_mean)*self.momentum
|
||||
self.running_var += (xvar.sum([0])-self.running_var)*self.momentum
|
||||
else:
|
||||
running_mean = self.running_mean.broadcast(x, [0])
|
||||
running_var = self.running_var.broadcast(x, [0])
|
||||
norm_x = (x-running_mean)/jt.sqrt(running_var+self.eps)
|
||||
w = self.weight.broadcast(x, [0])
|
||||
b = self.bias.broadcast(x, [0])
|
||||
return norm_x * w + b
|
||||
|
||||
def weights_init_normal(m):
|
||||
classname = m.__class__.__name__
|
||||
if (classname.find('Conv') != (- 1)):
|
||||
init.gauss_(m.weight.data, mean=0.0, std=0.02)
|
||||
elif (classname.find('BatchNorm') != (- 1)):
|
||||
init.gauss_(m.weight.data, mean=1.0, std=0.02)
|
||||
init.constant_(m.bias.data, value=0.0)
|
||||
|
||||
class Generator(nn.Module):
|
||||
|
||||
def __init__(self):
|
||||
super(Generator, self).__init__()
|
||||
self.init_size = (opt.img_size // 4)
|
||||
self.l1 = nn.Sequential(nn.Linear(opt.latent_dim, (128 * (self.init_size ** 2))))
|
||||
self.conv_blocks = nn.Sequential(
|
||||
nn.Upsample(scale_factor=2),
|
||||
nn.Conv(128, 128, 3, stride=1, padding=1),
|
||||
nn.BatchNorm(128,0.8),
|
||||
nn.Leaky_relu(0.2),
|
||||
nn.Upsample(scale_factor=2),
|
||||
nn.Conv(128, 64, 3, stride=1, padding=1),
|
||||
nn.BatchNorm(64, 0.8),
|
||||
nn.Leaky_relu(0.2),
|
||||
nn.Conv(64, opt.channels, 3, stride=1, padding=1),
|
||||
nn.Tanh()
|
||||
)
|
||||
|
||||
def execute(self, noise):
|
||||
out = self.l1(noise)
|
||||
#print(out)
|
||||
out = out.reshape((out.shape[0], 128, self.init_size, self.init_size))
|
||||
img = self.conv_blocks(out)
|
||||
#print(img)
|
||||
return img
|
||||
|
||||
class Discriminator(nn.Module):
|
||||
|
||||
def __init__(self):
|
||||
super(Discriminator, self).__init__()
|
||||
self.down = nn.Sequential(nn.Conv(opt.channels, 64, 3, stride=2, padding=1), nn.ReLU())
|
||||
self.down_size = (opt.img_size // 2)
|
||||
down_dim = (64 * ((opt.img_size // 2) ** 2))
|
||||
self.embedding = nn.Linear(down_dim, 32)
|
||||
self.fc = nn.Sequential(
|
||||
BatchNorm1d(32, 0.8),
|
||||
nn.ReLU(),
|
||||
nn.Linear(32, down_dim),
|
||||
BatchNorm1d(down_dim),
|
||||
nn.ReLU()
|
||||
)
|
||||
self.up = nn.Sequential(nn.Upsample(scale_factor=2), nn.Conv(64, opt.channels, 3, stride=1, padding=1))
|
||||
|
||||
def execute(self, img):
|
||||
out = self.down(img)
|
||||
embedding = self.embedding(out.reshape((out.shape[0], (- 1))))
|
||||
#print(embedding.shape)
|
||||
out = self.fc(embedding)
|
||||
out = self.up(out.reshape((out.shape[0], 64, self.down_size, self.down_size)))
|
||||
return (out, embedding)
|
||||
|
||||
# Reconstruction loss of AE
|
||||
pixelwise_loss = nn.MSELoss()
|
||||
|
||||
# Initialize generator and discriminator
|
||||
generator = Generator()
|
||||
discriminator = Discriminator()
|
||||
|
||||
#generator.apply(weights_init_normal)
|
||||
#discriminator.apply(weights_init_normal)
|
||||
|
||||
optimizer_G = nn.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
|
||||
optimizer_D = nn.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
|
||||
|
||||
def pullaway_loss(embeddings):
|
||||
norm = jt.sqrt((embeddings ** 2).sum(1,keepdims=True))
|
||||
normalized_emb = embeddings / norm
|
||||
similarity = jt.matmul(normalized_emb, normalized_emb.transpose(1, 0))
|
||||
batch_size = embeddings.size(0)
|
||||
loss_pt = (jt.sum(similarity) - batch_size) / (batch_size * (batch_size - 1))
|
||||
return loss_pt
|
||||
|
||||
|
||||
# ----------
|
||||
# Training
|
||||
# ----------
|
||||
|
||||
# BEGAN hyper parameters
|
||||
lambda_pt = 0.1
|
||||
margin = max(1, opt.batch_size / 64.0)
|
||||
for epoch in range(opt.n_epochs):
|
||||
for i, (imgs, _) in enumerate(dataloader):
|
||||
|
||||
# Configure input
|
||||
real_imgs = jt.array(imgs).float32()
|
||||
# -----------------
|
||||
# Train Generator
|
||||
# -----------------
|
||||
# Sample noise as generator input
|
||||
z = jt.array((np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim))).astype('float32'))
|
||||
|
||||
# Generate a batch of images
|
||||
gen_imgs = generator(z)
|
||||
recon_imgs, img_embeddings = discriminator(gen_imgs)
|
||||
|
||||
# Loss measures generator's ability to fool the discriminator
|
||||
g_loss = pixelwise_loss(recon_imgs, gen_imgs.detach()) + lambda_pt * pullaway_loss(img_embeddings)
|
||||
g_loss.sync()
|
||||
optimizer_G.step(g_loss)
|
||||
# ---------------------
|
||||
# Train Discriminator
|
||||
# ---------------------
|
||||
# Measure discriminator's ability to classify real from generated samples
|
||||
real_recon, _ = discriminator(real_imgs)
|
||||
fake_recon, _ = discriminator(gen_imgs.detach())
|
||||
|
||||
d_loss_real = pixelwise_loss(real_recon, real_imgs)
|
||||
d_loss_fake = pixelwise_loss(fake_recon, gen_imgs.detach())
|
||||
|
||||
d_loss = d_loss_real
|
||||
if (margin - d_loss_fake.data).item() > 0:
|
||||
d_loss += margin - d_loss_fake
|
||||
d_loss.sync()
|
||||
optimizer_D.step(d_loss)
|
||||
# --------------
|
||||
# Log Progress
|
||||
# --------------
|
||||
|
||||
print(
|
||||
"[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]"
|
||||
% (epoch, opt.n_epochs, i, len(dataloader), d_loss.data, g_loss.data)
|
||||
)
|
||||
|
||||
batches_done = epoch * len(dataloader) + i
|
||||
if batches_done % opt.sample_interval == 0:
|
||||
save_image(gen_imgs.data[:25], "images/%d.png" % batches_done, nrow=5)
|
||||
|
|
@ -0,0 +1 @@
|
|||
python3.7 EBGAN.py
|
||||
|
|
@ -0,0 +1,154 @@
|
|||
import jittor as jt
|
||||
from jittor import init,nn
|
||||
import argparse
|
||||
import os
|
||||
import numpy as np
|
||||
import math
|
||||
import sys
|
||||
from jittor.dataset.mnist import MNIST
|
||||
import jittor.transform as transform
|
||||
import cv2
|
||||
os.makedirs('images', exist_ok=True)
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument('--n_epochs', type=int, default=200, help='number of epochs of training')
|
||||
parser.add_argument('--batch_size', type=int, default=64, help='size of the batches')
|
||||
parser.add_argument('--lr', type=float, default=0.0002, help='adam: learning rate')
|
||||
parser.add_argument('--b1', type=float, default=0.5, help='adam: decay of first order momentum of gradient')
|
||||
parser.add_argument('--b2', type=float, default=0.999, help='adam: decay of first order momentum of gradient')
|
||||
parser.add_argument('--n_cpu', type=int, default=8, help='number of cpu threads to use during batch generation')
|
||||
parser.add_argument('--latent_dim', type=int, default=100, help='dimensionality of the latent space')
|
||||
parser.add_argument('--img_size', type=int, default=28, help='size of each image dimension')
|
||||
parser.add_argument('--channels', type=int, default=1, help='number of image channels')
|
||||
parser.add_argument('--n_critic', type=int, default=5, help='number of training steps for discriminator per iter')
|
||||
parser.add_argument('--clip_value', type=float, default=0.01, help='lower and upper clip value for disc. weights')
|
||||
parser.add_argument('--sample_interval', type=int, default=400, help='interval betwen image samples')
|
||||
opt = parser.parse_args()
|
||||
print(opt)
|
||||
img_shape = (opt.channels, opt.img_size, opt.img_size)
|
||||
transforms = transform.Compose([
|
||||
transform.Resize(opt.img_size),
|
||||
transform.Gray(),
|
||||
transform.ImageNormalize(mean=[0.5],std=[0.5]),
|
||||
])
|
||||
dataloader = MNIST(train=True,transform=transforms).set_attrs(batch_size=opt.batch_size,shuffle=True)
|
||||
jt.flags.use_cuda = 1
|
||||
os.makedirs("images",exist_ok=True)
|
||||
def save_image(img, path, nrow=10):
|
||||
N,C,W,H = img.shape
|
||||
img2=img.reshape([-1,W*nrow*nrow,H])
|
||||
img=img2[:,:W*nrow,:]
|
||||
for i in range(1,nrow):
|
||||
img=np.concatenate([img,img2[:,W*nrow*i:W*nrow*(i+1),:]],axis=2)
|
||||
min_=img.min()
|
||||
max_=img.max()
|
||||
img=(img-min_)/(max_-min_)*255
|
||||
img=img.transpose((1,2,0))
|
||||
cv2.imwrite(path,img)
|
||||
|
||||
class BatchNorm1d(nn.Module):
|
||||
def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=None, is_train=True, sync=True):
|
||||
assert affine == None
|
||||
|
||||
self.sync = sync
|
||||
self.num_features = num_features
|
||||
self.is_train = is_train
|
||||
self.eps = eps
|
||||
self.momentum = momentum
|
||||
self.weight = init.constant((num_features,), "float32", 1.0)
|
||||
self.bias = init.constant((num_features,), "float32", 0.0)
|
||||
self.running_mean = init.constant((num_features,), "float32", 0.0).stop_grad()
|
||||
self.running_var = init.constant((num_features,), "float32", 1.0).stop_grad()
|
||||
|
||||
def execute(self, x):
|
||||
if self.is_train:
|
||||
xmean = jt.mean(x, dims=[0], keepdims=1)
|
||||
x2mean = jt.mean(x*x, dims=[0], keepdims=1)
|
||||
if self.sync and jt.mpi:
|
||||
xmean = xmean.mpi_all_reduce("mean")
|
||||
x2mean = x2mean.mpi_all_reduce("mean")
|
||||
|
||||
xvar = x2mean-xmean*xmean
|
||||
norm_x = (x-xmean)/jt.sqrt(xvar+self.eps)
|
||||
self.running_mean += (xmean.sum([0])-self.running_mean)*self.momentum
|
||||
self.running_var += (xvar.sum([0])-self.running_var)*self.momentum
|
||||
else:
|
||||
running_mean = self.running_mean.broadcast(x, [0])
|
||||
running_var = self.running_var.broadcast(x, [0])
|
||||
norm_x = (x-running_mean)/jt.sqrt(running_var+self.eps)
|
||||
w = self.weight.broadcast(x, [0])
|
||||
b = self.bias.broadcast(x, [0])
|
||||
return norm_x * w + b
|
||||
|
||||
class Generator(nn.Module):
|
||||
|
||||
def __init__(self):
|
||||
super(Generator, self).__init__()
|
||||
|
||||
def block(in_feat, out_feat, normalize=True):
|
||||
layers = [nn.Linear(in_feat, out_feat)]
|
||||
if normalize:
|
||||
layers.append(BatchNorm1d(out_feat, 0.8))
|
||||
layers.append(nn.LeakyReLU(0.2))
|
||||
return layers
|
||||
self.model = nn.Sequential(*block(opt.latent_dim, 128, normalize=False), *block(128, 256), *block(256, 512), *block(512, 1024), nn.Linear(1024, int(np.prod(img_shape))), nn.Tanh())
|
||||
|
||||
def execute(self, z):
|
||||
img = self.model(z)
|
||||
img = img.view((img.shape[0], *img_shape))
|
||||
return img
|
||||
|
||||
class Discriminator(nn.Module):
|
||||
|
||||
def __init__(self):
|
||||
super(Discriminator, self).__init__()
|
||||
self.model = nn.Sequential(nn.Linear(int(np.prod(img_shape)), 512),
|
||||
nn.LeakyReLU(0.2),
|
||||
nn.Linear(512, 256),
|
||||
nn.LeakyReLU(0.2),
|
||||
nn.Linear(256, 1),
|
||||
)
|
||||
|
||||
def execute(self, img):
|
||||
img_flat = img.reshape((img.shape[0], (- 1)))
|
||||
validity = self.model(img_flat)
|
||||
return validity
|
||||
|
||||
lambda_gp = 10
|
||||
generator = Generator()
|
||||
discriminator = Discriminator()
|
||||
|
||||
optimizer_G = nn.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
|
||||
optimizer_D = nn.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
|
||||
|
||||
def compute_gradient_penalty(D, real_samples, fake_samples):
|
||||
'Calculates the gradient penalty loss for WGAN GP'
|
||||
alpha = jt.array(np.random.random((real_samples.shape[0], 1, 1, 1)).astype('float32'))
|
||||
interpolates = ((alpha * real_samples) + ((1 - alpha) * fake_samples))
|
||||
d_interpolates = D(interpolates)
|
||||
gradients = jt.grad(d_interpolates, interpolates)
|
||||
gradients = gradients.reshape((gradients.shape[0], (- 1)))
|
||||
gp =((jt.sqrt((gradients**2).sum(1))-1)**2).mean()
|
||||
return gp
|
||||
|
||||
batches_done = 0
|
||||
for epoch in range(opt.n_epochs):
|
||||
for i, (imgs, _) in enumerate(dataloader):
|
||||
real_imgs = jt.array(imgs).float32()
|
||||
z = jt.array((np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim))).astype('float32'))
|
||||
fake_imgs = generator(z)
|
||||
real_validity = discriminator(real_imgs)
|
||||
fake_validity = discriminator(fake_imgs)
|
||||
gradient_penalty = compute_gradient_penalty(discriminator, real_imgs.data, fake_imgs.data)
|
||||
d_loss = (- real_validity.mean() + fake_validity.mean() + lambda_gp * gradient_penalty)
|
||||
d_loss.sync()
|
||||
optimizer_D.step(d_loss)
|
||||
if ((i % opt.n_critic) == 0):
|
||||
fake_img = generator(z)
|
||||
fake_validityg = discriminator(fake_img)
|
||||
g_loss = -fake_validityg.mean()
|
||||
g_loss.sync()
|
||||
optimizer_G.step(g_loss)
|
||||
print(('[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]' % (epoch, opt.n_epochs, i, len(dataloader), d_loss.data, g_loss.data)))
|
||||
if ((batches_done % opt.sample_interval) == 0):
|
||||
save_image(fake_imgs.data[:25], ('images/%d.png' % batches_done), nrow=5)
|
||||
batches_done += opt.n_critic
|
||||
|
|
@ -0,0 +1 @@
|
|||
python3.7 WGAN.py
|
||||
Loading…
Reference in New Issue