def reg_init(p):
if p.lambda_reg == 0:
return None
# load the gan
gan_dir = '/accounts/projects/vision/chandan/gan/mnist_dcgan'
sys.path.insert(1, gan_dir)
from dcgan import Discriminator
D = Discriminator(
ngpu=1 if torch.cuda.is_available() else 0).to(device)
D.load_state_dict(
torch.load(oj(gan_dir, 'weights/netD_epoch_99.pth'),
map_location=device))
D = D.eval()
return D
# Create the generator. netG = Generator(params).to(device) # Apply the weights_init() function to randomly initialize all # weights to mean=0.0, stddev=0.2 ############# #netG.apply(weights_init) netG.load_state_dict(checkpoint['generator']) #print(netGt) # Print the model. #print(netG) ############ # Create the discriminator. netD = Discriminator(params).to(device) # Apply the weights_init() function to randomly initialize all # weights to mean=0.0, stddev=0.2 netD.load_state_dict(checkpoint['discriminator']) #netD.apply(weights_init) # Print the model. #print(netD) # Binary Cross Entropy loss function. criterion = nn.BCELoss() fixed_noise = torch.randn(128, params['nz'], 1, 1, device=device) real_label = 1 fake_label = 0 # Optimizer for the discriminator. optimizerD = optim.Adam(netD.parameters(), lr=params['lr'], betas=(params['beta1'], 0.999)) # Optimizer for the generator.
class Discriminator_trimmed(nn.Module):
def __init__(self,list_children):
super(Discriminator_trimmed, self).__init__()
self.main = nn.Sequential(*list_children[0:-2])
def forward(self, input):
x = self.main(input)
return x.view(-1,512*4*4)
if __name__ == "__main__":
model_save_path = r"./saved_data/ImageNet/discriminator_ImageNet"
disc_model = Discriminator()
disc_model.load_state_dict(torch.load(model_save_path))
list_children = []
for i in disc_model.children():
list_children.extend([*i])
# print(list_children)
disc_new = Discriminator_trimmed(list_children)
food_101_path = r"/home/data/food-101/images"
index_arr = np.arange(1000)
np.random.shuffle(index_arr)
# print(index_arr)
training_idx = index_arr[0:75]
validation_idx = index_arr[75:100]
for i in range(1,101):
def train(z_channels,
c_channels,
epoch_num,
batch_size,
lr=0.0002,
beta1=0.5,
model_path='models/dcgan_checkpoint.pth'):
use_cuda = torch.cuda.is_available()
device = torch.device('cuda' if use_cuda else 'cpu')
if use_cuda:
cudnn.benchmark = True
else:
print("***** Warning: Cuda isn't available! *****")
loader = load_mnist(batch_size)
generator = Generator(z_channels, c_channels).to(device)
discriminator = Discriminator(c_channels).to(device)
g_optimizer = optim.Adam(generator.parameters(),
lr=lr,
betas=(beta1, 0.999))
d_optimizer = optim.Adam(discriminator.parameters(),
lr=lr,
betas=(beta1, 0.999))
start_epoch = 0
if os.path.exists(model_path):
checkpoint = torch.load(model_path)
generator.load_state_dict(checkpoint['g'])
discriminator.load_state_dict(checkpoint['d'])
g_optimizer.load_state_dict(checkpoint['g_optim'])
d_optimizer.load_state_dict(checkpoint['d_optim'])
start_epoch = checkpoint['epoch'] + 1
criterion = nn.BCELoss().to(device)
generator.train()
discriminator.train()
std = 0.1
for epoch in range(start_epoch, start_epoch + epoch_num):
d_loss_sum, g_loss_sum = 0, 0
print('---- epoch: %d ----' % (epoch, ))
for i, (real_image, number) in enumerate(loader):
real_image = real_image.to(device)
image_noise = torch.randn(real_image.size(),
device=device).normal_(0, std)
d_optimizer.zero_grad()
real_label = torch.randn(number.size(),
device=device).normal_(0.9, 0.1)
real_image.add_(image_noise)
out = discriminator(real_image)
d_real_loss = criterion(out, real_label)
d_real_loss.backward()
noise_z = torch.randn((number.size(0), z_channels, 1, 1),
device=device)
fake_image = generator(noise_z)
fake_label = torch.zeros(number.size(), device=device)
fake_image = fake_image.add(image_noise)
out = discriminator(fake_image.detach())
d_fake_loss = criterion(out, fake_label)
d_fake_loss.backward()
d_optimizer.step()
g_optimizer.zero_grad()
out = discriminator(fake_image)
g_loss = criterion(out, real_label)
g_loss.backward()
g_optimizer.step()
d_loss_sum += d_real_loss.item() + d_fake_loss.item()
g_loss_sum += g_loss.item()
# if i % 10 == 0:
# print(d_loss, g_loss)
print('d_loss: %f \t\t g_loss: %f' % (d_loss_sum /
(i + 1), g_loss_sum / (i + 1)))
std *= 0.9
if epoch % 1 == 0:
checkpoint = {
'g': generator.state_dict(),
'd': discriminator.state_dict(),
'g_optim': g_optimizer.state_dict(),
'd_optim': d_optimizer.state_dict(),
'epoch': epoch,
}
save_image(fake_image,
'out/fake_samples_epoch_%03d.png' % (epoch, ),
normalize=False)
torch.save(checkpoint, model_path)
os.system('cp ' + model_path + ' models/model%d' % (epoch, ))
print('saved!')
class Solver(object):
def __init__(self, train_loader, test_loader, config):
# 训练集DataLoader
self.train_loader = train_loader
# 测试集DataLoader
self.test_loader = test_loader
# config配置
self.config = config
# 展示信息epoch次数
self.show_every = config.show_every
# 学习率衰退epoch数
self.lr_decay_epoch = [
15,
]
# 创建模型
self.build_model()
# Loss function
self.adversarial_loss = torch.nn.BCELoss()
# 进入test模式
if config.mode == 'test':
print('Loading pre-trained model from %s...' % self.config.model)
# 载入预训练模型并放入相应位置
if self.config.cuda:
self.netG.load_state_dict(torch.load(self.config.model))
self.netD.load_state_dict(torch.load(self.config.model))
else:
self.netG.load_state_dict(
torch.load(self.config.model, map_location='cpu'))
self.netD.load_state_dict(
torch.load(self.config.model, map_location='cpu'))
# 打印网络信息和参数数量
def print_network(self, model, name):
num_params = 0
for p in model.parameters():
num_params += p.numel()
print(name)
print(model)
print("The number of parameters: {}".format(num_params))
# 建立模型
def build_model(self):
self.netG = Generator(nz=self.config.nz,
ngf=self.config.ngf,
nc=self.config.nc)
self.netD = Discriminator(nz=self.config.nz,
ndf=self.config.ndf,
nc=self.config.nc)
# 是否将网络搬运至cuda
if self.config.cuda:
self.netG = self.net.cuda()
self.netD = self.net.cuda()
cudnn.benchmark = True
# self.net.train()
# 设置eval状态
self.netG.eval() # use_global_stats = True
self.netD.eval()
# 载入预训练模型或自行训练模型
if self.config.load == '':
self.netG.load_state_dict(torch.load(self.config.pretrained_model))
self.netD.load_state_dict(torch.load(self.config.pretrained_model))
else:
self.netG.load_state_dict(torch.load(self.config.load))
self.netD.load_state_dict(torch.load(self.config.load))
# 设置优化器
self.optimizerD = Adam(self.netD.parameters(),
lr=self.config.lr,
betas=(self.config.beta1, self.config.beta2),
weight_decay=self.config.wd)
self.optimizerG = Adam(self.netG.parameters(),
lr=self.config.lr,
betas=(self.config.beta1, self.config.beta2),
weight_decay=self.config.wd)
# 打印网络结构
self.print_network(self.netG, 'Generator Structure')
self.print_network(self.netD, 'Discriminator Structure')
# testing状态
def test(self):
# 训练模式
mode_name = 'enhanced'
# 开始时间
time_s = time.time()
# images数量
img_num = len(self.test_loader)
for i, data_batch in enumerate(self.test_loader):
# 获取image数据和name
phone_image, _, name = data_batch['phone_image'], data_batch[
'dslr_image'], data_batch['name']
# testing状态
with torch.no_grad():
# 获取tensor数据并搬运指定设备
images = torch.Tensor(phone_image)
if self.config.cuda:
images = images.cuda()
# 预测值
preds = self.netG(images).cpu().data.numpy()
# 创建image
cv2.imwrite(
os.path.join(self.config.test_fold,
name[:-4] + '_' + mode_name + '.png'), preds)
# 结束时间
time_e = time.time()
print('Speed: %f FPS' % (img_num / (time_e - time_s)))
print('Test Done!')
# training状态
def train(self):
for epoch in range(self.config.epochs):
for i, data_batch in enumerate(self.train_loader):
# 获取image数据和name
phone_image, _, _ = data_batch['phone_image'], data_batch[
'dslr_image'], data_batch['name']
# Adversarial ground truths
valid = torch.Tensor(phone_image.size(0), 1).fill_(1.0)
fake = torch.Tensor(phone_image.size(0), 1).fill_(0.0)
# -----------------
# Train Generator
# -----------------
self.optimizerG.zero_grad()
# Sample noise as generator input
z = torch.Tensor(
np.random.normal(0, 1,
(phone_image.shape[0], self.config.nz)))
# Generate a batch of images
gen_imgs = self.generator(z)
# Loss measures generator's ability to fool the discriminator
g_loss = self.adversarial_loss(self.discriminator(gen_imgs),
valid)
g_loss.backward()
self.optimizerG.step()
# ---------------------
# Train Discriminator
# ---------------------
self.optimizerD.zero_grad()
# Measure discriminator's ability to classify real from generated samples
real_loss = self.adversarial_loss(
self.discriminator(phone_image), valid)
fake_loss = self.adversarial_loss(
self.discriminator(gen_imgs.detach()), fake)
d_loss = (real_loss + fake_loss) / 2
d_loss.backward()
self.optimizerD.step()
# 展示此时信息
if i % (self.show_every // self.config.batch_size) == 0:
print(
"[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]"
% (epoch, self.config.epochs, i, len(
self.train_loader), d_loss.item(), g_loss.item()))
print('Learning rate: ' + str(self.config.lr))
# 保存训练模型
if (epoch + 1) % self.config.epoch_save == 0:
torch.save(
self.netG.state_dict(),
'%s/models/generator/epoch_%d.pth' %
(self.config.save_folder, epoch + 1))
torch.save(
self.netD.state_dict(),
'%s/models/discriminator/epoch_%d.pth' %
(self.config.save_folder, epoch + 1))
# 学习率衰退
if epoch in self.lr_decay_epoch:
self.lr = self.lr * 0.1
# 设置优化器
self.optimizerG = Adam(
filter(lambda p: p.requires_grad,
self.netG.parameters(),
lr=self.config.lr,
betas=(self.config.beta1, self.config.beta2),
weight_decay=self.config.wd))
self.optimizerD = Adam(
filter(lambda p: p.requires_grad,
self.netD.parameters(),
lr=self.config.lr,
betas=(self.config.beta1, self.config.beta2),
weight_decay=self.config.wd))
# 保存训练模型
torch.save(self.net.state_dict(),
'%s/models/generator/final.pth' % self.config.save_folder)
torch.save(
self.net.state_dict(),
'%s/models/discriminator/final.pth' % self.config.save_folder)
shuffle=True,
num_workers=workers)
# Decide which device we want to run on
device = torch.device("cuda:0" if (
torch.cuda.is_available() and ngpu > 0) else "cpu")
print(device)
# custom weights initialization called on netG and netD
# Create the generator
netG = Generator(ngpu, nz + t_in, ngf, nc).to(device)
netD = Discriminator(ngpu, nc, ndf, t_in).to(device)
net_g_path = '/content/drive/MyDrive/Deep_Learning/projects/dcgan/continue_train/checkpoints/netG_latest.pth'
net_d_path = '/content/drive/MyDrive/Deep_Learning/projects/dcgan/continue_train/checkpoints/netD_latest.pth'
netG.load_state_dict(torch.load(net_g_path))
netD.load_state_dict(torch.load(net_d_path))
print("Done loading model!")
# Handle multi-gpu if desired
# if (device.type == 'cuda') and (ngpu > 1):
# netG = nn.DataParallel(netG, list(range(ngpu)))
# netD = nn.DataParallel(netD, list(range(ngpu)))
# netG.apply(weights_init)
# netD.apply(weights_init)
netG = netG.cuda()
netD = netD.cuda()
# Initialize BCELoss function
criterion = nn.BCELoss()
fixed_noise = torch.randn(64, nz + t_in, 1, 1, device=device)
real_label = 1.
import matplotlib.pyplot as plt
import pylab
import numpy as np
num_gpu = 1 if torch.cuda.is_available() else 0
# load the models
from dcgan import Discriminator, Generator
D = Discriminator(ngpu=1).eval()
G = Generator(ngpu=1).eval()
D = D.double()
# load weights
D.load_state_dict(torch.load('weights/netD_epoch_199.pth', map_location='cpu'))
G.load_state_dict(torch.load('weights/netG_epoch_199.pth', map_location='cpu'))
if torch.cuda.is_available():
D = D.cuda()
G = G.cuda()
batch_size = 25
latent_size = 100
fixed_noise = torch.randn(batch_size, latent_size, 1, 1)
if torch.cuda.is_available():
fixed_noise = fixed_noise.cuda()
fake_images = G(fixed_noise)
# z = torch.randn(batch_size, latent_size).cuda()
# z = Variable(z)
import sys
sys.path
sys.path.append(
'/home/luca/GitRepositories/gan-pretrained-pytorch/cifar10_dcgan')
num_gpu = 1 if torch.cuda.is_available() else 0
# load the models
from dcgan import Discriminator, Generator
D = Discriminator(ngpu=1).eval()
G = Generator(ngpu=1).eval()
# load weights
D.load_state_dict(
torch.load(
'/home/luca/GitRepositories/gan-pretrained-pytorch/cifar10_dcgan/weights/netD_epoch_199.pth',
map_location=torch.device('cpu')))
G.load_state_dict(
torch.load(
'/home/luca/GitRepositories/gan-pretrained-pytorch/cifar10_dcgan/weights/netG_epoch_199.pth',
map_location=torch.device('cpu')))
if torch.cuda.is_available():
D = D.cuda()
G = G.cuda()
batch_size = 25
latent_size = 100
from brancher.variables import ProbabilisticModel
from brancher.standard_variables import NormalVariable, DeterministicVariable
import brancher.functions as BF
vutils.make_grid(sample_batch[0].to(device)[:64],
padding=2,
normalize=True).cpu(), (1, 2, 0)))
plt.show()
# Create the generator network.
netG = Generator(params).to(device)
# Load the trained generator weights.
netG.load_state_dict(state_dict['generator'])
print(netG)
# Create the discriminator network.
netD = Discriminator(params).to(device)
# Load the trained discriminator weights.
netD.load_state_dict(state_dict['discriminator'])
print(netD)
# Binary Cross Entropy loss function.
criterion = nn.BCELoss()
fixed_noise = torch.randn(64, params['nz'], 1, 1, device=device)
real_label = 1
fake_label = 0
# Optimizer for the discriminator.
optimizerD = optim.Adam(netD.parameters(),
lr=params['lr'],
betas=(params['beta1'], 0.999))
optimizerD.load_state_dict(state_dict['optimizerD'])
netG = Generator(ngpu, ngf, nc, imageSize, nz).to(device)
netG.apply(weights_init)
if opt.netG != '':
netG.load_state_dict(torch.load(opt.netG))
print(netG)
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
netD = Discriminator(ngpu, ndf, nc, imageSize).to(device)
netD.apply(weights_init)
if opt.netD != '':
netD.load_state_dict(torch.load(opt.netD))
print(netD)
print("Generator n. params: {}".format(count_parameters(netG)))
print("Discriminator n. params: {}".format(count_parameters(netD)))
criterion = nn.BCELoss()
fixed_noise = torch.randn(opt.batchSize, nz, 1, 1, device=device)
real_label = 1
fake_label = 0
# setup optimizer
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))
def main():
dataSize = 128
batchSize = 8
# imageSize = 32
imageSize = 64
initWithCats = True
# discCheckpointPath = r'E:\projects\visus\PyTorch-GAN\implementations\dcgan\checkpoints\2020_07_10_15_53_34\disc_step4800.pth'
# discCheckpointPath = r'E:\projects\visus\pytorch-examples\dcgan\out\netD_epoch_24.pth'
discCheckpointPath = None
gpu = torch.device('cuda')
imageRootPath = r'E:\data\cat-vs-dog\cat'
catDataset = CatDataset(
imageSubdirPath=imageRootPath,
transform=transforms.Compose([
transforms.Resize((imageSize, imageSize)),
# torchvision.transforms.functional.to_grayscale,
transforms.ToTensor(),
# transforms.Lambda(lambda x: torch.reshape(x, x.shape[1:])),
transforms.Normalize([0.5], [0.5])
]))
sampler = InfiniteSampler(catDataset)
catLoader = DataLoader(catDataset, batch_size=batchSize, sampler=sampler)
# Generate a random distance matrix.
# # Make a matrix with positive values.
# distancesCpu = np.clip(np.random.normal(0.5, 1.0 / 3, (dataSize, dataSize)), 0, 1)
# # Make it symmetrical.
# distancesCpu = np.matmul(distancesCpu, distancesCpu.T)
# Generate random points and compute distances, guaranteeing that the triangle rule isn't broken.
randomPoints = generate_points(dataSize)
distancesCpu = scipy.spatial.distance_matrix(randomPoints,
randomPoints,
p=2)
if initWithCats:
imagePaths = random.choices(glob.glob(os.path.join(imageRootPath,
'*')),
k=dataSize)
catImages = []
for p in imagePaths:
image = skimage.transform.resize(imageio.imread(p),
(imageSize, imageSize),
1).transpose(2, 0, 1)
catImages.append(image)
imagesInitCpu = np.asarray(catImages)
else:
imagesInitCpu = np.clip(
np.random.normal(0.5, 0.5 / 3,
(dataSize, 3, imageSize, imageSize)), 0, 1)
images = torch.tensor(imagesInitCpu,
requires_grad=True,
dtype=torch.float32,
device=gpu)
scale = torch.tensor(1.0,
requires_grad=True,
dtype=torch.float32,
device=gpu)
lossModel = models.PerceptualLoss(model='net-lin', net='vgg',
use_gpu=True).to(gpu)
lossBce = torch.nn.BCELoss()
# discriminator = Discriminator(imageSize, 3)
discriminator = Discriminator(3, 64, 1)
if discCheckpointPath:
discriminator.load_state_dict(torch.load(discCheckpointPath))
else:
discriminator.init_params()
discriminator = discriminator.to(gpu)
optimizerImages = torch.optim.Adam([images, scale],
lr=1e-3,
betas=(0.9, 0.999))
# optimizerDisc = torch.optim.Adam(discriminator.parameters(), lr=2e-4, betas=(0.9, 0.999))
optimizerDisc = torch.optim.Adam(discriminator.parameters(),
lr=0.0002,
betas=(0.5, 0.999))
import matplotlib.pyplot as plt
fig, axes = plt.subplots(nrows=2, ncols=batchSize // 2)
fig2 = plt.figure()
ax2 = fig2.add_subplot(1, 1, 1)
outPath = os.path.join(
'images',
datetime.datetime.today().strftime('%Y_%m_%d_%H_%M_%S'))
os.makedirs(outPath)
catIter = iter(catLoader)
for batchIndex in range(10000):
realImageBatch, _ = next(catIter) # type: Tuple(torch.Tensor, Any)
realImageBatch = realImageBatch.to(gpu)
# realImageBatch = torch.tensor(realImageBatchCpu, device=gpu)
# noinspection PyTypeChecker
randomIndices = np.random.randint(
0, dataSize, batchSize).tolist() # type: List[int]
# randomIndices = list(range(dataSize)) # type: List[int]
distanceBatch = torch.tensor(
distancesCpu[randomIndices, :][:, randomIndices],
dtype=torch.float32,
device=gpu)
imageBatch = images[randomIndices].contiguous()
distPred = lossModel.forward(
imageBatch.repeat(repeats=(batchSize, 1, 1, 1)).contiguous(),
imageBatch.repeat_interleave(repeats=batchSize,
dim=0).contiguous(),
normalize=True)
distPredMat = distPred.reshape((batchSize, batchSize))
lossDist = torch.sum((distanceBatch - distPredMat * scale)**2) # MSE
discPred = discriminator(imageBatch)
lossRealness = lossBce(discPred,
torch.ones(imageBatch.shape[0], 1, device=gpu))
lossImages = lossDist + 100.0 * lossRealness # todo
# lossImages = lossRealness # todo
optimizerImages.zero_grad()
lossImages.backward()
optimizerImages.step()
lossDiscReal = lossBce(
discriminator(realImageBatch),
torch.ones(realImageBatch.shape[0], 1, device=gpu))
lossDiscFake = lossBce(discriminator(imageBatch.detach()),
torch.zeros(imageBatch.shape[0], 1, device=gpu))
lossDisc = (lossDiscFake + lossDiscReal) / 2
# lossDisc = torch.tensor(0)
optimizerDisc.zero_grad()
lossDisc.backward()
optimizerDisc.step()
with torch.no_grad():
# todo We're clamping all the images every batch, can we do clamp only the ones updated?
# images = torch.clamp(images, 0, 1) # For some reason this was making the training worse.
images.data = torch.clamp(images.data, 0, 1)
if batchIndex % 100 == 0:
msg = 'iter {}, loss images {:.3f}, loss dist {:.3f}, loss real {:.3f}, loss disc {:.3f}, scale: {:.3f}'.format(
batchIndex, lossImages.item(), lossDist.item(),
lossRealness.item(), lossDisc.item(), scale.item())
print(msg)
# print(discPred.tolist())
imageBatchCpu = imageBatch.cpu().data.numpy().transpose(0, 2, 3, 1)
for i, ax in enumerate(axes.flatten()):
ax.imshow(imageBatchCpu[i])
fig.suptitle(msg)
imagesAllCpu = images.cpu().data.numpy().transpose(0, 2, 3, 1)
plot_image_scatter(ax2,
randomPoints,
imagesAllCpu,
downscaleRatio=2)
fig.savefig(
os.path.join(outPath, 'batch_{}.png'.format(batchIndex)))
fig2.savefig(
os.path.join(outPath, 'scatter_{}.png'.format(batchIndex)))
def main():
dataSize = 32
batchSize = 8
elpipsBatchSize = 1
# imageSize = 32
imageSize = 64
nz = 100
# discCheckpointPath = r'E:\projects\visus\PyTorch-GAN\implementations\dcgan\checkpoints\2020_07_10_15_53_34\disc_step4800.pth'
discCheckpointPath = r'E:\projects\visus\pytorch-examples\dcgan\out\netD_epoch_24.pth'
genCheckpointPath = r'E:\projects\visus\pytorch-examples\dcgan\out\netG_epoch_24.pth'
gpu = torch.device('cuda')
# For now we normalize the vectors to have norm 1, but don't make sure
# that the data has certain mean/std.
pointDataset = AuthorDataset(
jsonPath=r'E:\out\scripts\metaphor-vis\authors-all.json'
)
# Take top N points.
points = np.asarray([pointDataset[i][0] for i in range(dataSize)])
distPointsCpu = l2_sqr_dist_matrix(torch.tensor(points)).numpy()
latents = torch.tensor(np.random.normal(0.0, 1.0, (dataSize, nz)),
requires_grad=True, dtype=torch.float32, device=gpu)
scale = torch.tensor(2.7, requires_grad=True, dtype=torch.float32, device=gpu) # todo Re-check!
bias = torch.tensor(0.0, requires_grad=True, dtype=torch.float32, device=gpu) # todo Re-check!
lpips = models.PerceptualLoss(model='net-lin', net='vgg', use_gpu=True).to(gpu)
# lossModel = lpips
config = elpips.Config()
config.batch_size = elpipsBatchSize # Ensemble size for ELPIPS.
config.set_scale_levels_by_image_size(imageSize, imageSize)
lossModel = elpips.ElpipsMetric(config, lpips).to(gpu)
discriminator = Discriminator(3, 64, 1)
if discCheckpointPath:
discriminator.load_state_dict(torch.load(discCheckpointPath))
else:
discriminator.init_params()
discriminator = discriminator.to(gpu)
generator = Generator(nz=nz, ngf=64)
if genCheckpointPath:
generator.load_state_dict(torch.load(genCheckpointPath))
else:
generator.init_params()
generator = generator.to(gpu)
# optimizerImages = torch.optim.Adam([images, scale], lr=1e-2, betas=(0.9, 0.999))
optimizerScale = torch.optim.Adam([scale, bias], lr=0.001)
# optimizerGen = torch.optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))
# optimizerDisc = torch.optim.Adam(discriminator.parameters(), lr=2e-4, betas=(0.9, 0.999))
# optimizerDisc = torch.optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999))
optimizerLatents = torch.optim.Adam([latents], lr=5e-3, betas=(0.9, 0.999))
fig, axes = plt.subplots(nrows=2, ncols=batchSize // 2)
fig2 = plt.figure()
ax2 = fig2.add_subplot(1, 1, 1)
outPath = os.path.join('runs', datetime.datetime.today().strftime('%Y_%m_%d_%H_%M_%S'))
os.makedirs(outPath)
summaryWriter = SummaryWriter(outPath)
for batchIndex in range(10000):
# noinspection PyTypeChecker
randomIndices = np.random.randint(0, dataSize, batchSize).tolist() # type: List[int]
# # randomIndices = list(range(dataSize)) # type: List[int]
distTarget = torch.tensor(distPointsCpu[randomIndices, :][:, randomIndices], dtype=torch.float32, device=gpu)
latentsBatch = latents[randomIndices]
imageBatchFake = generator(latentsBatch[:, :, None, None].float())
# todo It's possible to compute this more efficiently, but would require re-implementing lpips.
# For now, compute the full BSxBS matrix row-by-row to avoid memory issues.
lossDistTotal = torch.tensor(0.0, device=gpu)
distanceRows = []
for iRow in range(batchSize):
distPredFlat = lossModel(imageBatchFake[iRow].repeat(repeats=(batchSize, 1, 1, 1)).contiguous(),
imageBatchFake, normalize=True)
distPred = distPredFlat.reshape((1, batchSize))
distanceRows.append(distPred)
lossDist = torch.sum((distTarget[iRow] - (distPred * scale + bias)) ** 2) # MSE
lossDistTotal += lossDist
lossDistTotal /= batchSize * batchSize # Compute the mean.
distPredFull = torch.cat(distanceRows, dim=0)
# print('{} - {} || {} - {}'.format(
# torch.min(distPred).item(),
# torch.max(distPred).item(),
# torch.min(distTarget).item(),
# torch.max(distTarget).item()
# ))
# discPred = discriminator(imageBatchFake)
# lossRealness = bceLoss(discPred, torch.ones(imageBatchFake.shape[0], device=gpu))
# lossGen = lossDist + 1.0 * lossRealness
lossLatents = lossDistTotal
# optimizerGen.zero_grad()
# optimizerScale.zero_grad()
# lossGen.backward()
# optimizerGen.step()
# optimizerScale.step()
optimizerLatents.zero_grad()
# optimizerScale.zero_grad()
lossLatents.backward()
optimizerLatents.step()
# optimizerScale.step()
# with torch.no_grad():
# # todo We're clamping all the images every batch, can we clamp only the ones updated?
# # images = torch.clamp(images, 0, 1) # For some reason this was making the training worse.
# images.data = torch.clamp(images.data, 0, 1)
if batchIndex % 100 == 0:
msg = 'iter {} loss dist {:.3f} scale: {:.3f} bias: {:.3f}'.format(batchIndex, lossDistTotal.item(), scale.item(), bias.item())
print(msg)
summaryWriter.add_scalar('loss-dist', lossDistTotal.item(), global_step=batchIndex)
def gpu_images_to_numpy(images):
imagesNumpy = images.cpu().data.numpy().transpose(0, 2, 3, 1)
imagesNumpy = (imagesNumpy + 1) / 2
return imagesNumpy
# print(discPred.tolist())
imageBatchFakeCpu = gpu_images_to_numpy(imageBatchFake)
# imageBatchRealCpu = gpu_images_to_numpy(imageBatchReal)
for iCol, ax in enumerate(axes.flatten()[:batchSize]):
ax.imshow(imageBatchFakeCpu[iCol])
fig.suptitle(msg)
with torch.no_grad():
images = gpu_images_to_numpy(generator(latents[..., None, None]))
authorVectorsProj = umap.UMAP(n_neighbors=min(5, dataSize), random_state=1337).fit_transform(points)
plot_image_scatter(ax2, authorVectorsProj, images, downscaleRatio=2)
fig.savefig(os.path.join(outPath, f'batch_{batchIndex}.png'))
fig2.savefig(os.path.join(outPath, f'scatter_{batchIndex}.png'))
plt.close(fig)
plt.close(fig2)
with torch.no_grad():
imagesGpu = generator(latents[..., None, None])
imageNumber = imagesGpu.shape[0]
# Compute LPIPS distances, batch to avoid memory issues.
bs = min(imageNumber, 8)
assert imageNumber % bs == 0
distPredEval = np.zeros((imagesGpu.shape[0], imagesGpu.shape[0]))
for iCol in range(imageNumber // bs):
startA, endA = iCol * bs, (iCol + 1) * bs
imagesA = imagesGpu[startA:endA]
for j in range(imageNumber // bs):
startB, endB = j * bs, (j + 1) * bs
imagesB = imagesGpu[startB:endB]
distBatchEval = lossModel(imagesA.repeat(repeats=(bs, 1, 1, 1)).contiguous(),
imagesB.repeat_interleave(repeats=bs, dim=0).contiguous(),
normalize=True).cpu().numpy()
distPredEval[startA:endA, startB:endB] = distBatchEval.reshape((bs, bs))
distPredEval = (distPredEval * scale.item() + bias.item())
# Move to the CPU and append an alpha channel for rendering.
images = gpu_images_to_numpy(imagesGpu)
images = [np.concatenate([im, np.ones(im.shape[:-1] + (1,))], axis=-1) for im in images]
distPoints = distPointsCpu
assert np.abs(distPoints - distPoints.T).max() < 1e-5
distPoints = np.minimum(distPoints, distPoints.T) # Remove rounding errors, guarantee symmetry.
config = DistanceMatrixConfig()
config.dataRange = (0., 4.)
_, pointIndicesSorted = render_distance_matrix(
os.path.join(outPath, f'dist_point_{batchIndex}.png'),
distPoints,
images,
config=config
)
# print(np.abs(distPredFlat - distPredFlat.T).max())
# assert np.abs(distPredFlat - distPredFlat.T).max() < 1e-5
# todo The symmetry doesn't hold for E-LPIPS, since it's stochastic.
distPredEval = np.minimum(distPredEval, distPredEval.T) # Remove rounding errors, guarantee symmetry.
config = DistanceMatrixConfig()
config.dataRange = (0., 4.)
render_distance_matrix(
os.path.join(outPath, f'dist_images_{batchIndex}.png'),
distPredEval,
images,
config=config
)
config = DistanceMatrixConfig()
config.dataRange = (0., 4.)
render_distance_matrix(
os.path.join(outPath, f'dist_images_aligned_{batchIndex}.png'),
distPredEval,
images,
predefinedOrder=pointIndicesSorted,
config=config
)
fig, axes = plt.subplots(ncols=2)
axes[0].matshow(distTarget.cpu().numpy(), vmin=0, vmax=4)
axes[1].matshow(distPredFull.cpu().numpy() * scale.item(), vmin=0, vmax=4)
fig.savefig(os.path.join(outPath, f'batch_dist_{batchIndex}.png'))
plt.close(fig)
surveySize = 30
fig, axes = plt.subplots(nrows=3, ncols=surveySize, figsize=(surveySize, 3))
assert len(images) == dataSize
allIndices = list(range(dataSize))
with open(os.path.join(outPath, f'survey_{batchIndex}.txt'), 'w') as file:
for iCol in range(surveySize):
randomIndices = random.sample(allIndices, k=3)
leftToMid = distPointsCpu[randomIndices[0], randomIndices[1]]
rightToMid = distPointsCpu[randomIndices[2], randomIndices[1]]
correctAnswer = 'left' if leftToMid < rightToMid else 'right'
file.write("{}\t{}\t{}\t{}\t{}\n".format(iCol, correctAnswer, leftToMid, rightToMid,
str(tuple(randomIndices))))
for iRow in (0, 1, 2):
axes[iRow][iCol].imshow(images[randomIndices[iRow]])
fig.savefig(os.path.join(outPath, f'survey_{batchIndex}.png'))
plt.close(fig)
torch.save(generator.state_dict(), os.path.join(outPath, 'gen_{}.pth'.format(batchIndex)))
torch.save(discriminator.state_dict(), os.path.join(outPath, 'gen_{}.pth'.format(batchIndex)))
summaryWriter.close()
def main():
dataSize = 128
batchSize = 8
# imageSize = 32
imageSize = 64
# discCheckpointPath = r'E:\projects\visus\PyTorch-GAN\implementations\dcgan\checkpoints\2020_07_10_15_53_34\disc_step4800.pth'
# discCheckpointPath = r'E:\projects\visus\pytorch-examples\dcgan\out\netD_epoch_24.pth'
discCheckpointPath = None
gpu = torch.device('cuda')
# imageDataset = CatDataset(
# imageSubdirPath=r'E:\data\cat-vs-dog\cat',
# transform=transforms.Compose(
# [
# transforms.Resize((imageSize, imageSize)),
# transforms.ToTensor(),
# transforms.Normalize([0.5], [0.5])
# ]
# )
# )
imageDataset = datasets.CIFAR10(root=r'e:\data\images\cifar10', download=True,
transform=transforms.Compose([
transforms.Resize((imageSize, imageSize)),
transforms.ToTensor(),
# transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
transforms.Normalize([0.5], [0.5]),
]))
# For now we normalize the vectors to have norm 1, but don't make sure
# that the data has certain mean/std.
pointDataset = AuthorDataset(
jsonPath=r'E:\out\scripts\metaphor-vis\authors-all.json'
)
imageLoader = DataLoader(imageDataset, batch_size=batchSize, sampler=InfiniteSampler(imageDataset))
pointLoader = DataLoader(pointDataset, batch_size=batchSize, sampler=InfiniteSampler(pointDataset))
# Generate a random distance matrix.
# # Make a matrix with positive values.
# distancesCpu = np.clip(np.random.normal(0.5, 1.0 / 3, (dataSize, dataSize)), 0, 1)
# # Make it symmetrical.
# distancesCpu = np.matmul(distancesCpu, distancesCpu.T)
# Generate random points and compute distances, guaranteeing that the triangle rule isn't broken.
# randomPoints = generate_points(dataSize)
# distancesCpu = scipy.spatial.distance_matrix(randomPoints, randomPoints, p=2)
# catImagePath = os.path.expandvars(r'${DEV_METAPHOR_DATA_PATH}/cats/cat.247.jpg')
# catImage = skimage.transform.resize(imageio.imread(catImagePath), (64, 64), 1).transpose(2, 0, 1)
# imagesInitCpu = np.clip(np.random.normal(0.5, 0.5 / 3, (dataSize, 3, imageSize, imageSize)), 0, 1)
# imagesInitCpu = np.clip(np.tile(catImage, (dataSize, 1, 1, 1)) + np.random.normal(0., 0.5 / 6, (dataSize, 3, 64, 64)), 0, 1)
# images = torch.tensor(imagesInitCpu, requires_grad=True, dtype=torch.float32, device=gpu)
scale = torch.tensor(4.0, requires_grad=True, dtype=torch.float32, device=gpu)
lossModel = models.PerceptualLoss(model='net-lin', net='vgg', use_gpu=True).to(gpu)
bceLoss = torch.nn.BCELoss()
# discriminator = Discriminator(imageSize, 3)
discriminator = Discriminator(3, 64, 1)
if discCheckpointPath:
discriminator.load_state_dict(torch.load(discCheckpointPath))
else:
discriminator.init_params()
discriminator = discriminator.to(gpu)
generator = Generator(nz=pointDataset[0][0].shape[0], ngf=64)
generator.init_params()
generator = generator.to(gpu)
# todo init properly, if training
# discriminator.apply(weights_init_normal)
# optimizerImages = torch.optim.Adam([images, scale], lr=1e-2, betas=(0.9, 0.999))
optimizerScale = torch.optim.Adam([scale], lr=0.001)
optimizerGen = torch.optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))
# optimizerDisc = torch.optim.Adam(discriminator.parameters(), lr=2e-4, betas=(0.9, 0.999))
optimizerDisc = torch.optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999))
import matplotlib.pyplot as plt
fig, axes = plt.subplots(nrows=2 * 2, ncols=batchSize // 2)
fig2 = plt.figure()
ax2 = fig2.add_subplot(1, 1, 1)
outPath = os.path.join('runs', datetime.datetime.today().strftime('%Y_%m_%d_%H_%M_%S'))
os.makedirs(outPath)
imageIter = iter(imageLoader)
pointIter = iter(pointLoader)
for batchIndex in range(10000):
imageBatchReal, _ = next(imageIter) # type: Tuple(torch.Tensor, Any)
imageBatchReal = imageBatchReal.to(gpu)
# imageBatchReal = torch.tensor(realImageBatchCpu, device=gpu)
# noinspection PyTypeChecker
# randomIndices = np.random.randint(0, dataSize, batchSize).tolist() # type: List[int]
# # randomIndices = list(range(dataSize)) # type: List[int]
# distanceBatch = torch.tensor(distancesCpu[randomIndices, :][:, randomIndices], dtype=torch.float32, device=gpu)
# imageBatchFake = images[randomIndices].contiguous()
vectorBatch, _ = next(pointIter)
vectorBatch = vectorBatch.to(gpu)
distanceBatch = l2_sqr_dist_matrix(vectorBatch) # In-batch vector distances.
imageBatchFake = generator(vectorBatch[:, :, None, None].float())
# todo It's possible to compute this more efficiently, but would require re-implementing lpips.
distImages = lossModel.forward(imageBatchFake.repeat(repeats=(batchSize, 1, 1, 1)).contiguous(),
imageBatchFake.repeat_interleave(repeats=batchSize, dim=0).contiguous(), normalize=True)
distPredMat = distImages.reshape((batchSize, batchSize))
lossDist = torch.sum((distanceBatch - distPredMat * scale) ** 2) # MSE
discPred = discriminator(imageBatchFake)
lossRealness = bceLoss(discPred, torch.ones(imageBatchFake.shape[0], device=gpu))
lossGen = lossDist + 1.0 * lossRealness
optimizerGen.zero_grad()
optimizerScale.zero_grad()
lossGen.backward()
optimizerGen.step()
optimizerScale.step()
lossDiscReal = bceLoss(discriminator(imageBatchReal), torch.ones(imageBatchReal.shape[0], device=gpu))
lossDiscFake = bceLoss(discriminator(imageBatchFake.detach()), torch.zeros(imageBatchFake.shape[0], device=gpu))
lossDisc = (lossDiscFake + lossDiscReal) / 2
# lossDisc = torch.tensor(0)
optimizerDisc.zero_grad()
lossDisc.backward()
optimizerDisc.step()
# with torch.no_grad():
# # todo We're clamping all the images every batch, can we clamp only the ones updated?
# # images = torch.clamp(images, 0, 1) # For some reason this was making the training worse.
# images.data = torch.clamp(images.data, 0, 1)
if batchIndex % 100 == 0:
msg = 'iter {}, loss gen {:.3f}, loss dist {:.3f}, loss real {:.3f}, loss disc {:.3f}, scale: {:.3f}'.format(
batchIndex, lossGen.item(), lossDist.item(), lossRealness.item(), lossDisc.item(), scale.item()
)
print(msg)
def gpu_images_to_numpy(images):
imagesNumpy = images.cpu().data.numpy().transpose(0, 2, 3, 1)
imagesNumpy = (imagesNumpy + 1) / 2
return imagesNumpy
# print(discPred.tolist())
imageBatchFakeCpu = gpu_images_to_numpy(imageBatchFake)
imageBatchRealCpu = gpu_images_to_numpy(imageBatchReal)
for i, ax in enumerate(axes.flatten()[:batchSize]):
ax.imshow(imageBatchFakeCpu[i])
for i, ax in enumerate(axes.flatten()[batchSize:]):
ax.imshow(imageBatchRealCpu[i])
fig.suptitle(msg)
with torch.no_grad():
points = np.asarray([pointDataset[i][0] for i in range(200)], dtype=np.float32)
images = gpu_images_to_numpy(generator(torch.tensor(points[..., None, None], device=gpu)))
authorVectorsProj = umap.UMAP(n_neighbors=5, random_state=1337).fit_transform(points)
plot_image_scatter(ax2, authorVectorsProj, images, downscaleRatio=2)
fig.savefig(os.path.join(outPath, f'batch_{batchIndex}.png'))
fig2.savefig(os.path.join(outPath, f'scatter_{batchIndex}.png'))
plt.close(fig)
plt.close(fig2)
with torch.no_grad():
imageNumber = 48
points = np.asarray([pointDataset[i][0] for i in range(imageNumber)], dtype=np.float32)
imagesGpu = generator(torch.tensor(points[..., None, None], device=gpu))
# Compute LPIPS distances, batch to avoid memory issues.
bs = 8
assert imageNumber % bs == 0
distImages = np.zeros((imagesGpu.shape[0], imagesGpu.shape[0]))
for i in range(imageNumber // bs):
startA, endA = i * bs, (i + 1) * bs
imagesA = imagesGpu[startA:endA]
for j in range(imageNumber // bs):
startB, endB = j * bs, (j + 1) * bs
imagesB = imagesGpu[startB:endB]
distBatch = lossModel.forward(imagesA.repeat(repeats=(bs, 1, 1, 1)).contiguous(),
imagesB.repeat_interleave(repeats=bs, dim=0).contiguous(),
normalize=True).cpu().numpy()
distImages[startA:endA, startB:endB] = distBatch.reshape((bs, bs))
# Move to the CPU and append an alpha channel for rendering.
images = gpu_images_to_numpy(imagesGpu)
images = [np.concatenate([im, np.ones(im.shape[:-1] + (1,))], axis=-1) for im in images]
distPoints = l2_sqr_dist_matrix(torch.tensor(points, dtype=torch.double)).numpy()
assert np.abs(distPoints - distPoints.T).max() < 1e-5
distPoints = np.minimum(distPoints, distPoints.T) # Remove rounding errors, guarantee symmetry.
config = DistanceMatrixConfig()
config.dataRange = (0., 4.)
render_distance_matrix(os.path.join(outPath, f'dist_point_{batchIndex}.png'),
distPoints,
images,
config)
assert np.abs(distImages - distImages.T).max() < 1e-5
distImages = np.minimum(distImages, distImages.T) # Remove rounding errors, guarantee symmetry.
config = DistanceMatrixConfig()
config.dataRange = (0., 1.)
render_distance_matrix(os.path.join(outPath, f'dist_images_{batchIndex}.png'),
distImages,
images,
config)
torch.save(generator.state_dict(), os.path.join(outPath, 'gen_{}.pth'.format(batchIndex)))
torch.save(discriminator.state_dict(), os.path.join(outPath, 'disc_{}.pth'.format(batchIndex)))
def train():
os.makedirs('log', exist_ok=True)
ds = datasets.ImageFolder(root=data_root,
transform=transforms.Compose([
transforms.Resize(image_size),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
dataloader = DataLoader(ds, batch_size=batch_size, shuffle=True)
net_g = Generator(n_latent_vector, n_g_filters).to(device)
net_g.apply(weight_init)
net_d = Discriminator(n_d_filters).to(device)
net_d.apply(weight_init)
if os.path.exists(model_save_path):
all_state_dict = torch.load(model_save_path)
net_d.load_state_dict(all_state_dict['d_state_dict'])
net_g.load_state_dict(all_state_dict['g_state_dict'])
print('model restored from {}'.format(model_save_path))
criterion = nn.BCELoss()
fixed_noise = torch.randn(1, n_latent_vector, 1, 1, device=device)
real_label = 1
fake_label = 0
optimizer_d = optim.Adam(net_d.parameters(), lr=lr, betas=(0.5, 0.999))
optimizer_g = optim.Adam(net_g.parameters(), lr=lr, betas=(0.5, 0.999))
print('start training...')
try:
for epoch in range(epochs):
for i, data in enumerate(dataloader, 0):
# update Discrinimator, maximize d loss
net_d.zero_grad()
real_cpu = data[0].to(device)
b_size = real_cpu.size(0)
label = torch.full((b_size,), real_label, device=device)
output = net_d(real_cpu).view(-1)
err_d_real = criterion(output, label)
err_d_real.backward()
d_x = output.mean().item()
# train with fake batch
noise = torch.randn(b_size, n_latent_vector, 1, 1, device=device)
fake = net_g(noise)
label.fill_(fake_label)
output = net_d(fake.detach()).view(-1)
err_d_fake = criterion(output, label)
err_d_fake.backward()
d_g_z1 = output.mean().item()
err_d = err_d_real + err_d_fake
optimizer_d.step()
# update Generator
net_g.zero_grad()
label.fill_(real_label)
output = net_d(fake).view(-1)
err_g = criterion(output, label)
err_g.backward()
d_g_z2 = output.mean().item()
optimizer_d.step()
if i % 50 == 0:
print(f'Epoch: {epoch}, loss_d: {err_d.item()}, loss_g: {err_g.item()}')
if epoch % 2 == 0 and epoch != 0:
with torch.no_grad():
fake = net_g(fixed_noise).detach().cpu().numpy()
print(fake.shape)
fake = np.transpose(np.squeeze(fake, axis=0), (1, 2, 0))
print(fake.shape)
cv2.imwrite('log/{}_fake.png'.format(epoch), fake)
print('record a fake image to local.')
except KeyboardInterrupt:
print('interrupted, try saving the model')
all_state_dict = {
'd_state_dict': net_d.state_dict(),
'g_state_dict': net_g.state_dict(),
}
torch.save(all_state_dict, model_save_path)
print('model saved...')