def test_simple(args):
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
g_n_blocks = 8
G = Generator(args.input_nc,
args.output_nc,
args.ngf,
args.nz,
n_blocks=g_n_blocks).to(device)
G.load_state_dict(
torch.load(args.modelG_state_path,
map_location=lambda storage, loc: storage))
G.eval()
input_img = Image.open(args.input_img_path).convert('RGB')
input_tensor = get_input_tensor(input_img).unsqueeze(0).to(device)
z_random = sample_z(1, args.nz, 'gauss').to(device)
with torch.no_grad():
now = time.time()
out = G(input_tensor, z_random)
end = time.time()
print('elapsed: {}'.format(end - now))
out_denormalized = denormalize(out.squeeze()).cpu()
out_img = toPIL(out_denormalized)
out_img.show()
def _combined_threshold(image):
color_thresholded, h_thresholded, l_thresholded, s_thresholded = color_threshold(image)
xy_thresholded, x_thresholded, y_thresholded = absolute_threshold_(image)
gd_thresholded, g_thresholded, d_thresholded = gradient_threshold(image)
combined = (color_thresholded == 1) | (xy_thresholded == 1) | (g_thresholded == 1)
combined = util.denormalize(combined)
return combined, color_thresholded, xy_thresholded, g_thresholded
def test_recursive(args):
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
g_n_blocks = 8
G = Generator(args.input_nc,
args.output_nc,
args.ngf,
args.nz,
n_blocks=g_n_blocks).to(device)
G.load_state_dict(
torch.load(args.modelG_state_path,
map_location=lambda storage, loc: storage))
G.eval()
input_img = Image.open(args.input_img_path).convert('RGB')
input_tensor = get_input_tensor(input_img).unsqueeze(0).to(device)
z_random = sample_z(1, args.nz, 'gauss').to(device)
cap = cv2.VideoCapture(0)
while cap.isOpened():
with torch.no_grad():
now = time.time()
out = G(input_tensor, z_random)
end = time.time()
print('elapsed: {}'.format(end - now))
out_denormalized = denormalize(out.squeeze())
out_denormalized = out_denormalized.cpu().numpy().transpose(
1, 2, 0)
out_denormalized = out_denormalized[:, :, ::-1]
cv2.imshow('Result', out_denormalized)
input_tensor = out
z_random = sample_z(1, args.nz, 'gauss').to(device)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cv2.destroyAllWindows()
def save_video(self, video_dir, global_step):
output_dir = os.path.join(video_dir, 'step_{}'.format(global_step))
os.mkdir(output_dir)
input_img = Image.open('imgs/test.png').convert('RGB').resize(
(self.img_size, self.img_size), Image.BICUBIC)
input_tensor = get_input_tensor(input_img).unsqueeze(0).to(self.device)
self.G.eval()
for i in range(450):
with torch.no_grad():
out = self.G(input_tensor)
out_denormalized = denormalize(out.squeeze()).cpu()
out_img = toPIL(out_denormalized)
out_img.save('{0}/{1:04d}.png'.format(output_dir, i))
input_tensor = out
self.G.train()
cmd = 'ffmpeg -r 30 -i {}/%04d.png -vcodec libx264 -pix_fmt yuv420p -r 30 {}/movie.mp4'.format(
output_dir, output_dir)
subprocess.call(cmd.split())
def optimize(self, A, B, global_step):
A = A.to(self.device)
B = B.to(self.device)
# Logging the input images
if global_step % self.log_freq == 0:
log_real_A = torchvision.utils.make_grid(A)
log_real_A = denormalize(log_real_A)
self.writer.add_image('real_A', log_real_A, global_step)
log_real_B = torchvision.utils.make_grid(B)
log_real_B = denormalize(log_real_B)
self.writer.add_image('real_B', log_real_B, global_step)
# Forward pass
fake_B = self.G(A)
if global_step % self.log_freq == 0:
log_fake_B = torchvision.utils.make_grid(fake_B)
log_fake_B = denormalize(log_fake_B)
self.writer.add_image('fake_B', log_fake_B, global_step)
# ==================================================================
# 1. Train D
# ==================================================================
self._set_requires_grad(self.D, True)
# Real
real_pair = torch.cat([A, B], dim=1)
real_D = self.D(real_pair)
loss_real_D = gan_loss(real_D, target=1)
# Fake
fake_pair = torch.cat([A, fake_B], dim=1)
fake_D = self.D(fake_pair.detach())
loss_fake_D = gan_loss(fake_D, target=0)
loss_D = (loss_real_D + loss_fake_D) * 0.5
self._all_zero_grad()
loss_D.backward()
self.optim_D.step()
# Logging
self.writer.add_scalar('loss/loss_D', loss_D.item(), global_step)
# ==================================================================
# 2. Train G
# ==================================================================
self._set_requires_grad(self.D, False)
# Fake
fake_D2 = self.D(fake_pair)
loss_G_GAN = gan_loss(fake_D2, target=1)
loss_G_L1 = l1_loss(fake_B, B)
loss_G = loss_G_GAN + loss_G_L1 * self.lambda_l1
self._all_zero_grad()
loss_G.backward()
self.optim_G.step()
# Logging
self.writer.add_scalar('loss/loss_G_GAN', loss_G_GAN.item(),
global_step)
self.writer.add_scalar('loss/loss_G_L1', loss_G_L1.item(), global_step)
self.writer.add_scalar('loss/loss_G', loss_G.item(), global_step)
def optimize(self, A, B, global_step):
if A.size(0) <= 1:
return
A = A.to(self.device)
B = B.to(self.device)
cVAE_data = {'A': A[0:self.half_size], 'B': B[0:self.half_size]}
cLR_data = {'A': A[self.half_size:], 'B': B[self.half_size:]}
# Logging the input images
log_imgs = torch.cat([cVAE_data['A'], cVAE_data['B']], 0)
log_imgs = torchvision.utils.make_grid(log_imgs)
log_imgs = denormalize(log_imgs)
self.writer.add_image('cVAE_input', log_imgs, global_step)
log_imgs = torch.cat([cLR_data['A'], cLR_data['B']], 0)
log_imgs = torchvision.utils.make_grid(log_imgs)
log_imgs = denormalize(log_imgs)
self.writer.add_image('cLR_input', log_imgs, global_step)
# ----------------------------------------------------------------
# 1. Train D
# ----------------------------------------------------------------
# -----------------------------
# Optimize D in cVAE-GAN
# -----------------------------
# Generate encoded latent vector
mu, logvar = self.E(cVAE_data['B'])
std = torch.exp(logvar / 2)
random_z = sample_z(self.half_size, self.nz, 'gauss').to(self.device)
encoded_z = (random_z * std) + mu
# Generate fake image
fake_img_cVAE = self.G(cVAE_data['A'], encoded_z)
log_imgs = torchvision.utils.make_grid(fake_img_cVAE)
log_imgs = denormalize(log_imgs)
self.writer.add_image('cVAE_fake_encoded', log_imgs, global_step)
real_pair_cVAE = torch.cat([cVAE_data['A'], cVAE_data['B']], dim=1)
fake_pair_cVAE = torch.cat([cVAE_data['A'], fake_img_cVAE], dim=1)
real_D_cVAE_1, real_D_cVAE_2 = self.D_cVAE(real_pair_cVAE)
fake_D_cVAE_1, fake_D_cVAE_2 = self.D_cVAE(fake_pair_cVAE.detach())
# The loss for small patch & big patch
loss_D_cVAE_1 = mse_loss(real_D_cVAE_1, target=1) + mse_loss(
fake_D_cVAE_1, target=0)
loss_D_cVAE_2 = mse_loss(real_D_cVAE_2, target=1) + mse_loss(
fake_D_cVAE_2, target=0)
self.writer.add_scalar('loss/loss_D_cVAE_1', loss_D_cVAE_1.item(),
global_step)
self.writer.add_scalar('loss/loss_D_cVAE_2', loss_D_cVAE_2.item(),
global_step)
# -----------------------------
# Optimize D in cLR-GAN
# -----------------------------
# Generate fake image
fake_img_cLR = self.G(cLR_data['A'], random_z)
log_imgs = torchvision.utils.make_grid(fake_img_cLR)
log_imgs = denormalize(log_imgs)
self.writer.add_image('cLR_fake_random', log_imgs, global_step)
real_pair_cLR = torch.cat([cLR_data['A'], cLR_data['B']], dim=1)
fake_pair_cLR = torch.cat([cVAE_data['A'], fake_img_cLR], dim=1)
real_D_cLR_1, real_D_cLR_2 = self.D_cLR(real_pair_cLR)
fake_D_cLR_1, fake_D_cLR_2 = self.D_cLR(fake_pair_cLR.detach())
# Loss for small patch & big patch
loss_D_cLR_1 = mse_loss(real_D_cLR_1, target=1) + mse_loss(
fake_D_cLR_1, target=0)
loss_D_cLR_2 = mse_loss(real_D_cLR_2, target=1) + mse_loss(
fake_D_cLR_2, target=0)
self.writer.add_scalar('loss/loss_D_cVAE_1', loss_D_cVAE_1.item(),
global_step)
self.writer.add_scalar('loss/loss_D_cVAE_2', loss_D_cVAE_2.item(),
global_step)
loss_D = loss_D_cVAE_1 + loss_D_cVAE_2 + loss_D_cLR_1 + loss_D_cLR_2
self.writer.add_scalar('loss/loss_D', loss_D.item(), global_step)
# -----------------------------
# Update D
# -----------------------------
# set_requires_grad([], False)
self.all_zero_grad()
loss_D.backward()
self.optim_D_cVAE.step()
self.optim_D_cLR.step()
# ----------------------------------------------------------------
# 2. Train G & E
# ----------------------------------------------------------------
# -----------------------------
# GAN loss
# -----------------------------
# Generate encoded latent vector
mu, logvar = self.E(cVAE_data['B'])
std = torch.exp(logvar / 2)
random_z = sample_z(self.half_size, self.nz, 'gauss').to(self.device)
encoded_z = (random_z * std) + mu
# Generate fake image
fake_img_cVAE = self.G(cVAE_data['A'], encoded_z)
# self.writer.add_images('cVAE_output', fake_img_cVAE.add(1.0).mul(0.5), global_step)
fake_pair_cVAE = torch.cat([cVAE_data['A'], fake_img_cVAE], dim=1)
# Fool D_cVAE
fake_D_cVAE_1, fake_D_cVAE_2 = self.D_cVAE(fake_pair_cVAE)
# Loss for small patch & big patch
loss_G_cVAE_1 = mse_loss(fake_D_cVAE_1, target=1)
loss_G_cVAE_2 = mse_loss(fake_D_cVAE_2, target=1)
# Random latent vector and generate fake image
random_z = sample_z(self.half_size, self.nz, 'gauss').to(self.device)
fake_img_cLR = self.G(cLR_data['A'], random_z)
fake_pair_cLR = torch.cat([cLR_data['A'], fake_img_cLR], dim=1)
# Fool D_cLR
fake_D_cLR_1, fake_D_cLR_2 = self.D_cLR(fake_pair_cLR)
# Loss for small patch & big patch
loss_G_cLR_1 = mse_loss(fake_D_cLR_1, target=1)
loss_G_cLR_2 = mse_loss(fake_D_cLR_2, target=1)
loss_G = loss_G_cVAE_1 + loss_G_cVAE_2 + loss_G_cLR_1 + loss_G_cLR_2
self.writer.add_scalar('loss/loss_G', loss_G.item(), global_step)
# -----------------------------
# KL-divergence (cVAE-GAN)
# -----------------------------
kl_div = torch.sum(
0.5 * (mu**2 + torch.exp(logvar) - logvar - 1)) * self.lambda_kl
self.writer.add_scalar('loss/kl_div', kl_div.item(), global_step)
# -----------------------------
# Reconstruction of image B (|G(A, z) - B|) (cVAE-GAN)
# -----------------------------
loss_img_recon = l1_loss(fake_img_cVAE,
cVAE_data['B']) * self.lambda_img
self.writer.add_scalar('loss/loss_img_recon', loss_img_recon.item(),
global_step)
loss_E_G = loss_G + kl_div + loss_img_recon
self.writer.add_scalar('loss/loss_E_G', loss_E_G.item(), global_step)
# -----------------------------
# Update E & G
# -----------------------------
self.all_zero_grad()
loss_E_G.backward(retain_graph=True)
self.optim_E.step()
self.optim_G.step()
# ----------------------------------------------------------------
# 3. Train only G
# ----------------------------------------------------------------
# -----------------------------
# Reconstruction of random latent code (|E(G(A, z)) - z|) (cLR-GAN)
# -----------------------------
# This step should update only G.
# See https://github.com/junyanz/BicycleGAN/issues/5 for details.
mu, logvar = self.E(fake_img_cLR)
loss_z_recon = l1_loss(mu, random_z) * self.lambda_z
self.writer.add_scalar('loss/loss_z_recon', loss_z_recon.item(),
global_step)
# -----------------------------
# Update G
# -----------------------------
self.all_zero_grad()
loss_z_recon.backward()
self.optim_G.step()
def main():
args = parse_arg()
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
writer = SummaryWriter(os.path.join(args.save_dir, 'tb'))
if args.use_cuda:
device = torch.device('cuda')
else:
device = torch.device('cpu')
model = net.InPainting(args.use_cuda, mask_clipping=False).to(device)
model_dis = net.Discriminator().to(device)
model.train()
model_dis.train()
train_params = [
p for (n, p) in model.named_parameters() if 'fine_painter' not in n
]
optimizer = torch.optim.Adam(train_params, lr=args.lr)
optimizer_dis = torch.optim.Adam(model_dis.parameters(), lr=args.lr)
if args.resume:
print('Resume training from ', args.resume)
ckpt = torch.load(args.resume)
try:
model.load_state_dict(ckpt['ckpt'])
model.load_state_dict(ckpt['ckpt_dis'])
optimizer.load_state_dict(ckpt['optim'])
optimizer_dis.load_state_dict(ckpt['optim_dis'])
except Exception as e:
print(traceback.format_exc())
print('Missing keys')
model.load_state_dict(
{k: v
for k, v in ckpt['ckpt'].items() if 'encoder' in k},
strict=False)
if args.fix_mask:
print('fix mask prediction')
for p in model.encoder.parameters():
p.requires_grad = False
elif args.fix_recon:
print('fix painter')
for p in model.painter.parameters():
p.requires_grad = False
loss = torch.nn.L1Loss()
loss_bce = torch.nn.BCELoss()
ssim_window = ssim.create_window(11, 3).to(device)
#data = dataset.Dataset(args.path)
data = dataset.LargeScaleWatermarkDataset(
folder_origin=os.path.join(args.path, args.path_origin),
folder_watermarked=os.path.join(args.path, args.path_wm),
anno_file=os.path.join(args.path, args.path_anno))
train_loader = torch.utils.data.DataLoader(dataset=data,
batch_size=args.batchsize,
shuffle=True,
num_workers=4)
try:
batch_per_epoch = len(train_loader)
best_loss = 100
for i in range(args.epochs):
epoch_loss = 0
print('Epoch %d' % i)
for j, item in enumerate(train_loader):
img_raw, img_wm, mask_wm = item
img_raw, img_wm, mask_wm = img_raw.to(device), img_wm.to(
device), mask_wm.to(device)
mask, recon = model(img_wm)
# 加入discriminator
if args.gan_method:
# optimize D
dis_real, dis_recon = dis_forward(model_dis, img_raw,
recon.detach())
dis_wm = torch.sigmoid(model_dis(img_wm))
assert dis_recon.size() == dis_wm.size()
dis_fake = 0.5 * (dis_recon + dis_wm)
loss_disc = torch.mean(-1 * torch.log(1 - dis_fake) -
torch.log(dis_real))
loss_gp = net.calc_gradient_penalty(
model_dis, img_raw, recon.detach())
loss_d = loss_gp + loss_disc
# optimize G through D
dis_real, dis_recon = dis_forward(model_dis, img_raw,
recon)
# print('dis_real:', dis_real.size(), 'dis_recon:', dis_recon.size())
loss_g = 0.001 * torch.mean(-1 * torch.log(dis_recon))
loss_mask_reg = 0.1 * mask.clamp(0, 1).mean()
# loss_mask = 1000*util.exclusion_loss(mask)
try:
loss_mask = loss_bce(mask.clamp(0., 1.),
mask_wm.float().clamp(0., 1.))
except Exception:
import pdb
pdb.set_trace()
if not (mask >= 0. & mask <= 1.).all():
print('错误出在生成的mask')
if not (mask_wm >= 0. & mask_wm <= 1.).all():
print('错误出在gt水印')
loss_recon = loss(recon, img_raw)
loss_ssim = 1 - ssim._ssim(0.5 * (1 + img_raw), 0.5 *
(1 + recon), ssim_window, 11, 3,
True)
loss_weighted_recon = util.weighted_l1(recon, img_raw, mask)
loss_ = loss_recon + loss_mask + loss_ssim
if args.gan_method:
loss_ += loss_g
optimizer_dis.zero_grad()
loss_d.backward()
optimizer_dis.step()
optimizer.zero_grad()
loss_.backward()
optimizer.step()
epoch_loss += loss_.item()
step = i * batch_per_epoch + j
if j % 5 == 0:
writer.add_scalars(
'loss', {
'recon_l1': loss_recon.item(),
'ssim': loss_ssim.item(),
'exclusion': loss_mask.item(),
'mask_reg': loss_mask_reg.item()
}, step)
if j % 10 == 0:
print(
'Loss: %.3f (recon: %.3f \t ssim: %.3f \t mask: %.3f \t)'
% (loss_.item(), loss_recon.item(), loss_ssim.item(),
loss_mask.item()))
if args.gan_method:
print(
'disc: %.3f \t gp: %.3f \t gen: %.3f' %
(loss_disc.item(), loss_gp.item(), loss_g.item()))
# 记录mask和原图
if j % 50 == 0:
#import pdb; pdb.set_trace()
writer.add_images('images', [
torch.cat(3 * [mask[0].float().to(device)]),
torch.cat(
3 * [mask_wm[0].float().to(device).unsqueeze(0)]),
util.denormalize(img_wm[0]),
util.denormalize(recon[0]).clamp(0, 1)
],
global_step=step,
dataformats='CHW')
'''
writer.add_image('mask', mask[0], step)
writer.add_image('img', util.denormalize(img_wm[0]), step)
writer.add_image('recon_c', util.denormalize(recon_coarse[0]).clamp(0, 1), step)
writer.add_image('recon_f', util.denormalize(recon_fine[0]).clamp(0, 1), step)
writer.add_image('mask_gt', mask_wm[0], step)
'''
# 画各层的梯度分布图
if j % 100 == 0:
writer.add_figure('grad_flow',
util.plot_grad_flow_v2(
model.named_parameters()),
global_step=step)
if args.gan_method:
writer.add_figure('discriminator_grad',
util.plot_grad_flow_v2(
model_dis.named_parameters()),
global_step=step)
ckpt = {
'ckpt': model.state_dict(),
'optim': optimizer.state_dict()
}
torch.save(ckpt, os.path.join(args.save_dir, 'latest.pth'))
# 记录所有最好的epoch weight
if epoch_loss / (j + 1) < best_loss:
best_loss = epoch_loss / (j + 1)
shutil.copy(
os.path.join(args.save_dir, 'latest.pth'),
os.path.join(args.save_dir, 'epoch_' + str(i) + '.pth'))
except Exception as e:
ckpt = {'ckpt': model.state_dict(), 'optim': optimizer.state_dict()}
torch.save(ckpt, os.path.join(args.save_dir, 'latest.pth'))
print('Save temporary checkpoints to %s' % args.save_dir)
print(str(e), traceback.print_exc())
sys.exit(0)
print('Done training.')
shutil.copyfile(os.path.join(args.save_dir, 'latest.pth'),
os.path.join(args.save_dir, 'epoch_%d.pth' % (i + 1)))
def PredictMotion(self):
print('Motion: ')
P_m = ConditionalMotionNet()
param = torch.load(self.model_path + '/PMNet_weight_' +
self.model_epoch + '.pth')
P_m.load_state_dict(param)
if self.gpu > -1:
P_m.cuda(self.gpu)
with open(self.model_path + '/codebook_m_' + self.model_epoch + '.pkl',
'rb') as f:
codebook_m = pickle.load(
f) if sys.version_info[0] == 2 else pickle.load(
f, encoding='latin1')
id1 = int(np.floor((len(codebook_m) - 1) * self.t_m))
id2 = int(np.ceil((len(codebook_m) - 1) * self.t_m))
z_weight = (len(codebook_m) - 1) * self.t_m - np.floor(
(len(codebook_m) - 1) * self.t_m)
z_m = (1. - z_weight
) * codebook_m[id1:id1 + 1] + z_weight * codebook_m[id2:id2 + 1]
z_m = Variable(torch.from_numpy(z_m.astype(np.float32)))
if self.gpu > -1:
z_m = z_m.cuda(self.gpu)
initial_coordinate = np.array([
np.meshgrid(np.linspace(-1, 1, self.w + 2 * self.pad),
np.linspace(-1, 1, self.h + 2 * self.pad),
sparse=False)
]).astype(np.float32)
initial_coordinate = Variable(torch.from_numpy(initial_coordinate))
if self.gpu > -1:
initial_coordinate = initial_coordinate.cuda(self.gpu)
with torch.no_grad():
test_img = cv2.imread(self.input_path)
test_img = cv2.resize(test_img, (self.w, self.h))
test_input = np.array([normalize(test_img)])
test_input = Variable(
torch.from_numpy(test_input.transpose(0, 3, 1, 2)))
if self.gpu > -1:
test_input = test_input.cuda(self.gpu)
padded_test_input = F.pad(test_input,
(self.pad, self.pad, self.pad, self.pad),
mode='reflect')
test_img_large = cv2.imread(self.input_path)
if self.fw == None or self.fh == None:
self.fh, self.fw = test_img_large.shape[:2]
test_img_large = cv2.resize(test_img_large, (self.fw, self.fh))
padded_test_input_large = np.array([normalize(test_img_large)])
padded_test_input_large = Variable(
torch.from_numpy(padded_test_input_large.transpose(0, 3, 1,
2)))
if self.gpu > -1:
padded_test_input_large = padded_test_input_large.cuda(
self.gpu)
scaled_pads = (int(self.pad * self.fh / float(self.h)),
int(self.pad * self.fw / float(self.w)))
padded_test_input_large = F.pad(padded_test_input_large,
(scaled_pads[1], scaled_pads[1],
scaled_pads[0], scaled_pads[0]),
mode='reflect')
V_m = list()
V_f = list()
old_correpondence = None
for t in range(self.TM):
sys.stdout.write("\rProcessing frame %d, " % (t + 1))
sys.stdout.flush()
flow = P_m(test_input, z_m)
flow[:, 0, :, :] = flow[:, 0, :, :] * (
self.w / float(self.pad * 2 + self.w))
flow[:, 1, :, :] = flow[:, 1, :, :] * (
self.h / float(self.pad * 2 + self.h))
flow = F.pad(flow, (self.pad, self.pad, self.pad, self.pad),
mode='reflect')
flow = self.s_m * flow
correspondence = initial_coordinate + flow
if old_correpondence is not None:
correspondence = F.grid_sample(old_correpondence,
correspondence.permute(
0, 2, 3, 1),
padding_mode='border')
correspondence_large = F.upsample(
correspondence,
size=(self.fh + scaled_pads[0] * 2,
self.fw + scaled_pads[1] * 2),
mode='bilinear',
align_corners=True)
y_large = F.grid_sample(padded_test_input_large,
correspondence_large.permute(
0, 2, 3, 1),
padding_mode='border')
outimg = y_large.data.cpu().numpy()[0].transpose(1, 2, 0)
outimg = denormalize(outimg)
outimg = outimg[scaled_pads[0]:outimg.shape[0] -
scaled_pads[0],
scaled_pads[1]:outimg.shape[1] -
scaled_pads[1]]
V_m.append(outimg)
outflowimg = flow.data.cpu().numpy()[0].transpose(1, 2, 0)
outflowimg = outflowimg[self.pad:outflowimg.shape[0] -
self.pad,
self.pad:outflowimg.shape[1] -
self.pad]
mag, ang = cv2.cartToPolar(outflowimg[..., 1], outflowimg[...,
0])
hsv = np.zeros_like(test_img)
hsv[..., 1] = cv2.normalize(mag, None, 0, 255, cv2.NORM_MINMAX)
hsv[..., 0] = ang * 180 / np.pi / 2
hsv[..., 2] = 255
outflowimg = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
outflowimg = cv2.resize(outflowimg, (self.fw, self.fh))
V_f.append(outflowimg)
y = F.grid_sample(padded_test_input,
correspondence.permute(0, 2, 3, 1),
padding_mode='border')
test_input = y[:, :, self.pad:y.shape[2] - self.pad,
self.pad:y.shape[3] - self.pad]
old_correpondence = correspondence
V_mloop = generateLoop(V_m)
return V_mloop, V_f
def PredictAppearance(self, V_mloop):
print('\nAppearance: ', )
minimum_loop_num = int(1 / self.s_a)
P_a = ConditionalAppearanceNet(8)
param = torch.load(self.model_path + '/PANet_weight_' +
self.model_epoch + '.pth')
P_a.load_state_dict(param)
if self.gpu > -1:
P_a.cuda(self.gpu)
E_a = define_E(3, 8, 64, which_model_netE='resnet_128', vaeLike=True)
param = torch.load(self.model_path + '/EANet_weight_' +
self.model_epoch + '.pth')
E_a.load_state_dict(param)
if self.gpu > -1:
E_a.cuda(self.gpu)
with torch.no_grad():
interpolated_za_seq = list()
input_conditional_test = cv2.resize(V_mloop[0], (128, 128))
input_conditional_test = np.array(
[normalize(input_conditional_test)])
input_conditional_test = Variable(
torch.from_numpy(input_conditional_test.transpose(0, 3, 1, 2)))
if self.gpu > -1:
input_conditional_test = input_conditional_test.cuda(self.gpu)
za_input, _ = E_a(input_conditional_test)
interpolated_za_seq.append(za_input.clone())
with open(
self.model_path + '/codebook_a_' + self.model_epoch +
'.pkl', 'rb') as f:
codebook_a = pickle.load(
f) if sys.version_info[0] == 2 else pickle.load(
f, encoding='latin1')
za_seq = codebook_a[int((len(codebook_a) - 1) * self.t_a)]
za_seq = [torch.from_numpy(np.array([za])) for za in za_seq]
if self.gpu > -1:
za_seq = [za.cuda(self.gpu) for za in za_seq]
start_fid = None
min_dist = float('inf')
for t, mu in enumerate(za_seq):
dist = F.mse_loss(za_input, mu).cpu().numpy()
if dist < min_dist:
min_dist = dist
start_fid = t
TA = len(za_seq)
loop_num = max(minimum_loop_num,
int(np.ceil(len(za_seq) / float(len(V_mloop)))))
interpolation_size = int((loop_num * len(V_mloop) - TA) / TA)
za1 = za_input.clone()
for t in range(start_fid + 1, TA):
za2 = za_seq[t]
for ti in range(interpolation_size):
lambd = (ti + 1) / float(interpolation_size + 1)
z = (1. - lambd) * za1 + lambd * za2
interpolated_za_seq.append(z)
interpolated_za_seq.append(za2)
za1 = za2
za1 = za_input.clone()
for t in range(start_fid - 1, -1, -1):
za2 = za_seq[t]
for ti in range(interpolation_size - 1, -1, -1):
lambd = (ti + 1) / float(interpolation_size + 1)
z = (1. - lambd) * za2 + lambd * za1
interpolated_za_seq.insert(0, z)
interpolated_za_seq.insert(0, za2)
za1 = za2
loop_num = int(
np.ceil(TA * (interpolation_size + 1) / float(len(V_mloop))))
interpolation_size2 = int(interpolation_size +
loop_num * len(V_mloop) - TA *
(interpolation_size + 1))
z_start = za_input.clone() if start_fid == 0 else za_seq[0]
z_final = za_input.clone() if start_fid == TA - 1 else za_seq[-1]
for ti in range(interpolation_size2):
lambd = (ti + 1) / float(interpolation_size2 + 1)
z = (1. - lambd) * z_final + lambd * z_start
interpolated_za_seq.append(z)
zaid = (interpolation_size + 1) * start_fid
V = list()
t = 0
for loop in range(loop_num):
for frame in V_mloop:
sys.stdout.write("\rProcessing frame %d, " % (t + 1))
sys.stdout.flush()
t += 1
test_input = cv2.resize(frame, (self.w, self.h))
test_input = np.array([normalize(test_input)])
test_input = Variable(
torch.from_numpy(test_input.transpose(0, 3, 1, 2)))
if self.gpu > -1:
test_input = test_input.cuda(self.gpu)
test_input_large = np.array([normalize(frame)])
test_input_large = Variable(
torch.from_numpy(test_input_large.transpose(
0, 3, 1, 2)))
if self.gpu > -1:
test_input_large = test_input_large.cuda(self.gpu)
z = interpolated_za_seq[zaid]
y, al, bl = P_a(test_input, z)
al_large = F.upsample(al,
size=(self.fh, self.fw),
mode='bilinear',
align_corners=True)
bl_large = F.upsample(bl,
size=(self.fh, self.fw),
mode='bilinear',
align_corners=True)
y = F.tanh(al_large * test_input_large + bl_large)
outimg = y.data.cpu().numpy()[0].transpose(1, 2, 0)
V.append(denormalize(outimg))
zaid += 1
if zaid > len(interpolated_za_seq) - 1:
zaid = 0
return V
kl_loss = -0.5 * K.sum(
1 + z_log_var - K.square(z_mean) - K.square(K.exp(z_log_var)), axis=-1)
# return the average loss over all images in batch
vae_loss = K.mean(reconstruction_loss + kl_loss)
# Compile
reduce_lr = ReduceLROnPlateau(monitor='loss',
factor=0.2,
patience=5,
min_lr=0.001)
model.add_loss(vae_loss)
optimizer = optimizers.Adam(lr=learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=None,
decay=0.0,
amsgrad=False)
model.compile(optimizer=optimizer)
#checkpoint = ModelCheckpoint(filepath='x_best_weight.hdf5', verbose=1, save_best_only=True)
#for epoch in range(epochs):
history = model.fit(X_train, batch_size=batch_size, epochs=epochs)
# Predicting
pred = model.predict(X[10000:15000], batch_size=20)
pred = denormalize(pred, X_max, X_min).flatten()
set_audio('{}_{}_epoch.wav'.format(name, pct), rate, pred)