def build_whole_net_with_loss(self):
i_t, i_s = self.i_t, self.i_s
t_sk, t_t, t_b, t_f, mask_t = self.t_sk, self.t_t, self.t_b, self.t_f, self.mask_t
inputs = [i_t, i_s]
labels = [t_sk, t_t, t_b, t_f]
o_sk, o_t, o_b, o_f = self.build_generator(inputs)
self.o_sk = tf.identity(o_sk, name='o_sk')
self.o_t = tf.identity(o_t, name='o_t')
self.o_b = tf.identity(o_b, name='o_b')
self.o_f = tf.identity(o_f, name='o_f')
i_db_true = tf.concat([t_b, i_s], axis=-1, name='db_true_concat')
i_db_pred = tf.concat([o_b, i_s], axis=-1, name='db_pred_concat')
i_db = tf.concat([i_db_true, i_db_pred], axis=0, name='db_concat')
i_df_true = tf.concat([t_f, i_t], axis=-1, name='df_true_concat')
i_df_pred = tf.concat([o_f, i_t], axis=-1, name='df_pred_concat')
i_df = tf.concat([i_df_true, i_df_pred], axis=0, name='df_concat')
o_db = self.build_discriminator(i_db, name='db')
o_df = self.build_discriminator(i_df, name='df')
i_vgg = tf.concat([t_f, o_f], axis=0, name='vgg_concat')
vgg_graph_def = tf.GraphDef()
vgg_graph_path = cfg.vgg19_weights
with open(vgg_graph_path, 'rb') as f:
vgg_graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(vgg_graph_def,
input_map={"inputs:0": i_vgg})
with tf.Session() as sess:
o_vgg_1 = sess.graph.get_tensor_by_name(
"import/block1_conv1/Relu:0")
o_vgg_2 = sess.graph.get_tensor_by_name(
"import/block2_conv1/Relu:0")
o_vgg_3 = sess.graph.get_tensor_by_name(
"import/block3_conv1/Relu:0")
o_vgg_4 = sess.graph.get_tensor_by_name(
"import/block4_conv1/Relu:0")
o_vgg_5 = sess.graph.get_tensor_by_name(
"import/block5_conv1/Relu:0")
out_g = [o_sk, o_t, o_b, o_f, mask_t]
out_d = [o_db, o_df]
out_vgg = [o_vgg_1, o_vgg_2, o_vgg_3, o_vgg_4, o_vgg_5]
db_loss = build_discriminator_loss(o_db, name='db_loss')
df_loss = build_discriminator_loss(o_df, name='df_loss')
self.d_loss_detail = [db_loss, df_loss]
self.d_loss = tf.add(db_loss, df_loss, name='d_loss')
self.g_loss, self.g_loss_detail = build_generator_loss(out_g,
out_d,
out_vgg,
labels,
name='g_loss')
def main():
os.environ['CUDA_VISIBLE_DEVICES'] = str(cfg.gpu)
train_name = get_train_name()
print_log('Initializing SRNET', content_color = PrintColor['yellow'])
train_data = datagen_srnet(cfg)
train_data = DataLoader(dataset = train_data, batch_size = cfg.batch_size, shuffle = False, collate_fn = custom_collate, pin_memory = True)
trfms = To_tensor()
example_data = example_dataset(transform = trfms)
example_loader = DataLoader(dataset = example_data, batch_size = 1, shuffle = False)
print_log('training start.', content_color = PrintColor['yellow'])
G = Generator(in_channels = 3).cuda()
D1 = Discriminator(in_channels = 6).cuda()
D2 = Discriminator(in_channels = 6).cuda()
vgg_features = Vgg19().cuda()
G_solver = torch.optim.Adam(G.parameters(), lr=cfg.learning_rate, betas = (cfg.beta1, cfg.beta2))
D1_solver = torch.optim.Adam(D1.parameters(), lr=cfg.learning_rate, betas = (cfg.beta1, cfg.beta2))
D2_solver = torch.optim.Adam(D2.parameters(), lr=cfg.learning_rate, betas = (cfg.beta1, cfg.beta2))
#g_scheduler = torch.optim.lr_scheduler.MultiStepLR(G_solver, milestones=[30, 200], gamma=0.5)
#d1_scheduler = torch.optim.lr_scheduler.MultiStepLR(D1_solver, milestones=[30, 200], gamma=0.5)
#d2_scheduler = torch.optim.lr_scheduler.MultiStepLR(D2_solver, milestones=[30, 200], gamma=0.5)
try:
checkpoint = torch.load(cfg.ckpt_path)
G.load_state_dict(checkpoint['generator'])
D1.load_state_dict(checkpoint['discriminator1'])
D2.load_state_dict(checkpoint['discriminator2'])
G_solver.load_state_dict(checkpoint['g_optimizer'])
D1_solver.load_state_dict(checkpoint['d1_optimizer'])
D2_solver.load_state_dict(checkpoint['d2_optimizer'])
'''
g_scheduler.load_state_dict(checkpoint['g_scheduler'])
d1_scheduler.load_state_dict(checkpoint['d1_scheduler'])
d2_scheduler.load_state_dict(checkpoint['d2_scheduler'])
'''
print('Resuming after loading...')
except FileNotFoundError:
print('checkpoint not found')
pass
requires_grad(G, False)
requires_grad(D1, True)
requires_grad(D2, True)
disc_loss_val = 0
gen_loss_val = 0
grad_loss_val = 0
trainiter = iter(train_data)
example_iter = iter(example_loader)
K = torch.nn.ZeroPad2d((0, 1, 1, 0))
for step in tqdm(range(cfg.max_iter)):
D1_solver.zero_grad()
D2_solver.zero_grad()
if ((step+1) % cfg.save_ckpt_interval == 0):
torch.save(
{
'generator': G.state_dict(),
'discriminator1': D1.state_dict(),
'discriminator2': D2.state_dict(),
'g_optimizer': G_solver.state_dict(),
'd1_optimizer': D1_solver.state_dict(),
'd2_optimizer': D2_solver.state_dict(),
#'g_scheduler' : g_scheduler.state_dict(),
#'d1_scheduler':d1_scheduler.state_dict(),
#'d2_scheduler':d2_scheduler.state_dict(),
},
cfg.checkpoint_savedir+f'train_step-{step+1}.model',
)
try:
i_t, i_s, t_sk, t_t, t_b, t_f, mask_t = trainiter.next()
except StopIteration:
trainiter = iter(train_data)
i_t, i_s, t_sk, t_t, t_b, t_f, mask_t = trainiter.next()
i_t = i_t.cuda()
i_s = i_s.cuda()
t_sk = t_sk.cuda()
t_t = t_t.cuda()
t_b = t_b.cuda()
t_f = t_f.cuda()
mask_t = mask_t.cuda()
#inputs = [i_t, i_s]
labels = [t_sk, t_t, t_b, t_f]
o_sk, o_t, o_b, o_f = G(i_t, i_s)
o_sk = K(o_sk)
o_t = K(o_t)
o_b = K(o_b)
o_f = K(o_f)
#print(o_sk.shape, o_t.shape, o_b.shape, o_f.shape)
#print('------')
#print(i_s.shape)
i_db_true = torch.cat((t_b, i_s), dim = 1)
i_db_pred = torch.cat((o_b, i_s), dim = 1)
i_df_true = torch.cat((t_f, i_t), dim = 1)
i_df_pred = torch.cat((o_f, i_t), dim = 1)
o_db_true = D1(i_db_true)
o_db_pred = D1(i_db_pred)
o_df_true = D2(i_df_true)
o_df_pred = D2(i_df_pred)
i_vgg = torch.cat((t_f, o_f), dim = 0)
out_vgg = vgg_features(i_vgg)
db_loss = build_discriminator_loss(o_db_true, o_db_pred)
df_loss = build_discriminator_loss(o_df_true, o_df_pred)
db_loss.backward()
df_loss.backward()
D1_solver.step()
D2_solver.step()
#d1_scheduler.step()
#d2_scheduler.step()
clip_grad(D1)
clip_grad(D2)
if ((step+1) % 5 == 0):
requires_grad(G, True)
requires_grad(D1, False)
requires_grad(D2, False)
G_solver.zero_grad()
o_sk, o_t, o_b, o_f = G(i_t, i_s)
o_sk = K(o_sk)
o_t = K(o_t)
o_b = K(o_b)
o_f = K(o_f)
#print(o_sk.shape, o_t.shape, o_b.shape, o_f.shape)
#print('------')
#print(i_s.shape)
i_db_true = torch.cat((t_b, i_s), dim = 1)
i_db_pred = torch.cat((o_b, i_s), dim = 1)
i_df_true = torch.cat((t_f, i_t), dim = 1)
i_df_pred = torch.cat((o_f, i_t), dim = 1)
o_db_pred = D1(i_db_pred)
o_df_pred = D2(i_df_pred)
i_vgg = torch.cat((t_f, o_f), dim = 0)
out_vgg = vgg_features(i_vgg)
out_g = [o_sk, o_t, o_b, o_f, mask_t]
out_d = [o_db_pred, o_df_pred]
g_loss, detail = build_generator_loss(out_g, out_d, out_vgg, labels)
g_loss.backward()
G_solver.step()
#g_scheduler.step()
requires_grad(G, False)
requires_grad(D1, True)
requires_grad(D2, True)
if ((step+1) % cfg.write_log_interval == 0):
print('Iter: {}/{} | Gen: {} | D_bg: {} | D_fus: {}'.format(step+1, cfg.max_iter, g_loss.item(), db_loss.item(), df_loss.item()))
if ((step+1) % cfg.gen_example_interval == 0):
savedir = os.path.join(cfg.example_result_dir, train_name, 'iter-' + str(step+1).zfill(len(str(cfg.max_iter))))
with torch.no_grad():
try:
inp = example_iter.next()
except StopIteration:
example_iter = iter(example_loader)
inp = example_iter.next()
i_t = inp[0].cuda()
i_s = inp[1].cuda()
name = str(inp[2][0])
o_sk, o_t, o_b, o_f = G(i_t, i_s)
o_sk = o_sk.squeeze(0).to('cpu')
o_t = o_t.squeeze(0).to('cpu')
o_b = o_b.squeeze(0).to('cpu')
o_f = o_f.squeeze(0).to('cpu')
if not os.path.exists(savedir):
os.makedirs(savedir)
o_sk = F.to_pil_image(o_sk)
o_t = F.to_pil_image((o_t + 1)/2)
o_b = F.to_pil_image((o_b + 1)/2)
o_f = F.to_pil_image((o_f + 1)/2)
o_f.save(os.path.join(savedir, name + 'o_f.png'))
o_sk.save(os.path.join(savedir, name + 'o_sk.png'))
o_t.save(os.path.join(savedir, name + 'o_t.png'))
o_b.save(os.path.join(savedir, name + 'o_b.png'))
def main():
os.environ['CUDA_VISIBLE_DEVICES'] = str(cfg.gpu)
train_name = get_train_name()
print_log('Initializing SRNET', content_color=PrintColor['yellow'])
model = SRNet(shape=cfg.data_shape, name=train_name)
print_log('model compiled.', content_color=PrintColor['yellow'])
train_data = datagen_srnet(cfg)
train_data = DataLoader(dataset=train_data,
batch_size=cfg.batch_size,
shuffle=False,
collate=custom_collate,
pin_memory=True)
trfms = To_tensor()
example_data = example_dataset(transform=trfms)
example_loader = DataLoader(dataset=example_data,
batch_size=len(example_data),
shuffle=False)
print_log('training start.', content_color=PrintColor['yellow'])
G = Generator(in_channels=3).cuda()
D1 = discriminator(in_channels=6).cuda()
D2 = discriminator(in_channels=6).cuda()
vgg_features = Vgg19().cuda()
G_solver = torch.optim.Adam(G.parameters(),
lr=cfg.learning_rate,
betas=(cfg.beta1, cfg.beta2))
D1_solver = torch.optim.Adam(D1.parameters(),
lr=cfg.learning_rate,
betas=(cfg.beta1, cfg.beta2))
D2_solver = torch.optim.Adam(D2.parameters(),
lr=cfg.learning_rate,
betas=(cfg.beta1, cfg.beta2))
g_scheduler = torch.optim.lr_scheduler.MultiStepLR(G_solver,
milestones=[30, 200],
gamma=0.5)
d1_scheduler = torch.optim.lr_scheduler.MultiStepLR(D1_solver,
milestones=[30, 200],
gamma=0.5)
d2_scheduler = torch.optim.lr_scheduler.MultiStepLR(D2_solver,
milestones=[30, 200],
gamma=0.5)
requires_grad(G, False)
requires_grad(D1, True)
requires_grad(D2, True)
disc_loss_val = 0
gen_loss_val = 0
grad_loss_val = 0
trainiter = iter(train_data)
example_iter = iter(example_loader)
for step in tqdm(range(cfg.max_iter)):
D1.zero_grad()
D2.zero_grad()
if ((step + 1) % save_ckpt_interval == 0):
torch.save(
{
'generator': G.module.state_dict(),
'discriminator1': D1.module.state_dict(),
'discriminator2': D2.module.state_dict(),
'g_optimizer': G_solver.state_dict(),
'd1_optimizer': D1_solver.state_dict(),
'd2_optimizer': D2_solver.state_dict(),
},
f'checkpoint/train_step-{step+1}.model',
)
i_t, i_s, t_sk, t_t, t_b, t_f, mask_t = trainiter.next()
inputs = [i_t, i_s]
labels = [t_sk, t_t, t_b, t_f]
o_sk, o_t, o_b, o_f = G(inputs)
i_db_true = torch.cat((t_b, i_s), dim=1)
i_db_pred = torch.cat((o_b, i_s), dim=1)
i_df_true = torch.cat((t_f, i_t), dim=1)
i_df_pred = torch.cat((o_f, i_t), dim=1)
o_db_true = D1(i_db_true)
o_db_pred = D1(i_db_pred)
o_df_true = D2(i_df_true)
o_df_pred = D2(i_df_pred)
i_vgg = torch.cat((t_f, o_f), dim=0)
out_vgg = vgg_features(i_vgg)
db_loss = build_discriminator_loss(o_db_true, o_db_pred)
df_loss = build_discriminator_loss(o_df_true, o_df_pred)
out_g = [o_sk, o_t, o_b, o_f, mask_t]
out_d = [o_db_true, o_db_pred, o_df_true, o_df_pred]
g_loss, detail = build_generator_loss(out_g, out_d, out_vgg, labels)
db_loss.backward()
df_loss.backward()
D1_solver.step()
D2_solver.step()
d1_scheduler.step()
d2_scheduler.step()
clip_grad(D1)
clip_grad(D2)
if ((step + 1) % 5 == 0):
requires_grad(G, True)
requires_grad(D1, False)
requires_grad(D2, False)
G.zero_grad()
g_loss.backward()
G_solver.step()
g_scheduler.step()
clip_grad(G)
requires_grad(G, False)
requires_grad(D1, True)
requires_grad(D2, True)
if ((step + 1) % cfg.write_log_interval == 0):
print('Iter: {}/{} | Gen: {} | D_bg: {} | D_fus: {}'.format(
step + 1, cfg.max_iter, g_loss.item(), db_loss.item(),
df_loss.item()))
if ((step + 1) % gen_example_interval == 0):
savedir = os.path.join(
cfg.example_result_dir, train_name,
'iter-' + str(step + 1).zfill(len(str(cfg.max_iter))))
with torch.no_grad():
inp = example_iter.next()
o_sk, o_t, o_b, o_f = G(inp)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
o_sk = skimage.img_as_ubyte(o_sk)
o_t = skimage.img_as_ubyte(o_t + 1)
o_b = skimage.img_as_ubyte(o_b + 1)
o_f = skimage.img_as_ubyte(o_f + 1)
io.imsave(os.path.join(save_dir, name + 'o_f.png'), o_f)
io.imsave(os.path.join(save_dir, name + 'o_sk.png'), o_sk)
io.imsave(os.path.join(save_dir, name + 'o_t.png'), o_t)
io.imsave(os.path.join(save_dir, name + 'o_b.png'), o_b)