def train_gan(data_path, save_gen_path, save_disc_path, feedback_fn=None):
data_loader = load_dataset(path=data_path, batch_size=128)
gen = Generator().to('cuda')
disc = Discriminator().to('cuda')
criterion = torch.nn.BCEWithLogitsLoss()
lr = 3e-4
gen_opt = torch.optim.Adam(gen.parameters(), lr=lr, betas=(0.5, 0.999))
disc_opt = torch.optim.Adam(disc.parameters(), lr=lr, betas=(0.5, 0.999))
gen = gen.apply(weights_init)
disc = disc.apply(weights_init)
n_epochs = 5
gen, disc = _train_loop(data_loader=data_loader,
gen=gen,
disc=disc,
criterion=criterion,
gen_opt=gen_opt,
disc_opt=disc_opt,
n_epochs=5,
feedback_fn=feedback_fn)
torch.save(gen.state_dict(), save_gen_path)
torch.save(disc.state_dict(), save_disc_path)
def main(config):
# For fast training.
cudnn.benchmark = True
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# running preparation
prepareDirs(config)
writer = SummaryWriter(config.log_dir)
# Data preparation
data_loader = getProperLoader(config)
# Model Initialization
G = Generator(config.g_conv_dim, config.c_dim, config.g_repeat_num)
D = Discriminator(config.image_size, config.d_conv_dim, config.c_dim,
config.d_repeat_num)
g_optimizer = torch.optim.Adam(G.parameters(), config.g_lr,
[config.beta1, config.beta2])
d_optimizer = torch.optim.Adam(D.parameters(), config.d_lr,
[config.beta1, config.beta2])
G.to(device)
D.to(device)
# Training/Test
if config.mode == 'train':
train(config, G, D, g_optimizer, d_optimizer, data_loader, device,
writer)
elif config.mode == 'test':
test(config, G, data_loader, device)
criterion = criterion.to(device)
cri_perception = VGGFeatureExtractor().to(device)
cri_gan = GANLoss('vanilla', 1.0, 0.0).to(device)
# textual init
vgg19 = cri_perception.vggNet
loss_layer = [1, 6, 11, 20]
loss_fns = [GramMSELoss()] * len(loss_layer)
if torch.cuda.is_available():
loss_fns = [loss_fn.cuda() for loss_fn in loss_fns]
style_weights = [1e3 / n**2 for n in [64, 128, 256, 512]]
optimizer = torch.optim.RMSprop(
filter(lambda p: p.requires_grad, model.parameters()), 0.0001)
optimizer_D = torch.optim.RMSprop(
filter(lambda p: p.requires_grad, netD.parameters()), 0.0002)
print('# generator parameters:',
sum(param.numel() for param in model.parameters()))
print('# discriminator parameters:',
sum(param.numel() for param in netD.parameters()))
print()
for i in range(opt.start_training_step, 4):
opt.nEpochs = training_settings[i - 1]['nEpochs']
opt.lr = training_settings[i - 1]['lr']
opt.step = training_settings[i - 1]['step']
opt.lr_decay = training_settings[i - 1]['lr_decay']
opt.lambda_db = training_settings[i - 1]['lambda_db']
opt.gated = training_settings[i - 1]['gated']
print(opt)
for epoch in range(opt.start_epoch, opt.nEpochs + 1):
else:
model = Net()
netD = Discriminator()
mkdir_steptraing()
model = model.to(device)
netD = netD.to(device)
criterion = torch.nn.L1Loss(size_average=True)
criterion = criterion.to(device)
cri_perception = VGGFeatureExtractor().to(device)
optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, model.parameters()), 0.0001,
[0.9, 0.999])
optimizer_D = torch.optim.Adam(
filter(lambda p: p.requires_grad, netD.parameters()), 0.0002, [0.9, 0.999])
print()
for i in range(opt.start_training_step, 4):
opt.nEpochs = training_settings[i - 1]['nEpochs']
opt.lr = training_settings[i - 1]['lr']
opt.step = training_settings[i - 1]['step']
opt.lr_decay = training_settings[i - 1]['lr_decay']
opt.lambda_db = training_settings[i - 1]['lambda_db']
opt.gated = training_settings[i - 1]['gated']
print(opt)
for epoch in range(opt.start_epoch, opt.nEpochs + 1):
lr = adjust_learning_rate(epoch - 1)
trainloader = CreateDataLoader(opt)
train(trainloader, model, netD, criterion, optimizer, epoch, lr)
if epoch % 5 == 0:
netD = netD.to(device)
criterion = torch.nn.MSELoss(size_average=True)
criterion = criterion.to(device)
cri_perception = VGGFeatureExtractor().to(device)
cri_gan = GANLoss('vanilla', 1.0, 0.0).to(device)
# textual init
vgg19 = cri_perception.vggNet
loss_layer = [1, 6, 11, 20]
loss_fns = [GramMSELoss()] * len(loss_layer)
if torch.cuda.is_available():
loss_fns = [loss_fn.cuda() for loss_fn in loss_fns]
style_weights = [1e3 / n ** 2 for n in [64, 128, 256, 512]]
optimizer = torch.optim.RMSprop(filter(lambda p: p.requires_grad, model.parameters()), 0.0001)
optimizer_D = torch.optim.RMSprop(filter(lambda p: p.requires_grad, netD.parameters()), 0.0002)
print('# generator parameters:', sum(param.numel() for param in model.parameters()))
print('# discriminator parameters:', sum(param.numel() for param in netD.parameters()))
print()
for i in range(opt.start_training_step, 4):
opt.nEpochs = training_settings[i-1]['nEpochs']
opt.lr = training_settings[i-1]['lr']
opt.step = training_settings[i-1]['step']
opt.lr_decay = training_settings[i-1]['lr_decay']
opt.lambda_db = training_settings[i-1]['lambda_db']
opt.gated = training_settings[i-1]['gated']
print(opt)
for epoch in range(opt.start_epoch, opt.nEpochs+1):
lr = adjust_learning_rate(epoch-1)
# model_ESR = RRDBNet(in_nc=3, out_nc=3, nf=64, nb=23, gc=32, norm_type=None)
netD = Discriminator()
mkdir_steptraing()
# model_msrn = model_msrn.to(device)
model_GFN_deblur.to(device)
# model_GFN_sr = model_GFN_sr.to(device)
# model_DebulrGAN = model_DebulrGAN.to(device)
# model_ESR = model_ESR.to(device)
netD = netD.to(device)
# print('# MSRN parameters:', sum(param.numel() for param in model_msrn.parameters()))
print('# GFN_deblur parameters:', sum(param.numel() for param in model_GFN_deblur.parameters()))
# print('# sr parameters:', sum(param.numel() for param in model_GFN_sr.parameters()))
# print('# deblurGAN parameters:', sum(param.numel() for param in model_GFN_deblur.parameters()))
# print('# ESR parameters:', sum(param.numel() for param in model_ESR.parameters()))
print('# discriminator parameters:', sum(param.numel() for param in netD.parameters()))
criterion = torch.nn.L1Loss(size_average=True)
criterion = criterion.to(device)
cri_perception = VGGFeatureExtractor().to(device)
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model_GFN_deblur.parameters()), 0.0001, [0.9, 0.999])
optimizer_D = torch.optim.Adam(filter(lambda p: p.requires_grad, netD.parameters()), 0.0004, [0.9, 0.999])
print()
# for i in range(opt.start_training_step, 4):
# opt.nEpochs = training_settings[i-1]['nEpochs']
# opt.lr = training_settings[i-1]['lr']
# opt.step = training_settings[i-1]['step']
# opt.lr_decay = training_settings[i-1]['lr_decay']
# opt.lambda_db = training_settings[i-1]['lambda_db']
netD = torch.load(opt.resumeD)
netD.load_state_dict(netD.state_dict())
opt.start_training_step, opt.start_epoch = which_trainingstep_epoch(opt.resume)
else:
model = Net()
netD = Discriminator()
mkdir_steptraing()
model = model.to(device)
netD = netD.to(device)
criterion = torch.nn.L1Loss(size_average=True)
criterion = criterion.to(device)
cri_perception = VGGFeatureExtractor().to(device)
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), 0.0001)
optimizer_D = torch.optim.Adam(filter(lambda p: p.requires_grad, netD.parameters()), 0.0004)
print()
for i in range(opt.start_training_step, 4):
opt.nEpochs = training_settings[i-1]['nEpochs']
opt.lr = training_settings[i-1]['lr']
opt.step = training_settings[i-1]['step']
opt.lr_decay = training_settings[i-1]['lr_decay']
opt.lambda_db = training_settings[i-1]['lambda_db']
opt.gated = training_settings[i-1]['gated']
print(opt)
for epoch in range(opt.start_epoch, opt.nEpochs+1):
lr = adjust_learning_rate(epoch-1)
trainloader = CreateDataLoader(opt)
train(trainloader, model, netD, criterion, optimizer, epoch, lr)
class LstmGan(object):
def __init__(self,
config=None,
train_loader=None,
test_loader=None,
gt_loader=None):
self.config = config
self.train_loader = train_loader
self.test_loader = test_loader
self.gt_loader = gt_loader
# build lstm-gan model
def build(self):
# build Summarizier
self.summarizer = Summarizer(self.config.input_size,
self.config.hidden_size,
self.config.num_layers).cuda()
# build Discriminator
self.discriminator = Discriminator(self.config.input_size,
self.config.hidden_size,
self.config.num_layers).cuda()
# hold and register submodules in a list
self.model = nn.ModuleList([self.summarizer, self.discriminator])
if self.config.mode == 'train':
# build optimizers
self.s_e_optimizer = optim.Adam(
list(self.summarizer.slstm.parameters()) +
list(self.summarizer.vae.elstm.parameters()),
lr=self.config.sum_learning_rate)
self.d_optimizer = optim.Adam(list(
self.summarizer.vae.dlstm.parameters()),
lr=self.config.sum_learning_rate)
self.c_optimizer = optim.Adam(list(
self.discriminator.parameters()),
lr=self.config.dis_learning_rate)
# set the model in training mode
self.model.train()
# initialize log writer
self.writer = LogWriter(self.config.log_dir)
"""
reconstruct loss
Args:
fea_h_last: given original vodieo feature, the output of the last hidden (top) layer of cLSTM, (1, hidden) = (1, 1024)
dec_h_last: given decoded video feature, the output of the last hidden (top) layer of cLSTM, (1, hidden) = (1, 1024)
"""
def get_reconst_loss(self, fea_h_last, dec_h_last):
# L-2 norm
return torch.norm(fea_h_last - dec_h_last, p=2)
"""
summary-length regularization
Args:
scores: (seq_len, 1)
"""
def get_sparsity_loss(self, scores):
# convert scores to 0-1
#scores = scores.ge(0.5).float()
print(scores)
return torch.abs(torch.mean(scores) - self.config.summary_rate)
"""
dpp loss
"""
def get_dpp_loss(self, scores):
# convert scores to 0-1
#scores = scores.ge(0.5).float()
seq_len = len(scores)
dpp_loss = 0
for i in range((seq_len - 1)):
dpp_loss += (scores[i] * scores[i + 1] + (1 - scores[i]) *
(1 - scores[i + 1]))
return torch.log(1 + dpp_loss)
"""
gan loss of cLSTM
Args:
fea_prob: a scalar of original video feature
dec_prob: a scalar of keyframe-based reconstruction video feature
rand_dec_prob: a scalar of random-frame-based reconstruction video feature
"""
def get_gan_loss(self, fea_prob, dec_prob, rand_dec_prob):
gan_loss = torch.log(fea_prob) + torch.log(1 - dec_prob) + torch.log(
1 - rand_dec_prob)
return gan_loss
# train model
def train(self):
step = 0
for epoch_i in range(self.config.n_epochs):
#for epoch_i in trange(self.config.n_epochs, desc = 'Epoch'):
s_e_loss_history = []
d_loss_history = []
c_loss_history = []
# one video is a batch
for batch_i, feature in enumerate(
tqdm(self.train_loader,
desc='Batch',
ncols=240,
leave=False)):
# feature: (1, seq_len, input_size) -> (seq_len, 1, 2048)
feature = feature.view(-1, 1, self.config.input_size)
# from cpu tensor to gpu Variable
feature = Variable(feature).cuda()
""" train sLSTM and eLSTM """
if self.config.detail_flag:
tqdm.write('------Training sLSTM and eLSTM------')
# decoded: decoded feature generated by dLSTM
scores, decoded = self.summarizer(feature)
#scores.masked_fill_(scores >= 0.5, 1)
#scores.masked_fill_(scores < 0.5, 0)
# shape of feature and decoded: (seq_len, 1, 2048)
# shape of fea_h_last and dec_h_last: (1, hidden_size) = (1, 2048)
fea_h_last, fea_prob = self.discriminator(feature)
dec_h_last, dec_prob = self.discriminator(decoded)
tqdm.write('fea_prob: %.3f, dec_prob: %.3f}' %
(fea_prob.data[0], dec_prob.data[0]))
# reconstruction loss
reconst_loss = self.get_reconst_loss(fea_h_last, dec_h_last)
tqdm.write('reconst_loss: %.3f' % reconst_loss.data[0])
# sparsity loss
sparsity_loss = self.get_sparsity_loss(scores)
tqdm.write('sparsity_loss: %.3f' % sparsity_loss.data[0])
# dpp loss
dpp_loss = self.get_dpp_loss(scores)
tqdm.write('diversity_loss: %.3f' % dpp_loss.data[0])
# minimize
s_e_loss = reconst_loss + sparsity_loss + dpp_loss
self.s_e_optimizer.zero_grad()
s_e_loss.backward(retain_graph=True)
self.s_e_optimizer.step()
# add to loss history
s_e_loss_history.append(s_e_loss.data.cpu())
""" train dLSTM """
if self.config.detail_flag:
tqdm.write('------Training dLSTM------')
# randomly select a subset of frames
_, rand_decoded = self.summarizer(feature,
random_score_flag=True)
# shape of rand_dec_h_last: (1, hidden_size) = (1, 2048)
rand_dec_h_last, rand_dec_prob = self.discriminator(
rand_decoded)
# gan loss
gan_loss = self.get_gan_loss(fea_prob, dec_prob, rand_dec_prob)
tqdm.write('gan_loss: %.3f' % gan_loss.data[0])
# minimize
d_loss = reconst_loss + gan_loss
self.d_optimizer.zero_grad()
d_loss.backward(retain_graph=True)
self.d_optimizer.step()
# add to loss history
d_loss_history.append(d_loss.data.cpu())
""" train cLSTM """
if batch_i > self.config.dis_start_batch:
if self.config.detail_flag:
tqdm.write('------Training cLSTM------')
# maximize
c_loss = -1 * self.get_gan_loss(fea_prob, dec_prob,
rand_dec_prob)
self.c_optimizer.zero_grad()
c_loss.backward()
self.c_optimizer.step()
# add to loss history
c_loss_history.append(c_loss.data.cpu())
if self.config.detail_flag:
tqdm.write('------plotting------')
self.writer.update_loss(reconst_loss.cpu().data.numpy(), step,
'reconst_loss')
self.writer.update_loss(sparsity_loss.cpu().data.numpy(), step,
'sparsity_loss')
self.writer.update_loss(gan_loss.cpu().data.numpy(), step,
'gan_loss')
self.writer.update_loss(fea_prob.cpu().data.numpy(), step,
'fea_prob')
self.writer.update_loss(dec_prob.cpu().data.numpy(), step,
'dec_prob')
self.writer.update_loss(rand_dec_prob.cpu().data.numpy(), step,
'rand_dec_prob')
# bacth over
step += 1
# epoch
s_e_loss = torch.stack(s_e_loss_history).mean()
d_loss = torch.stack(d_loss_history).mean()
c_loss = torch.stack(c_loss_history).mean()
# plot
if self.config.detail_flag:
tqdm.write('------plotting------')
self.writer.update_loss(s_e_loss, epoch_i, 's_e_epoch')
self.writer.update_loss(d_loss, epoch_i, 'd_loss_epoch')
self.writer.update_loss(c_loss, epoch_i, 'c_loss_epoch')
# save parameters at checkpoint
model_path = str(self.config.model_save_dir) + '/' + 'model.pkl'
tqdm.write('------save model parameters at %s' % model_path)
torch.save(self.model.state_dict(), model_path)
def test(self):
model_path = str(self.config.model_save_dir) + '/' + 'model.pkl'
# load model
self.model.load_state_dict(torch.load(model_path))
self.model.eval()
ground_truth = self.gt_loader
pred = {}
cnt = 0
precision = []
for sample_i, feature in enumerate(
tqdm(self.test_loader, desc='Evaluate', leave=False)):
# feautre: (seq_len, 1, 2048)
feature = feature.view(-1, 1, self.config.input_size)
scores = self.summarizer.slstm(
Variable(feature.cuda(), volatile=True))
# convert scores to a list
scores = scores.data.cpu().numpy().flatten()
pred[sample_i] = scores
print(scores)
# sort socres in descending order, and return index
idx_sorted_scores = sorted(range(len(scores)),
key=lambda k: scores[k],
reverse=True)
num_keyframes = int(len(scores) * self.config.summary_rate)
idx_keyframes = idx_sorted_scores[:num_keyframes]
# sort key frames index in ascending order
idx_keyframes = sorted(idx_keyframes)
# precision for a single video
single_precision = float(
len(set(ground_truth[cnt]) & set(idx_keyframes))) / float(
len(ground_truth[cnt]))
# add single video precision to the whole precision list
precision.append(single_precision)
cnt += 1
# average precision for the dataset
avg_precision = sum(precision) / float(len(precision))
model.load_state_dict(model.state_dict())
netD = torch.load(opt.resumeD)
netD.load_state_dict(netD.state_dict())
# opt.start_epoch = which_trainingstep_epoch(opt.resume)
else:
model = Generator()
netD = Discriminator()
mkdir_steptraing()
model = model.to(device)
netD = netD.to(device)
criterion = torch.nn.MSELoss(size_average=True)
criterion = criterion.to(device)
optimizer = optim.Adam(model.parameters(), lr=0.0001)
optimizer_D = optim.Adam(netD.parameters(), lr=0.0002)
print()
# opt.start_training_step = 2
# for i in range(opt.start_training_step, 4):
# opt.nEpochs = training_settings[i-1]['nEpochs']
# opt.lr = training_settings[i-1]['lr']
# opt.step = training_settings[i-1]['step']
# opt.lr_decay = training_settings[i-1]['lr_decay']
# opt.lambda_db = training_settings[i-1]['lambda_db']
# opt.gated = training_settings[i-1]['gated']
# print(opt)
opt.start_epoch = 2
opt.nEpochs = 1000
# trainloader = CreateDataLoader(opt)
return NotImplementedError(
'learning rate policy [%s] is not implemented', opt.lr_policy)
return scheduler
# update learning rate (called once every epoch)
def update_learning_rate(scheduler, optimizer):
scheduler.step()
lr = optimizer.param_groups[0]['lr']
print('learning rate = %.7f' % lr)
optimizer_g = torch.optim.Adam(mysegnet.parameters(),
lr=opt.g_lr,
betas=(opt.beta1, 0.999))
optimizer_d = torch.optim.Adam(mydisnet.parameters(),
lr=opt.d_lr,
betas=(opt.beta1, 0.999))
net_g_scheduler = get_scheduler(optimizer_g, opt)
net_d_scheduler = get_scheduler(optimizer_d, opt)
alpha = opt.alpha
for epoch in range(opt.epoch_count, opt.niter + opt.niter_decay + 1):
# train
for iteration, ((batchX, batchY), (batchY1)) in enumerate(
zip(training_iterator, train_unpair_iterator)):
# forward
# input data
batchX = batchX.unsqueeze(1).to(device='cuda').float()
batchY = batchY.unsqueeze(1).to(device='cuda').float()