def train_dataloader(self):
dataset = TrainDataset(self.train_df, self.data_dir,
self.train_transforms)
loader = DataLoader(dataset,
batch_size=BATCH_SIZE,
num_workers=12,
shuffle=True)
return loader
def fit(self):
cfg = self.cfg
self.refiner.zero_grad()
self.refiner.train()
self.train_loader = TrainDataset(cfg.train_data_path,
size=cfg.patch_size,
scale=cfg.scale,
batch_size=cfg.batch_size)
while True:
self.train_loader.resample()
inputs = self.train_loader[0]
hr = customTransform(inputs, 0)
lr = customTransform(inputs, 1)
del inputs
if 1:
sr = self.refiner(lr)
l1loss = self.loss_fn(sr, hr) / cfg.update_every
loss = l1loss
loss.backward()
if self.step % cfg.update_every == 0:
nn.utils.clip_grad_norm_(self.refiner.parameters(),
cfg.clip)
self.optim.step()
self.refiner.zero_grad()
self.epoch += 1
learning_rate = self.decay_learning_rate()
for param_group in self.optim.param_groups:
param_group["lr"] = learning_rate
if 1:
cv2.imshow('sr', deTransform(sr[:1]))
cv2.imshow('hr', deTransform(hr[:1]))
cv2.waitKey(1)
self.step += 1
if cfg.verbose and self.step % (cfg.update_every * 10) == 0:
print('epoch', self.epoch, 'l1_loss', l1loss.item())
if cfg.verbose and self.step % (cfg.update_every *
100) == 0:
self.save()
makedirs_(opt.outf)
if opt.manualSeed is None:
opt.manualSeed = random.randint(1, 10000)
print("Random Seed: ", opt.manualSeed)
random.seed(opt.manualSeed)
torch.manual_seed(opt.manualSeed)
if opt.cuda:
torch.cuda.manual_seed_all(opt.manualSeed)
cudnn.benchmark = True
########### DATASET ###########
train_dataset = TrainDataset(opt.data_path, opt.size_w, opt.size_h)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=opt.batch_size,
shuffle=True,
num_workers=6)
########### MODEL ###########
# custom weights initialization called on netG and netD
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
m.bias.data.fill_(0)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
class Solver():
def __init__(self, model, cfg):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.cfg = cfg
self.refiner = model().to(self.device)
self.optim = optim.Adam(
filter(lambda p: p.requires_grad, self.refiner.parameters()),
cfg.lr)
self.epoch = 0
if cfg.ckpt_name != 'None':
self.load(cfg.ckpt_name)
self.step = self.epoch * cfg.update_every
learning_rate = self.decay_learning_rate()
#for param_group in self.optim.param_groups:
for param_group in self.optim.param_groups:
param_group["lr"] = learning_rate
self.loss_fn = nn.L1Loss()
if cfg.verbose:
num_params = sum(p.numel() for p in self.refiner.parameters()
if p.requires_grad)
print("# of params:", num_params)
os.makedirs(cfg.saved_ckpt_dir, exist_ok=True)
def fit(self):
cfg = self.cfg
self.refiner.zero_grad()
self.refiner.train()
self.train_loader = TrainDataset(cfg.train_data_path,
size=cfg.patch_size,
scale=cfg.scale,
batch_size=cfg.batch_size)
while True:
self.train_loader.resample()
inputs = self.train_loader[0]
hr = customTransform(inputs, 0)
lr = customTransform(inputs, 1)
del inputs
if 1:
sr = self.refiner(lr)
l1loss = self.loss_fn(sr, hr) / cfg.update_every
loss = l1loss
loss.backward()
if self.step % cfg.update_every == 0:
nn.utils.clip_grad_norm_(self.refiner.parameters(),
cfg.clip)
self.optim.step()
self.refiner.zero_grad()
self.epoch += 1
learning_rate = self.decay_learning_rate()
for param_group in self.optim.param_groups:
param_group["lr"] = learning_rate
if 1:
cv2.imshow('sr', deTransform(sr[:1]))
cv2.imshow('hr', deTransform(hr[:1]))
cv2.waitKey(1)
self.step += 1
if cfg.verbose and self.step % (cfg.update_every * 10) == 0:
print('epoch', self.epoch, 'l1_loss', l1loss.item())
if cfg.verbose and self.step % (cfg.update_every *
100) == 0:
self.save()
def load(self, path):
states = torch.load(path)
state_dict = self.refiner.state_dict()
for k, v in states.items():
if k in state_dict.keys():
state_dict.update({k: v})
self.refiner.load_state_dict(state_dict)
def save(self):
torch.save(
self.refiner.state_dict(), self.cfg.saved_ckpt_dir +
'/checkpoint_e' + str(self.epoch) + '.pth')
def decay_learning_rate(self):
lr = self.cfg.lr * (0.5**(self.epoch // self.cfg.decay))
return lr
def val_dataloader(self):
dataset = TrainDataset(self.valid_df, self.data_dir, self.valid_transforms)
loader = DataLoader(dataset, batch_size=BATCH_SIZE, num_workers=12)
return loader
train_cate1 = [cate[0] for cate in train_cate]
train_cate2 = [cate[1] for cate in train_cate]
train_cate3 = [cate[2] for cate in train_cate]
train_title = padding(train_title, max(train_t_len))
train_desc = padding(train_desc, max(train_d_len))
train_t_len = torch.tensor(train_t_len)
train_d_len = torch.tensor(train_d_len)
train_cate1 = torch.tensor(train_cate1)
train_cate2 = torch.tensor(train_cate2)
train_cate3 = torch.tensor(train_cate3)
train_set = TrainDataset(train_title, train_desc,
train_cate1, train_cate2, train_cate3,
train_t_len, train_d_len)
train_loader = data.DataLoader(train_set, batch_size=args.batch_size, shuffle=True, num_workers=4)
with open(valid_feature_path, 'rb') as fp:
valid_features = pickle.load(fp)
valid_title = [feature[0] for feature in valid_features]
valid_desc = [feature[1] for feature in valid_features]
valid_t_len = [len(feature[0]) for feature in valid_features]
valid_d_len = [len(feature[1]) for feature in valid_features]
with open(valid_cate_path, 'rb') as fp:
valid_cate = pickle.load(fp)
valid_cate1 = [cate[0] for cate in valid_cate]
valid_cate2 = [cate[1] for cate in valid_cate]
valid_cate3 = [cate[2] for cate in valid_cate]
with open(valid_cate_path, 'rb') as fp:
cate = pickle.load(fp)
cate1 = [ca[0] for ca in cate]
cate2 = [ca[1] for ca in cate]
cate3 = [ca[2] for ca in cate]
title = padding(title, max(t_len))
desc = padding(desc, max(d_len))
t_len = torch.tensor(t_len)
d_len = torch.tensor(d_len)
cate1 = torch.tensor(cate1)
cate2 = torch.tensor(cate2)
cate3 = torch.tensor(cate3)
valid_set = TrainDataset(title, desc, cate1, cate2, cate3, t_len, d_len)
valid_loader = data.DataLoader(valid_set,
batch_size=args.batch_size,
num_workers=4)
clf1_state = torch.load(os.path.join('checkpoint', 'glu_w_cate1.pth'))
clf1 = Cate1Classifier(WORDS_CNT + 1, clf1_state['args'])
clf1.load_state_dict(clf1_state['model'])
with open('./preproc/mask.pkl', 'rb') as fp:
mask1, mask2 = pickle.load(fp)
clf2_state = torch.load(os.path.join('checkpoint', 'glu_w_cate2.pth'))
clf2 = Cate2Classifier(WORDS_CNT + 1, clf2_state['args'], mask1=mask1)
clf2.load_state_dict(clf2_state['model'])