criterion = nn.CrossEntropyLoss() #multi-class classification task
model = model.to(device)
model.train()
# DONOTCHANGE: They are reserved for nsml
bind_model(model)
# below the nsml load
nsml.load(checkpoint='15', session='team_62/airush1/40')
nsml.save('stillgoing')
if args.pause:
nsml.paused(scope=locals())
if args.mode == "train":
# Warning: Do not load data before this line
dataloader = train_dataloader(args.input_size, args.batch_size,
args.num_workers)
for epoch_idx in range(1, args.epochs + 1):
total_loss = 0
total_correct = 0
for batch_idx, (image, tags) in enumerate(dataloader):
optimizer.zero_grad()
image = image.to(device) #torch.Size([64, 3, 224, 224])
tags = tags.to(device) #torch.Size([64])
output = model(image).double() # torch.Size([64, 350])
loss = criterion(output,
tags) # criterion ì ë¤ì íì¸íì
loss.backward()
optimizer.step()
if re_train_info is not None and args.mode == "train":
print(re_train_info)
nsml.load(checkpoint=re_train_info['checkpoint'],
session=re_train_info['session'])
nsml.save('dontgiveup')
if args.pause:
nsml.paused(scope=locals())
if args.mode == "train":
# Warning: Do not load data before this line
dataloader, valid_dataloader = train_dataloader(
args.input_size,
args.batch_size,
args.num_workers,
test_bs=test_bs,
br_multi_oh=use_train_time_multi_calss_info_add #)
,
print_nor_info=False,
use_last_fine_tune=use_last_fine_tune)
def validation(val_step_num):
total_valid_correct = 0
model.eval()
for batch_idx, (image, tags) in enumerate(valid_dataloader):
image = image.to(device)
tags = tags.to(device)
output = model(image).double()
output_prob = F.softmax(output, dim=1)
predict_vector = np.argmax(to_np(output_prob), axis=1)
label_vector = to_np(tags)
model.eval()
transform = None
batch_size = (256 if m_name == "Resnet18" else 32)
if args.transform == "5crop":
transform = transforms.Compose([transforms.Resize((args.input_size, args.input_size)),
transforms.FiveCrop((args.input_size, args.input_size)), transforms.Lambda(
lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops]))])
batch_size //= 5
elif args.transform == "10crop":
transform = transforms.Compose([transforms.TenCrop((args.input_size, args.input_size)), transforms.Lambda(
lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops]))])
batch_size //= 10
dataloader, val_dataloader = train_dataloader(args.input_size, batch_size,
args.num_workers,
infer_batch_size=batch_size,
transform=transform,
infer_transform=transform)
for nsml_cp in nsml_checkpoints:
# Warning: Do not load data before this line
nsml.load(checkpoint=nsml_cp, session="team_13/airush1/" + nsml_ss)
total_loss = 0.
total_correct = 0.
total_ranking_ap_score = 0.
total_ranking_loss = 0.
print(model.__class__.__name__)
print(criterion.__class__.__name__)
print("team_13/airush1/" + nsml_ss, nsml_cp)
# eval!
model = MyEnsembleTTA(modelA,modelB,modelC)
model = fuse_bn_recursively(model)
model = model.to(device) #use gpu
#summary(model, (3,args.input_size,args.input_size))
# DONOTCHANGE: They are reserved for nsml
bind_model(model)
#nsml.load(checkpoint='3', session='team_27/airush1/392 ')
#nsml.save('dontgiveup')
if args.pause:
nsml.paused(scope=locals())
if args.mode == "train":
# Warning: Do not load data before this line
dataloader, valid_dataloader = train_dataloader(args.input_size, args.batch_size*10, args.num_workers, test_bs = False
, br_multi_oh=True#)
,print_nor_info = False,use_last_fine_tune=True)
def validation(val_step_num):
total_valid_correct = 0
model.eval()
for batch_idx, (image, tags) in enumerate(valid_dataloader):
image = image.to(device)
tags = tags.to(device)
output = model(image).double()
output_prob = F.softmax(output, dim=1)
predict_vector = np.argmax(to_np(output_prob), axis=1)
label_vector = to_np(tags)
bool_vector = predict_vector == label_vector
accuracy = bool_vector.sum() / len(bool_vector)
if args.load_nsml_cp and args.nsml_checkpoint is not None and args.nsml_session is not None:
nsml.load(checkpoint=args.nsml_checkpoint,
session=args.nsml_session)
print("load", args.nsml_session, args.nsml_checkpoint)
if str.isnumeric(args.nsml_checkpoint):
epoch_start += int(args.nsml_checkpoint)
args.epochs += int(args.nsml_checkpoint)
if args.only_save:
nsml.save(args.nsml_session + "," + args.nsml_checkpoint)
else:
dataloader, val_dataloader = train_dataloader(
args.input_size,
batch_size,
args.num_workers,
infer_batch_size=infer_batch_size,
transform=transforms.Compose(transform_list),
infer_transform=transforms.Compose(infer_transform_list),
val_ratio=args.val_ratio,
use_random_label=args.use_random_label,
seed=args.seed)
if args.sava_step_ratio > 0:
save_step_interval = int(
len(dataloader) * args.sava_step_ratio)
print("save_step_interval", save_step_interval, "total steps",
len(dataloader))
else:
save_step_interval = None
for epoch_idx in range(epoch_start, args.epochs + 1):
if args.use_train:
total_loss = 0.