def main():
if args.device == 'cuda':
torch.backends.cudnn.benchmark = True # should add to speed up the code when input array shape doesn't vary
print('Using cudnn.benchmark.')
model = model_dispatcher(True, args.base_model, args.nclass)
model.to(args.device)
train_size = len(pd.read_csv(args.train_file))
print(train_size)
train_dataset = ImageSamplerDataset(phase='train',
train_file=args.train_file,
image_file_path=args.image_file,
image_height=args.image_height,
image_width=args.image_width,
mean=MODEL_MEAN,
std=MODEL_STD,
binclass=args.binclass)
train_dataloader = DataLoader(dataset=train_dataset,
batch_size=args.train_batch_size,
num_workers=args.num_workers)
optimizer = AdamW(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
scheduler_cosine = CosineAnnealingLR(optimizer, args.epochs)
if torch.cuda.device_count() > 1:
model = nn.DataParallel()
for epoch in range(args.epochs):
tr_loss = train(dataset_size=train_size,
dataloader=train_dataloader,
model=model,
optimizer=optimizer,
device=args.device,
loss_fn=focal_loss_fn)
print(f'Epoch_{epoch+1} Train Loss:{tr_loss}')
scheduler_cosine.step(epoch)
torch.save(
model.state_dict(),
os.path.join(args.save_dir, f'{args.base_model}_on_all_epoch11.bin'))
print('train on all is complete')
def main():
model = model_dispatcher(False, args.base_model, args.nclass)
model.to(args.device)
model.load_state_dict(
torch.load(os.path.join(args.save_dir, args.model_weights)))
model.eval()
print(f'Loading pretrained model: {args.base_model} for eval')
test_dataset = ImageTestDataset(file_path=args.image_file,
image_height=args.image_height,
image_width=args.image_width,
mean=MODEL_MEAN,
std=MODEL_STD)
test_dataloader = DataLoader(
dataset=test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers,
)
image_id_list = []
image_pred_list = []
with torch.no_grad():
for batch_id, d in enumerate(tqdm(test_dataloader)):
image = d['image']
img_id = d['image_id']
image = image.to(args.device, dtype=torch.float)
outputs = model(image)
pred_prob = torch.nn.Softmax(dim=1)(outputs)
image_id_list.append(img_id)
image_pred_list.append(pred_prob)
preds = torch.cat(image_pred_list).cpu().numpy()
ids = list(chain(*image_id_list))
df_pred = pd.DataFrame(preds, columns=['A', 'B', 'C'])
df_id = pd.DataFrame(ids, columns=['image_ids'])
sub = pd.concat([df_id, df_pred], axis=1)
sub.to_csv(f'{args.output_name}.csv', index=False)
def main():
model = model_dispatcher(False, args.base_model, args.nclass)
model.to(args.device)
model.load_state_dict(torch.load(os.path.join(args.save_dir,args.model_weights)))
model.eval()
print(f'Loading pretrained model: {args.base_model} for eval')
test_dataset = ImageTestDataset(
file_path = args.image_file,
image_height=args.image_height,
image_width=args.image_width,
mean = MODEL_MEAN,
std = MODEL_STD
)
test_dataloader = DataLoader(
dataset=test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers,
)
image_id_list = []
image_feat_list = []
with torch.no_grad():
for batch_id, d in enumerate(tqdm(test_dataloader)):
image = d['image']
img_id = d['image_id']
image = image.to(args.device, dtype=torch.float)
model(image)
img_features = model.imgfeatures
image_id_list.append(img_id)
image_feat_list.extend(img_features.cpu().numpy().tolist())
preds = image_feat_list
ids = list(chain(*image_id_list))
sub = pd.DataFrame({'image_ids':ids, 'features':preds})
sub.to_csv(f'{args.base_model}_imagefeatures.csv',index=False)
def main():
model = model_dispatcher(False, args.base_model, args.nclass)
model.to(args.device)
model.load_state_dict(
torch.load(os.path.join(args.save_dir, args.model_weights)))
# checkpoint = torch.load(os.path.join(args.save_dir,args.model_weights), map_location=args.device)
# model.load_state_dict(checkpoint)
model.eval()
print(f'Loading pretrained model: {args.base_model} for eval')
for num_tta in range(args.num_tta):
if num_tta == 0:
test_dataset = ImageTTADataset(file_path=args.image_file,
transform=data_transforms_test)
test_dataloader = DataLoader(test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers)
elif num_tta == 1:
test_dataset = ImageTTADataset(file_path=args.image_file,
transform=data_transforms_tta1)
test_dataloader = DataLoader(test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers)
elif num_tta == 2:
test_dataset = ImageTTADataset(file_path=args.image_file,
transform=data_transforms_tta2)
test_dataloader = DataLoader(test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers)
elif num_tta == 3:
test_dataset = ImageTTADataset(file_path=args.image_file,
transform=data_transforms_tta3)
test_dataloader = DataLoader(test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers)
elif num_tta < 8:
test_dataset = ImageTTADataset(file_path=args.image_file,
transform=data_transforms_tta0)
test_dataloader = DataLoader(test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers)
else:
test_dataset = ImageTTADataset(file_path=args.image_file,
transform=data_transforms)
test_dataloader = DataLoader(test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.num_workers)
image_id_list = []
image_pred_list = []
with torch.no_grad():
for d in test_dataloader:
image = d['image']
img_id = d['image_id']
image = image.to(args.device, dtype=torch.float)
outputs = model(image)
pred_prob = torch.nn.Softmax(dim=1)(outputs)
image_id_list.append(img_id)
image_pred_list.append(pred_prob / args.num_tta)
if num_tta == 0:
ids = list(chain(*image_id_list))
preds = torch.cat(image_pred_list).cpu().numpy()
else:
preds_tmp = torch.cat(image_pred_list).cpu().numpy()
preds += preds_tmp
print(num_tta)
df_pred = pd.DataFrame(preds, columns=['A', 'B', 'C'])
df_id = pd.DataFrame(ids, columns=['image_ids'])
sub = pd.concat([df_id, df_pred], axis=1)
sub.to_csv(f'{args.output_name}.csv', index=False)
def main():
if args.device == 'cuda':
torch.backends.cudnn.benchmark = True # should add to speed up the code when input array shape doesn't vary
print('Using cudnn.benchmark.')
model = model_dispatcher(True, args.base_model, args.nclass)
model.to(args.device)
# print(f'Loading pretrained model: {args.base_model}')
train_indices, val_indices = get_train_valid_indice(
test_size=args.test_size, random_state=args.random_state)
train_size = len(train_indices)
valid_size = len(val_indices)
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
train_dataset = ImageSamplerDataset(phase='train',
train_file=args.train_file,
image_file_path=args.image_file,
image_height=args.image_height,
image_width=args.image_width,
mean=MODEL_MEAN,
std=MODEL_STD,
binclass=args.binclass)
train_dataloader = DataLoader(dataset=train_dataset,
batch_size=args.train_batch_size,
num_workers=args.num_workers,
sampler=train_sampler)
valid_dataset = ImageSamplerDataset(phase='valid',
train_file=args.train_file,
image_file_path=args.image_file,
image_height=args.image_height,
image_width=args.image_width,
mean=MODEL_MEAN,
std=MODEL_STD,
binclass=args.binclass)
valid_dataloader = DataLoader(dataset=valid_dataset,
batch_size=args.test_batch_size,
num_workers=args.num_workers,
sampler=valid_sampler)
optimizer = AdamW(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
mode='max',
patience=5,
factor=0.3)
if torch.cuda.device_count() > 1:
model = nn.DataParallel()
val_accu_benchmark = 0.34
val_loss_list = []
val_accu_list = []
tr_loss_list = []
# tr_accu_list = []
best_epoch = 0
for epoch in range(args.epochs):
tr_loss = train(dataset_size=train_size,
dataloader=train_dataloader,
model=model,
optimizer=optimizer,
device=args.device,
loss_fn=focal_loss_fn)
val_loss, val_accu = evaluate(dataset_size=valid_size,
dataloader=valid_dataloader,
model=model,
device=args.device,
loss_fn=focal_loss_fn,
tag='valid')
print(f'Epoch_{epoch+1} Train Loss:{tr_loss}')
print(f'Epoch_{epoch+1} Valid Loss:{val_loss}')
scheduler.step(val_loss)
tr_loss_list.append(tr_loss)
# tr_accu_list.append(tr_accu)
val_loss_list.append(val_loss)
val_accu_list.append(val_accu)
if val_accu > val_accu_benchmark:
best_epoch = epoch + 1
print(f'save {args.base_model} model on epoch {epoch+1}')
torch.save(
model.state_dict(),
os.path.join(args.save_dir,
f'{args.base_model}_fold_{VALID_FOLDS[0]}.bin'))
val_accu_benchmark = val_accu
print(f'Save the best model on epoch {best_epoch}')
stored_metrics = {
'train': {
'tr_loss_list': tr_loss_list
# , 'tr_accu_list': tr_accu_list
},
'valid': {
'val_loss_list': val_loss_list,
'val_accu_list': val_accu_list
}
}
# pickle a variable to a file
file = open(os.path.join(args.save_dir, 'stored_metrics.pickle'), 'wb')
pickle.dump(stored_metrics, file)
file.close()