n_epochs = 10000
input_size = 256
batch_size = 30
lr = 3e-4
weight_decay = 5e-5
num_workers = 8
ignore_index = 250
""" Scribble2Label Params """
thr_epoch = 100
period_epoch = 5
thr_conf = 0.8
alpha = 0.2
if __name__ == '__main__':
seed_everything(config.seed)
model = Unet(encoder_name='resnet50', encoder_weights='imagenet', decoder_use_batchnorm=True,
decoder_attention_type='scse', classes=2, activation=None)
df = pd.read_csv(config.df_path)
train_df = df[df.fold != config.fold].reset_index(drop=True)
valid_df = df[df.fold == config.fold].reset_index(drop=True)
transforms = get_transforms(config.input_size, need=('train', 'val'))
train_dataset = dsbDataset(config.data_dir, config.scr_dir, config.mask_dir, train_df,
tfms=transforms['train'], return_id=False)
valid_dataset = dsbDataset(config.data_dir, config.scr_dir, config.mask_dir, valid_df,
tfms=transforms['val'], return_id=True)
train_loader = DataLoader(dataset=train_dataset, batch_size=config.batch_size, num_workers=config.num_workers,
shuffle=True)
distrib_config = {}
distrib_config['LOCAL_RANK'] = args.local_rank
root_dir = Path(train_config['DIRECTORY']['ROOT_DIRECTORY'])
data_dir = Path(train_config['DIRECTORY']['DATA_DIRECTORY'])
log_dir = root_dir / train_config['DIRECTORY']['LOGGER_DIRECTORY']
log_dir.mkdir(exist_ok=True, parents=True)
if distrib_config['LOCAL_RANK'] == 0:
main_logger = init_logger(log_dir, 'train_main.log')
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.enabled = True
torch.backends.cudnn.deterministic = True
SEED = train_config['SEED']
seed_everything(SEED)
if distrib_config['LOCAL_RANK'] == 0:
main_logger.info(train_config)
# if "DEVICE_LIST" in train_config:
# os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(map(str, train_config["DEVICE_LIST"]))
# if len(train_config['DEVICE_LIST']) > 1:
# distrib_config['DISTRIBUTED'] = True
# torch.cuda.set_device(distrib_config['LOCAL_RANK'])
# torch.distributed.init_process_group(backend='nccl',
# init_method='env://')
# distrib_config['WORLD_SIZE'] = torch.distributed.get_world_size()
# train_config['OPTIMIZER']['ARGS']['lr'] = train_config['OPTIMIZER']['ARGS']['lr'] * float(
# train_config['BATCH_SIZE'] * distrib_config['WORLD_SIZE']) / 256
# else:
if config.cat_emb_dim > 0: # cat_emb_dim이 0보다 크면 모델에서 cate col 사용
input_cols = cate_cols + input_cols
if config.use_autoencoder:
input_cols = input_cols + ae_cols
if config.normalize:
pass # TODO; add normalization
is_cuda = config.device == 'cuda'
oof_preds = []
oof_targets = []
scores = []
seed = config.seed[0] # use one seed
seed_everything(seed)
for search_iter in range(NUM_SEARCH):
# change
rp(config)
for fold in [0]:
# for fold in range(config.n_folds):
print(f'[[ ITER {search_iter} Fold {fold} ]]')
fold_train = train_features[
train_features.fold != fold].reset_index(drop=True)
fold_valid = train_features[train_features.fold ==
fold].reset_index(drop=True)
input_features = []
if config.cat_emb_dim > 0: # cat_emb_dim이 0보다 크면 모델에서 cate col 사용; 항상 제일 앞에 와야함