def main():
args = parse_args()
with open(args.config) as f:
cfg = yaml.load(f, Loader=yaml.FullLoader)
# We will set all the seeds we can, in vain ...
set_reproducibility(cfg['seed'])
# Set GPU
torch.cuda.set_device(args.gpu)
if args.num_workers > 0:
if 'transform' not in cfg.keys():
cfg['transform'] = {}
cfg['transform']['num_workers'] = args.num_workers
cfg = set_inference_batch_size(cfg)
if args.mode == 'predict':
predict(args, cfg)
return
elif args.mode == 'predict_ensemble':
predict_ensemble(args, cfg)
return
if 'mixup' not in cfg['train']['params'].keys():
cfg['train']['params']['mixup'] = None
# Make directory to save checkpoints
if not osp.exists(cfg['evaluation']['params']['save_checkpoint_dir']):
os.makedirs(cfg['evaluation']['params']['save_checkpoint_dir'])
# Load in labels with CV splits
df = pd.read_csv(cfg['dataset']['csv_filename'])
if 'chromosome' in cfg['dataset'].keys():
if cfg['dataset']['chromosome'].lower() == 'xy':
df = df[df['male']]
elif cfg['dataset']['chromosome'].lower() == 'xx':
df = df[~df['male']]
else:
raise Exception('`chromosome` must be one of [`XY`,`XX`]')
ofold = cfg['dataset']['outer_fold']
ifold = cfg['dataset']['inner_fold']
train_df, valid_df, test_df = get_train_valid_test(cfg, df, ofold, ifold)
logger = create_logger(cfg, args.mode)
logger.info('Saving to {} ...'.format(cfg['evaluation']['params']['save_checkpoint_dir']))
if args.mode == 'find_lr':
cfg['optimizer']['params']['lr'] = cfg['find_lr']['params']['start_lr']
find_lr(args, cfg, train_df, valid_df)
elif args.mode == 'train':
train(args, cfg, train_df, valid_df)
elif args.mode == 'test':
test(args, cfg, test_df)
def main():
args = parse_args()
with open(args.config) as f:
cfg = yaml.load(f, Loader=yaml.FullLoader)
if args.num_workers > 0:
cfg['transform']['num_workers'] = args.num_workers
cfg = set_inference_batch_size(cfg)
# We will set all the seeds we can, in vain ...
set_reproducibility(cfg['seed'])
# Set GPU
torch.cuda.set_device(args.gpu)
# Make directory to save checkpoints
if not osp.exists(cfg['evaluation']['params']['save_checkpoint_dir']):
os.makedirs(cfg['evaluation']['params']['save_checkpoint_dir'])
# Load in labels with CV splits
with open(cfg['dataset']['pickled'], 'rb') as f:
ann = pickle.load(f)
if 'positives_only' in cfg['dataset'].keys(
) and cfg['dataset']['positives_only']:
print('Filtering out negative images ...')
ann = [a for a in ann if a['ann']['bboxes'].shape[0] > 0]
ofold = cfg['dataset']['outer_fold']
ifold = cfg['dataset']['inner_fold']
if args.mode != 'predict':
train_ann, valid_ann, test_ann = get_train_valid_test(
cfg, ann, ofold, ifold, args.mode)
else:
pred_ann = ann
logger = create_logger(cfg, args.mode)
logger.info('Saving to {} ...'.format(
cfg['evaluation']['params']['save_checkpoint_dir']))
if args.mode == 'find_lr':
cfg['optimizer']['params']['lr'] = cfg['find_lr']['params']['start_lr']
if 'grad_clip' in cfg['train']['params'].keys():
cfg['find_lr']['params']['grad_clip'] = cfg['train']['params'][
'grad_clip']
find_lr(args, cfg, train_ann, valid_ann)
elif args.mode == 'train':
train(args, cfg, train_ann, valid_ann)
elif args.mode == 'test':
test()
def main():
args = parse_args()
with open(args.config) as f:
cfg = yaml.load(f, Loader=yaml.FullLoader)
if args.num_workers > 0:
if 'transform' not in cfg.keys():
cfg['transform'] = {}
cfg['transform']['num_workers'] = args.num_workers
if args.mode != 'predict_kfold':
if args.backbone:
cfg['model']['params']['backbone'] = args.backbone
cfg['evaluation']['params']['save_checkpoint_dir'] = osp.join(cfg['evaluation']['params']['save_checkpoint_dir'], args.backbone)
if args.batch_size:
cfg['train']['batch_size'] = args.batch_size
cfg['evaluation']['params']['save_checkpoint_dir'] = osp.join(cfg['evaluation']['params']['save_checkpoint_dir'], f'bs{args.batch_size}')
if args.fold >= 0:
if args.mode == 'train':
cfg['dataset']['outer_only'] = True
cfg['dataset']['outer_fold'] = args.fold
cfg['evaluation']['params']['save_checkpoint_dir'] = osp.join(cfg['evaluation']['params']['save_checkpoint_dir'], f'fold{args.fold}')
cfg['seed'] = int('{}{}'.format(cfg['seed'], args.fold))
elif args.mode == 'test':
cfg['dataset']['outer_only'] = True
cfg['dataset']['outer_fold'] = args.fold
cfg['test']['savefile'] = osp.join(osp.dirname(cfg['test']['savefile']), f'fold{args.fold}', cfg['test']['savefile'].split('/')[-1])
if args.checkpoint:
cfg['test']['checkpoint'] = args.checkpoint
if args.load_previous:
cfg['train']['load_previous'] = args.load_previous
if args.eps:
cfg['optimizer']['params']['eps'] = args.eps
cfg = set_inference_batch_size(cfg)
# We will set all the seeds we can, in vain ...
set_reproducibility(cfg['seed'])
# Set GPU
if len(args.gpu) == 1:
torch.cuda.set_device(args.gpu[0])
if 'predict' in args.mode:
eval(args.mode)(args, cfg)
return
if cfg['transform']['augment'] == 'grid_mask':
assert 'grid_mask' in cfg['train']['params']
if 'mixup' not in cfg['train']['params'].keys():
cfg['train']['params']['mixup'] = None
if 'cutmix' not in cfg['train']['params'].keys():
cfg['train']['params']['cutmix'] = None
# Make directory to save checkpoints
if not osp.exists(cfg['evaluation']['params']['save_checkpoint_dir']):
os.makedirs(cfg['evaluation']['params']['save_checkpoint_dir'])
# Load in labels with CV splits
df = pd.read_csv(cfg['dataset']['csv_filename'])
if cfg['dataset'].pop('exclude_unknown', False):
df = df[df['diagnosis'] != 'unknown']
if cfg['dataset'].pop('isic2020', False):
df = df[df['isic'] == 2020]
ofold = cfg['dataset']['outer_fold']
ifold = cfg['dataset']['inner_fold']
train_df, valid_df, test_df = get_train_valid_test(cfg, df, ofold, ifold)
logger = create_logger(cfg, args.mode)
logger.info('Saving to {} ...'.format(cfg['evaluation']['params']['save_checkpoint_dir']))
if args.mode == 'find_lr':
cfg['optimizer']['params']['lr'] = cfg['find_lr']['params']['start_lr']
find_lr(args, cfg, train_df, valid_df)
elif args.mode == 'train':
train(args, cfg, train_df, valid_df)
elif args.mode == 'test':
test(args, cfg, train_df, test_df)