def cross_fold_validation(cfg, result_dir, best_model=True):
model_file = cfg['model'] + '_best.model' if best_model else cfg[
'model'] + '_final.model'
# run five fold validation
model = get_model(cfg['model'], cfg['data']['channel'], 1).to(device)
saved_model = torch.load(os.path.join(result_dir, model_file),
map_location=device)
try:
model.load_state_dict(
saved_model['model_state_dict']) # maybe need to change
except RuntimeError:
print('convert model layer name!!!')
model_state_dict = convert_state_dict(saved_model['model_state_dict'])
model.load_state_dict(model_state_dict)
test_result = evaluation(model,
data_root=cfg['data']['root'],
csv_file=cfg['data']['csv_file'])
test_csv = pd.DataFrame({
'image_name': test_result['image_name'],
'target': test_result['target']
})
test_csv.to_csv(os.path.join(result_dir, 'submission.csv'), index=False)
def load_saved_model(self, fname, train=True):
self.print_to_log_file("loading checkpoint", fname, "train=", train)
saved_model = torch.load(fname, map_location=self.device)
# load model state dict
try:
self.model.load_state_dict(
saved_model['model_state_dict']) # maybe need to change
except RuntimeError:
model_state_dict = convert_state_dict(
saved_model['model_state_dict'])
self.model.load_state_dict(model_state_dict)
# load optimizer and lr_scheduler
if train:
optimizer_state_dict = saved_model['optimizer_state_dict']
if optimizer_state_dict is not None:
self.optimizer.load_state_dict(optimizer_state_dict)
if self.lr_scheduler is not None and isinstance(
self.lr_scheduler, optim.lr_scheduler.ReduceLROnPlateau):
self.lr_scheduler.load_state_dict(
saved_model['lr_scheduler_state_dict'])
# load other related training results
self.epoch = saved_model['epoch']
self.monitor_MA = saved_model['monitor_MA']
self.val_eval_criterion_MA = saved_model['val_eval_criterion_MA']
self.best_val_eval_criterion_MA = saved_model[
'best_val_eval_criterion_MA']
self.best_epoch_based_on_monitor_MA = saved_model[
'best_epoch_based_on_monitor_MA']
self.train_losses, self.train_eval_metrics, self.val_losses, self.val_eval_metrics = saved_model[
'plot_stuff']
def inference_csv(cfg, result_dir, device, best_model=True):
test_aug_parameters = OrderedDict({
'affine': None,
'hflip': None,
'vflip': None,
'color_trans': None,
'hair_mask': None,
'rotate': None,
'microscope': None,
'normalization': {
'mean': (0.485, 0.456, 0.406),
'std': (0.229, 0.224, 0.225)
},
'size': 384,
'scale': (1.0, 1.0),
'ratio': (1.0, 1.0)
})
model_file = cfg['model'] + '_best.model' if best_model else cfg[
'model'] + '_final.model'
# run five fold validation
cfg['data']['root'] = '/home/zyi/My_disk/ISIC_2020/data/test'
cfg['data']['csv_file'] = '/home/zyi/My_disk/ISIC_2020/test.csv'
model = get_model(cfg['model'], cfg['data']['channel'], 1).to(device)
saved_model = torch.load(os.path.join(result_dir, model_file),
map_location=device)
try:
model.load_state_dict(
saved_model['model_state_dict']) # maybe need to change
except RuntimeError:
print('convert model layer name!!!')
model_state_dict = convert_state_dict(saved_model['model_state_dict'])
model.load_state_dict(model_state_dict)
test_result = evaluation(model,
data_root=cfg['data']['root'],
csv_file=cfg['data']['csv_file'],
device=device,
aug_parameters=test_aug_parameters)
test_csv = pd.DataFrame({
'image_name': list(test_result['image_name']),
'target': list(test_result['target'])
})
test_csv.to_csv(os.path.join(result_dir, 'submission.csv'), index=False)