[Enhancement] Check and sync the num_classes in Config class #2477
Conversation
juncaipeng
commented
Aug 23, 2022
•
juncaipeng
commented
- Check and sync the num_classes in Config class
- Update the docs of quick_start demo
| train( | ||
| cfg.model, |
There was a problem hiding this comment.
test/train/eval 都需要增加sync_infor确认信息同步。
There was a problem hiding this comment.
我看了medicalseg中所有配置文件都提供了num_classes，而且medicalseg是使用自己定义的Config类，所以就没大改这里面的。
| if len(num_classes_set) == 0: | ||
| raise ValueError( | ||
| '`num_classes` is not found. Please set it in model, train_dataset or val_dataset' | ||
| ) |
There was a problem hiding this comment.
在matting里不存在num_classes概念，是否能正常运行呢
There was a problem hiding this comment.
matting中之前也没有调用config_check函数，所以现在修改后，matting下面的cfg示例没调用check_sync_info函数，所以不影响matting的代码。
| if hasattr(self.val_dataset_class, 'NUM_CLASSES'): | ||
| num_classes_set.add(self.val_dataset_class.NUM_CLASSES) | ||
| elif 'num_classes' in self.val_dataset_config: | ||
| num_classes_set.add(self.val_dataset_config['num_classes']) |
There was a problem hiding this comment.
这里是否存在必须构建val_dataset或者train_dataset呢，如果存在的话，执行predict.py时当dataset_root不存在能否正常运行呢？原来是报FileNotFoundError时给个警告而非报错
There was a problem hiding this comment.
这里只考虑config文件中信息的同步，如果train_dataset和val_dataset构建不成功，那就会另外报错啊
There was a problem hiding this comment.
predict.py不需要构建val_dataset或train_dataset就应正常预测，如果会报错的话，不符合预期。
已经全部检查了 |
