[feature request] allow checkpoint to save partial network

[vincentzlt]

It would be nice if user can define the weights of partial network that can be saved by checkpoint.

Possible user cases could be GAN network or BERT with various downstream finetune tasks.

[ShawnXuan]

Currently, OneFlow save weights in separate subfolders, and can be removed separately manually. Also we will consider this request to improve the experience of using oneflow checkpoint function.

BTW, OneFlow will try to load weights from user specified checkpoint/model folder, if not exists, weights will be initialized on demand. So, for BERT downstream tasks, we can set pretrain model as the model_load_dir, oneflow will load avaiable weights and init missing weights automatically with Waring for information. see more information in https://github.com/Oneflow-Inc/OneFlow-Benchmark/tree/master/LanguageModeling/BERT.