search

ssnl / dataset-distillation

Open-source code for paper "Dataset Distillation"
https://ssnl.github.io/dataset_distillation
MIT License
778 stars 115 forks source link

Cannot find definition of `local_n_nets` #32

Closed XinDongol closed 4 years ago

XinDongol commented 4 years ago

state.local_n_nets is used without definition. What's the meaning of it?

If train_nets_type=='known_init', why we want to build several networks with different initialization?

https://github.com/SsnL/dataset-distillation/blob/24e6958cf23bc98de20a7b2714fc76fa7d38b874/main.py#L90-L91

Thanks a lot!

XinDongol commented 4 years ago

Any Update?

ssnl commented 4 years ago

It is set here https://github.com/SsnL/dataset-distillation/blob/24e6958cf23bc98de20a7b2714fc76fa7d38b874/base_options.py#L444

It means the number of training models in a single rank of distributed training. In case of no distributed training, it equals the total number of training models.

With known_init, n_nets controls the number of known initializations, i.e., whether you want to overfit for one model, or for multiple models.