search

HobbitLong / CMC

[arXiv 2019] "Contrastive Multiview Coding", also contains implementations for MoCo and InstDis
BSD 2-Clause "Simplified" License
1.3k stars 179 forks source link

Train/Test split when training linear classifier (ImageNet) #53

Open jnyjxn opened 4 years ago

jnyjxn commented 4 years ago

Hello,

Thank you for this great work and well-written paper.

When training the linear classifier (e.g. Alexnet with ImageNet), how do you perform the train/test split? 1. Do you use all of the classes in your test set?

  1. How do you choose views for each sample?

Many thanks, Jonny

HobbitLong commented 4 years ago

Hi Jonny,

  1. ImageNet dataset itself has a "train" set and a "val" set, which will be used as train and test sets, respectively.
  2. The same as the CMC pre-training stage. If you train CMC with Y/DbDr split, then you use Y/DbDr split in linear evaluation stage as well.
jnyjxn commented 4 years ago

Thanks @HobbitLong - unfortunately from what I can tell ImageNet is defunct (I've had no response for > 3 months) so I can't access the dataset.

Could you briefly describe the approach they've used to do the split? Specifically, do they split according to e.g. 80 different chairs in train, 20 different chairs in val, or is it the same 100 chairs in train and val but with different perspectives?

Thank you!

HobbitLong commented 4 years ago

I believe it's the former way.