instance-balanced Training config file for Places365
jiequancui opened this issue · 5 comments
Hello,
is there instance-balanced training config file for Places365 ?
Sorry, it's not included in the current version. But one can easily create one by modifying from ImageNet_LT or iNaturalist. Note that for instance-balanced training on Places365, we only train the last block, which means one should modify the code correspondingly by adding condition here https://github.com/facebookresearch/classifier-balancing/blob/master/run_networks.py#L119.
Big thanks for your reply.
@bingykang I'm wondering how many epochs you're using to train Places-LT. Is it 30 epochs as in OLTR, or 90 epochs as described in your paper?
Note that for instance-balanced training on Places365, we only train the last block, which means one should modify the code correspondingly by adding condition here https://github.com/facebookresearch/classifier-balancing/blob/master/run_networks.py#L119.
In addition, by only train the last block, do you mean in stage 1 (feature training stage) you freeze everything before last BottleNeck or layer4?
Very interesting work. Thank you !