Support on Multi-GPU？ about yolact HOT 4 CLOSED

To train with multiple GPUs, you should turn off JIT (the way I implemented it doesn't play well with multiple GPUs): add --no_jit to the commandline arguments. Otherwise, you'll get the performance degradation you mentioned because it will retrace the backbone every iteration.

Also, if your batch size is less than 8 per GPU, you should also freeze batch norm by adding 'freeze_bn': True to the config you use in data/config.py (e.g., ctrl+f yolact_base_config and add it to the dict there).

The DataParallel instance is in train.py btw.

from yolact.

Hi,

So what are the changes one needs to make to make it train on multiple GPUs?
Any lead will be highly appreciated.

Thanks. :)

from yolact.

See #8 for all of the changes. One of these days I'll add the necessary changes as a command-line argument.

from yolact.

For future reference: I removed --no_jit and now it'll just turn JIT off if it detects more than one GPU.

from yolact.