batch_size>1 is not working in the testing phase

[yuzehui1996]

Hi，
Have you ever tried muti-gpus training? I simply add DataParallel but the AP and AR are lower than the training with single gpu.

Thanks!

[yizhou-wang]

Sorry, we didn't try multiple GPUs before. But the performance should not downgrade using multiple GPUs. You may need to tune some parameters or monitor the loss during the training.

[yuzehui1996]

Thanks for your reply!
I have trained the houglass model(HG) with multi-gpus. When I test the checkpoints and set batch-size from 1 to 4, the AP and AR downgrade greatly. Will the accuracy decrease due to the BatchNormalization layer during multi-gpus training?
BTW, I use 4 gpu cards for training.

[yizhou-wang]

I double-checked the inference code. Seems the batch size is hardcoded as 1. We will fix this bug in the future. But for now, you can just use batch_size = 1 in the testing phase.