Training SLOWFAST_16x8_R101_50_50

Hi, I have a question about training SLOWFAST_16x8_R101_50_50 model. If I understand correctly, this model requires a total input of 80 frames, 64 for fast branch and 16 for slow branch. So training this model consumes lots of GPU memory. For me, I can only use batch_size=3 per GPU (I use V100 with 16GB memory). However, I see in the [yaml](https://github.com/facebookresearch/SlowFast/blob/master/configs/Kinetics/c2/SLOWFAST_16x8_R101_50_50.yaml) file that you still use batch_size=8. Can you tell me how you accomplish this? And please correct me if my understanding is wrong. Thank you.