Junhui Wu

fc_out, relu, top_k_op

2

Hi, I think the last fc layer may should not use an activation unit. If the output of fc layer is all negative, after the ReLU unit, it would be...

Is it better to train from scratch on Kinetics-600?

36

Hi, I wonder why you only release checkpoint on Kinetics-600 trained from scratch but not from ImageNet pre-trained parameters. In the paper, performance is better with ImageNet pre-training on Kinetics-400...

What regularizers do you use when training?

1

I noticed that in Conv3D or BatchNorm modules, the default regularizers are None. Do you use regularizers in Conv3d or BatchNorm when training? If so, do you use L1 or...

about batch size

4

Hi, In the paper _"Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset"_, it is said you used 64 GPUs to train 3D ConvNets. But in readme, you...