https://hgpu.org/?p=14245
Towards Good Practices for Very Deep Two-Stream ConvNets