mobilenet-torch

Torch implementation of MobileNet (https://arxiv.org/abs/1704.04861), implemented by Jaewoo Song.

CUDNN and NN versions

Why two versions?

• MobileNet uses depthwise separable convolutions. They are implemented differently in Torch.
• Only nn module has function named nn.SpatialDepthWiseConvolution().
• On the other hand, cudnn module's function cudnn.SpatialConvolution() has one more parameter than its equivalent in nn. By setting nInputPlane, the very last parameter, one can implement depthwise separable convolutions.

So, any difference?

• It is much faster to train the cudnn version. Training the nn version is really slow.
• But under the circumstances without GPUs, nn version is the only choice.