Input Quantization

[abhivandit]

Hi,

I was looking at the networks present in binary models directory and also at their usage and implementation in image_classification.py. I would like to know where exactly is the input getting quantized. None of the networks seem to use QActivation() function. Only the weights seem to get quantized.

Thanks

[Jopyth]

Hi,

So far we only correctly converted the following models (the others were mostly copied from gluon, but not correctly adapt, as you noted - we should probably (re-)move them):

• resnet.py
• resnet_e.py
• densenet.py
• meliusnet.py
• naivenet.py

These models are binarized (both weights and activations) with the nn.activated_conv function, e.g. here. This function adds both QAct and QConv layers, by adding e.g. a BinaryConvolution.

If you want to train one of the other models as a binary one (which we have not done so far), you would currently need to adapt them accordingly (replace the QConv with nn.activated_conv).