Closed hancy16 closed 4 years ago
I think it should be effective as long as there is a convolutional layer before the first Relu nonlinearity. You can try setting the centers to positive values and see what happens.
Maybe I did not express myself very well. In the end of the encoder there is one relu nonlinearity, so feature maps fed into quantizer are non-negatives. Thus only three levels of quantization exist(this can be verified in the tensorboard). In this case the actual entropy should be smaller than the H/16W/16log(5)*C upper bound by a non-ignorable margin. But the authors claim that this upper bound is tight. This is confusing. Look forward to your favourable reply.
Sorry, I originally thought you were referring to the decoder...
Good point, I believe you are correct. I think the relu should be omitted in the final layer of the encoder for the bound to make sense! Good find.
In current implementation the "quantizer" quantizes the feature maps into 5 centers {-2,-1,0,1,2}. However, the encoder adopts relu as activation, after which negatives can never be achieved. So { -2,-1 } become useless here. Is there anything wrong with my understanding?