karenullrich / pytorch-binary-converter Goto Github PK

Hi,

when sending a tensor from a Relu activation function to float2bit, f can take the value 0. Then, the log2 will return -inf with all the complications it ensues (out of bound error for the gather function).

One quick fix is to add a small constant to the tensor f such that it won't change the e_scientific value.

Thanks for this library, been very helpful ! :)

After I add float2bit and bit2float to the network, the weight of the convolution model is always 0 when back propagating，I checked that the variables in the code are in cuda and required_grid=true, but the gradient seems to disappear.Can anyone tell me why this problem is happening?

I found a BUG in your code. The float2bit function can't convert a number bigger than 512. Hopefully, someone can fix it ASAP. By the way, the code is very helpful.