wuzhiyang2016

> @wuzhiyang2016 I've got the same problem. I added the ModifiedVGG16Model class in my test script, with no success. Do you still have the same issue? like the author, when...

pytorch version caused the problem, you can change to this: values = \ torch.sum((activation * grad), dim=0, keepdim = True). \ sum(dim=2, keepdim = True).sum(dim=3,keepdim = True)[0, :, 0, 0].data

> @wuzhiyang2016 I used keepdim=True, but I still get errors at the same point. > > RuntimeError: The size of tensor a (14) must match the size of tensor b...

yes, the operation of merge_quantize_weights() is out of calculate graph, It has no effect on the gradient of model parameters，，,assume a situation like this: last layer's output y, the current...

i am very glad to hear from you~ what i mean is when calculate gradient, we should use wq instead of w. i just check the code of merge_quantize_weights() function...

ok, so, jacinto-ai-devkit's implement code doesn't replace w by wq in the backward. but section3.2, the paper said : "The updated weights are quantized and used for subsequent **forward** and...

2018~2019 i participated in reimplemention of google's quantize training paper by pytorch in my working company, we implemented it strictly follow the formula besides with forward and backword. our precision...

ok, i will try that, how to dealwith dataLayer's clip layer ? if a sturct : dataLayer + conv , it will convert to dataLayer + clip +conv, this clip...

very thanks , i'll try that

the folder name is tidl_j7_01_00_00_00 @mathmanu