ruizhoud / distributionloss Goto Github PK

Hi Ruizhou,

Thanks for sharing your code.
I'm very impressed of your excellent work because it makes network tolerant of various types of optimizers and values of hyperparameters. I have a question about after the last layer whether there is Distrloss_layer or not.
In Cifar10 experiment, you use 7 Convolution layers(written as xC-xC-MP-2xC-2xC-MP-4xC-4xC-10C-GP) and I think the first layer is full precision convolution, whereas, the others are all binary convolutions including the last layer (10C). Is there a Distrloss_layer after the last binary convolution layer? I assume 10C-GP part as 10C-BN-Distrloss_layer-GP.

Can you commit model for cifar10 described in the paper xC-xC-MP-2xC-2xC-MP-4xC-4xC-10C-GP?
I have some difficulties to reproduce result of the papers.

Hello! The link of checkpoint is missed, could you upload it again? Thanks very much！

Hi Ruizhou,

This paper on weight regularization is quite interesting!

I have a question about the code here. After this line executed, all the weights are forced to be in [-1, 1]. If so, how could gradient mismatch be mitigated/solved?

Haichao

Hi,
BNN_DL is very nice paper, solving the fundamental problem of BNN.
However, I have some ploblem understanding Degeneration, Saturation and Gradient mismatch.

According to what I understand,
Intuitively,

1. Degeneration Loss makes mean of activation to zero.
2. Saturation Loss reduces the number of activation with A>1 and A<-1

Then, What is the role of Gradient Mismatch Loss?
It is difficult to understand the meaning of formula [ReLU(1-|u|-k*sigma)]^2.
I think the only way avoiding gradient missmatch problem is changing activation function. (BNN+, Self-binarizing network)

Could you explain in more detail?

Thanks.

Hi Ruizhou,

Thanks for sharing your code!

While going through your code, I found you used LeakyReLU for the first activation function and didn't quantize its output. Therefore it seems the second convolution layer takes full precision input instead of binary input. Previous works (i.e. XNOR-Net, DoReFa-Net) quantize the first activation as well.

Have you tried to quantize the first activation layer too?

Hi Ruizhou,

Thanks for sharing your code!
When I read your code, there are some problems that bother me. I cannot understand the codes for distribution loss, because they are inconsistent with the description in the paper. In addition, the hyper-parameter is also inconsistent with that in the paper.

Cloud you help me and make some explanations about the distribution loss in the code?

Thanks!

[distrloss_layer.py]
distrloss1 = (torch.min(2 - mean - std, 2 + mean - std).clamp(min=0) ** 2).mean() + ((std - 4).clamp(min=0) ** 2).mean()
distrloss2 = (mean ** 2 - std ** 2).clamp(min=0).mean()

According to the paper, it seems to be
distrloss1:

• Gradient Mismatch Loss: torch.min(2 - mean - std, 2 + mean - std).clamp(min=0) ** 2).mean()
• Saturation Loss: (std - 4).clamp(min=0) ** 2

Q1. Why 2 & 4 instead of 1 which is described in the paper?

distrloss2:
Degeneration Loss: (mean ** 2 - std ** 2).clamp(min=0)

Q2. Why not (torch.abs(mean) - std).clamp(min=0) ** 2?