eghouti/binaryconnect

Implementation of BinaryConnect on Pytorch

binaryconnect's Issues

binary weights?

Hi, thanks for your implementation of binary connect.

I don't know why but after running your project and readout all the weigths, it seems like they are not binary weights.

for example the kernel filter
[ 0.4462, -0.4897, 0.3968, -0.8253],
[ 0.8852, 0.3383, 0.6267, -0.2224],
[ 0.2346, -0.4480, -0.5962, -0.7025]]],

is there something wrong with my turning ?

thanks.

Same kernel filter at the 1st epoch and the final epoch

Hi there,

thanks for your implementation of binary connect.

I got a weird problem.
I use your binary connect to train a model based on the MNIST. however, although I can get a model with an accuracy of over 90%. I find the kernel filters are quite random which are not apparently like edge detection. in addition, the kernel filter trained after 1st epoch and final epoch is almost identical, which means the convolutional kernel filters are not updated during training.

Do you know why that happens?

thanks.

Comparison between ".data.clone()" and ".data.copy_()"

Thank you for your code!

I have a question about LINE 26 and LINE33 in the binaryconnect.py. I wonder why you choose two different functions ".data.clone()" and ".data.copy_()" in these lines?

I guess the "clone()" will create a new tensor with new address, while "copy_()" is a in-place operation, am I right?

I also wonder the Version of Pytorch that you implemented in this code.

Thank you again!

Loss criterion

Hello there,

I see that your code uses the CrossEntropy loss for BC, whereas their original paper used instead an L2-SVM with a squared hinge loss. When trying to do my own replica, I fail to obtain their performance by using the squared hinge loss, but manage just fine with CE. My question is: Have you tried to replicate their results using an L2SVM, and if so, did you succeed?

Thanks!

AttributeError: 'BC' object has no attribute 'num_of_params'

Hi, thanks for the code.

I think the title is quite self explanatory. The attribute 'num_of_parameters' only appears in the clip function of the BC-class. Hence I get a syntax error when trying to run the training script.