The python version should be 3.x ? about gabor_cnn_pytorch HOT 7 CLOSED

@gzhcv Thanks for you kind advise. Since I am quite busy these days, I will fix this issue later.

from gabor_cnn_pytorch.

@jxgu1016 Thanks! I just no sure if this is a issue, because i visualized the filter in GoFs and find it seems each filter is different with others.

• For the filters of first layer, in the beginning of training stage, the filters obtained by modulating a learned convolutional filter using different orientation Gabor filters are different with each other (see the following figure).

But after training for 0.x million steps , the filters are quite similar to each other (see the following figure), but actually the values are different.

• For the filters of second layer, the filters are different to each other even training for 0.x million steps.

So, it seems this is not a issue. I just not very clear about how the GConv.py is used for build the Python package.

PS: All the experiments are performed with python 2.x, because the python 3.x seems cannot build
successfully.

from gabor_cnn_pytorch.

Thanks for your detailed report, and I will check it later.

from gabor_cnn_pytorch.

@gzhcv My python version is 2.7. There should be some 0 in the filters. It seems not to be a issue. Thx for your report anyway.

from gabor_cnn_pytorch.

@jxgu1016 The parameter theta represents different orientation of Gabor filter, if the GoF is obtained by the code in GConv.py (since theta is always be 0 that computed in line 107 of GConv.py), why the learned filters are different with each other? I'm confused about this.

from gabor_cnn_pytorch.

@gzhcv you may miss the codes on the 1st line: from __future__ import division.

from gabor_cnn_pytorch.

@jxgu1016 Oh.... Sorry for my carelessness.

from gabor_cnn_pytorch.