Improper equals sign casues serious problem

[Stephen0808]

In the function of generate_filter, the code should be revised as “return [[0. if i + j > end or i + j < start else 1. for j in range(size)] for i in range(size)]”. In the original version, the code is “return [[0. if i + j > end or i + j <= start else 1. for j in range(size)] for i in range(size)]”, which will alter the weight of point(0,0) as 0. This error casues the dc component disapear in the filter matrix.

[Stephen0808]

I supppose this is the reason why the performance of the model decreases in the test dataset.

[Stephen0808]

Besides, the cut-off frequency setting in the generate_filter is not resonable. 1/16 of the length of the spectrum is unequal to the code"(row + col) == size // 16". Obviously, the condition of cut-off frequency is ((cut_f^2)/2) // (size^2) = 1 / 16.(Using the square ratio). So the cut_f_16 and cut_f_8 should be size/(8^(0.5)), size/(4^(0.5)). Plz pay attention to this!

[yyk-wew]

Thanks for your advices! It makes sense to me. I'll run some experiments recently and update the results here.

[yyk-wew]

Hi @Stephen0808 , I've fixed bugs as you mentioned above and rerun some experiments. With dc component, using FAD branch only, the result comes to 91.0 on validation and 89.9 on test. Slightly better than before but still not good enough. Besides, it's observed that models almost always perform better on validation than test during training. I suppose that the dataset partition has something to do with this.

[Stephen0808]

Thanks for your detailed experiments and reply. Frankly speaking, there are some unknown module parameters in the paper which cause the incomformity in the reproduction. It is a pity that the comments will remain to the day the author opens the source code. But fortunately, some latter papers also cited this paper's module, FAD as the prime structure. I will delve into this problem for anthor task in the following days, and share some clues for this paper's reproduction in this issue.

[yyk-wew]

I will delve into this problem for anthor task in the following days, and share some clues for this paper's reproduction in this issue.

Look forward to it :)