DCT结果貌似不太一样呀

[little-one]

刚刚使用dct_2d函数试了一下，结果和cv2的dct()函数结果出入较大。手动计算了DC系数，手算的结果和cv2.dct()是吻合的，但是与dct_2d()的DC系数相差较大，其它系数暂未验证。请问您遇到过这样的情况吗

[lilu1996]

我遇到了，所以我也问了，感觉是rfft到dct的过程出了问题？

[Ir1d]

The assert in README failed also. I get an error of 2e-5.

[Debatrix]

#13
Try this:
im_gray = cv2.imread('lena.jpg', cv2.IMREAD_GRAYSCALE)
a = cv2.dct(im_gray.astype("float"))
b = torch_dct.dct_2d(torch.from_numpy(im_gray.astype("float")), norm='ortho')
error = (torch.abs(torch.from_numpy(a) - b)).sum()

I got an error of 6.0581e-08.

[ChriisYoung]

跟matlab的dct2也差别很大

[mmmmayi]

和scipy.fftpack.dct计算的也不一样

[underdogliu]

I tried both options with program below. With torch 1.7.1 + torchaudio 0.7.2 (latest versions):

import torch
import torch_dct as dct
import scipy.fftpack

x = torch.randn(200)
X = dct.dct(x)   # DCT-II done through the last dimension
y = dct.idct(X)  # scaled DCT-III done through the last dimension
print(torch.abs(x-y))

XX = torch.Tensor(scipy.fftpack.dct(x.numpy()))
print(torch.abs(XX-X))

And according to the results it can be seen that most of absolute differences between this DCT implementation and scope's solution is adorn numerical range of [1e-6, 1e-7], occasionally 1e-5. I don't know if it's a good observation which can support the capability of this program. Although I have been used it for sometime as a DCT fixed module for my experiments, so far I did not observe anything weird.

我用上面的代码做了一个简单的测试，发现确实有一些出入，而且并非test模块assertion部分所声明的1e-10 而是大约在1e-6至1e-7之间。不过我用这个模块做了一段时间的实验，目前没有发现什么异常。

[underdogliu]

@zh217 So in general, to recap I think apart from updating modules itself we need to upgrade our criteria on testing module. Do you have any ideas on it? For example, maybe loosening the threshold from 10^-10 to a larger value?

[adaxidedakaonang]

torch-dct的结果除以1000以后与cv2的相符合，刚试了下

[underdogliu]

@adaxidedakaonang 是的，如果scale之后是符合的。所以我认为这个部分无伤大雅，可以loosen一些。@zh217 your thoughts?

[zhangzheng2242]

请问DCT是全局DCT还是在8×8的小块进行的

[adaxidedakaonang]

看你需求了。jpeg是选的8x8小块上做的 Adida ***@***.***> 于2022年7月12日周二 16:09写道：

…

请问DCT是全局DCT还是在8×8的小块进行的 — Reply to this email directly, view it on GitHub <#12 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AD47FES3ZKPYMRRJRDXY723VTUR27ANCNFSM4JIQUDPA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[kerryhhh]

torch-dct的结果除以1000以后与cv2的相符合，刚试了下
这个是什么意思？

[kerryhhh]

刚刚研究了下，好像误差大是因为pytorch默认的浮点数为float32？

[doomysday]

和scipy.fftpack.dct计算的也不一样

请问这个问题你现在解决了吗？

[doomysday]

跟matlab的dct2也差别很大

matlab的dct是和scipy.fftpack.dct计算的一样，请问您解决这个问题了吗

[adaxidedakaonang]

刚刚研究了下，好像误差大是因为pytorch默认的浮点数为float32？

好久没用了，不知道代码有无更新。当时作者代码的计算结果和cv的计算结果相差一个固定系数，为1000。所以对pytorch dct的结果进行与1000的相除即可与cv一致。这是当时的意思。

[yyeh26]

I tried to set the default type to float64 and the assertion passed.

`import torch
import torch_dct as dct

// This line is crucial for accuracy
torch.set_default_dtype(torch.float64)

x = torch.randn(500)
X = dct.dct(x) # DCT-II done through the last dimension
y = dct.idct(X) # scaled DCT-III done through the last dimension
assert (torch.abs(x - y)).sum() < 1e-10, (torch.abs(x - y)).sum() # x == y within numerical tolerance`

The error margin is around 1e-13.

[yyyoite]

为什么跑出来和cv2.dct完全不一样？

[JJy764]

为什么跑出来和cv2.dct完全不一样？

我也是，可视化结果不一样

[TimeFlysLeo]

与cv2 dct变换后结果差异确实很大，但是逆变换回图像后结果差不多