The enlarged result is different from waifu2x-caffe. about waifu2x-ncnn-vulkan

[nihui]

https://i.imgur.com/SFs1Ws8.gif

If you look closely and compare it, the contrast of the image seems different.
Will the issue be solved if the bug related to fp16a is fixed?

よく見比べてみると画像のコントラストが異なるようです。
fp16aに関する不具合が修正されるとこの問題は解決しますか？

yeah, I found the little difference too.
It is not because of fp16s or fp16a, but because of postprocess.
waifu2x-caffe add a small eps value to all pixels during postprocess, while waifu2x-ncnn-vulkan do not.
I may port this behavior to let the result pixel-wise identity to the caffe one.

[nihui]

20190414 package updated, please try the new binary
the result is still not pixel-wise identity to the caffe output due to loss-of-precision during floating point arithmetic and fp16 storage conversion
but it is close enough I think ^^:)

[f11894]

Yes, ver 20190414 is almost the same as waif2x-caffe.
You will not notice the difference unless you compare them fairly carefully.

はい、ver 20190414ではwaif2x-caffeと殆ど変わらなくなりました。
かなり注意深く比較しない限り違いに気づくことはありません。

Ryzen 2400G

waifu2x-caffe ver 1.2.0.2
waifu2x-caffe-cui.exe -i input.png -t 0 --crop_size 128 -m scale --scale_ratio 2 --model_dir models\CUnet -p cpu -o output.png

waifu2x-ncnn-vulkan ver 20190407 and 20190408
waifu2x.exe input.png output.png -1 2

waifu2x-ncnn-vulkan ver 20190414
waifu2x.exe input.png output.png -1 2 300

	PSNR(RGB)	SSIM(RGB)	Average time(s)
waifu2x-caffe	37.619	0.987	20.250
waifu2x-ncnn-vulkan(20190407)	34.469	0.975	8.557
waifu2x-ncnn-vulkan(20190408)	36.942	0.986	8.664
waifu2x-ncnn-vulkan(20190414)	37.334	0.986	3.641

[Dearmw]

Input

Running code waifu2x 1.jpg 2.jpg 2 2
[0 Radeon RX Vega] queueC=1 queueT=2 memU=2 memDL=0 memHV=1
[0 Radeon RX Vega] fp16s=1 fp16a=1 int8s=1 int8a=1

Output

[f11894]

In addition, Benchmark score seems to be worse than waifu2x-caffe.
Average score of 55 images.

waifu2x-caffe RGB
35.492dB

waifu2x-caffe CUnet
37.619dB

waifu2x-ncnn-vulkan CUnet
34.469dB

This is the bat file I used for benchmarking.
https://github.com/f11894/SR_benchmark

I use Google Translate because I am not good at English.

[nihui]

Input

Running code waifu2x 1.jpg 2.jpg 2 2
[0 Radeon RX Vega] queueC=1 queueT=2 memU=2 memDL=0 memHV=1
[0 Radeon RX Vega] fp16s=1 fp16a=1 int8s=1 int8a=1

Output

the incorrect output caused by fp16 arithmetic on gtx20/rtx20/vega issue is confirmed. A workaround will be applied for these gpu soon...

[nihui]

thanks for your report
will investigate ...

[Dearmw]

感谢您的报告
将进行调查......
Are you from Shanghai?

[nihui]

deconvolution operator produce wrong result on fp16 storage/arithmetic mode.
I will fix it soon ...

[nihui]

the underlying ncnn library get fixed in Tencent/ncnn@9643916
new binary will be rebuild and uploaded about an hour later, stay tuned ...

[nihui]

20190407 package updated, please try the new binary

[f11894]

There is no change in the output result using the new binary.
MD5 is also consistent with the previous.

[nihui]

There is no change in the output result using the new binary.
MD5 is also consistent with the previous.

hi, I just tried the new binary with your original image on AMD r9-m290x machine and my Intel UHD 620 laptop, the output image looks the same with caffe one.

waifu2x.exe
SHA1: 6892F871041F4DAE1BF74FF9692547317A99DFBD

[f11894]

I'm sorry I didn't have enough words.
MD5 is from the output PNG image, not the executable file.

When I converted the sample described in README.md, I noticed that I could not get the same result as you.
Apparently there are things that do not work well depending on the environment.

[nihui]

I'm sorry I didn't have enough words.
MD5 is from the output PNG image, not the executable file.

When I converted the sample described in README.md, I noticed that I could not get the same result as you.
Apparently there are things that do not work well depending on the environment.

another issue which affects modern gpu with fp16a capability is confirmed.
because of lack of these specific devices, I may try to figure it out with some workarounds.

[nihui]

new binary 20190408 has been uploaded with a workaround for gtx20 and vega series.
please try this new version and see if this issue still exists.

[kalcohol]

@nihui @f11894 @Dearmw
i tested version 20190408, it seems this issue was fixed @rtx2060 and RTX2080.

[f11894]

Even in my environment, I confirmed that this problem was solved.
The output result is equivalent to waifu2x-caffe.

[f11894]

In the case of visual inspection, I do not feel much difference, but there seems to be a little difference in the score of Benchmark.

waifu2x-caffe RGB
35.492dB

waifu2x-caffe CUnet
37.619dB

waifu2x-ncnn-vulkan(20190408) CUnet
36.942dB

[f11894]

https://i.imgur.com/SFs1Ws8.gif

If you look closely and compare it, the contrast of the image seems different.
Will the issue be solved if the bug related to fp16a is fixed?

よく見比べてみると画像のコントラストが異なるようです。
fp16aに関する不具合が修正されるとこの問題は解決しますか？