+- Selur's Little Message Board (https://forum.selur.net)
+-- Forum: Hybrid - Support (https://forum.selur.net/forum-1.html)
+--- Forum: Problems & Questions (https://forum.selur.net/forum-3.html)
+--- Thread: Weighted Resize CRASH (/thread-3062.html)
RE: Weighted Resize CRASH - Selur - 03.02.2023
No, the problem I reported way related to tiling, but it caused HolyWu to look into the general functionality and compatibility of the filter, which caused him to decide that there are issues with RT.
It does not matter whether I think RT is broken or not, since I'm not the developer of the filter.
Cu Selur
RE: Weighted Resize CRASH - andry62 - 03.02.2023
Unfortunately using FeMaSR without TRT support
almost impossible like most SR filters due to
processing speed tending to zero.
Very very sorry.
RE: Weighted Resize CRASH - Dan64 - 05.02.2023
Hello Selur,
I wrote to HolyWu as you suggested: https://github.com/HolyWu/vs-femasr/issues/3
HolyWu accepted to keep support to Tensor RT and today released a new version.
Dan
RE: Weighted Resize CRASH - Selur - 05.02.2023
Nice. Just got a notification of the new release a few minutes ago.
Cu Selur
RE: Weighted Resize CRASH - Dan64 - 05.02.2023
Hello Selur,
I installed the last dev version and I found some problems in using FeMASR.
I attached 2 archives showing different problems:
1) Test1
In this case, it seems that LWLibavSource changed inputs.
If you open the preview with vsViewer on "Test-001_v1.vpy"
Will be issued the following error:
clip = core.lsmas.LWLibavSource(source="E:/VideoTest/FeMaSR/Test1/Test-001.mkv", format="YUV420P8", stream_index=0, cache=0, fpsnum=24,repeat=true, prefer_hw=0)
NameError: name 'true' is not defined
the problem has been fixed in the script: "Test-001_v1_fixed.vpy". Also the problem can be solved using FFMS as in the script "Test-001_v3.vpy"
After having fixed this problem is shown another bigger problem regarding FeMaSR. It seems that the problem is related to Tensor RT.
I get the same problems even if I use the previous version of Hybrid (2023.01.31.1) and previous version of FeMaSR (v1.0)
2) Test2
In this case it seems that Hybrid is not able to recognize the right format conversion.
If you open the preview with vsViewer on "Test-002_fp16.vpy"
Will be issued the following error:
Failed to evaluate the script:
Python exception: Merge: only 8-16 bit integer and 32 bit float input supported, passed RGBH
the problem has been fixed in the script: "Test-002_fp16_fixed.vpy"
RE: Weighted Resize CRASH - Selur - 05.02.2023
No Debug output.
=> will look at it later
RE: Weighted Resize CRASH - Selur - 05.02.2023
First issue doesn't seem to be related to the torchAddon itself. Seems like a typo where 'repeat=true' is used, when 'repeat=True' should be used.
Second issue looks like some color space mixup.
RE: Weighted Resize CRASH - Dan64 - 05.02.2023
Hello Selur,
sorry for the missing debug outputs, I attached them to this post.
Regarding the problem on Tensor RT described in Test1 it seems that is related to the frame dimension (it is too big).
If I divide by 2 the size of the frame by using the option: Vapoursynth->Misc->Lower res. before resize, Tensor RT is working.
Dan
P.S.
It seems that the new RT version is not discovered. The name of the cache files have the previous version embedded for example: FeMaSR_SRX2_model_g.pth_NVIDIA GeForce RTX 3060_trt-8.5.2.2_640x288_fp16_workspace-33554432
RE: Weighted Resize CRASH - Selur - 05.02.2023
I also can't use 'trt_min_subgraph_size'.
=> Something is broken with the setup or the filter.
RE: Weighted Resize CRASH - Selur - 05.02.2023
Did you use the updated download? (first download had the old tensor version)
Does "Vapoursynth\Lib\site-packages\tensorrt-8.5.3.1.dist-info" in your setup?