Weighted Resize CRASH

[Dan64]

Hello Selur,

  I found a movie that CRASH the encoding when I apply the weighted resize.
  I attached an archive with all the necessary information.

Thanks,
Dan

[Selur]

clip = core.std.Merge(clipa=clipSmoothed, clipb=clip, weight=0.4)
                                             ^
SyntaxError: invalid non-printable character U+00A0

another non-printable character

-> I'll compile a new version and send you links once I uploaded everything.

Cu Selur

[Selur]

Nope, removing that character does not fix the problem.
-> looking into it

My ugess is that it's related to TRT, since when I call the call inside a Windows Command Prompt now, I get:

Warning: i:\Hybrid\64bit\Vapoursynth\Lib\site-packages\torch_tensorrt\fx\tracer\acc_tracer\acc_tracer.py:584: UserWarning: acc_tracer does not support currently support models for training. Calling eval on model before tracing.
  warnings.warn(

Information: == Log pass <function fuse_permute_matmul at 0x000001582EA78670> before/after graph to C:\Users\Selur\AppData\Local\Temp\tmpku34hqog, before/after are the same = True
Information: == Log pass <function fuse_permute_linear at 0x000001582EA78430> before/after graph to C:\Users\Selur\AppData\Local\Temp\tmpc3o7u8pf, before/after are the same = True
Information: Now lowering submodule _run_on_acc_0
Information: split_name=_run_on_acc_0, input_specs=[InputTensorSpec(shape=torch.Size([1, 3, 308, 720]), dtype=torch.float16, device=device(type='cuda', index=0), shape_ranges=[], has_batch_dim=True)]
Information: Timing cache is used!
x265 [error]: unable to open input file <->
Information: TRT INetwork construction elapsed time: 0:00:00.400953
Information: Build TRT engine elapsed time: 0:01:01.174254
Information: Lowering submodule _run_on_acc_0 elapsed time 0:01:04.743047
Information: Now lowering submodule _run_on_acc_2
Information: split_name=_run_on_acc_2, input_specs=[InputTensorSpec(shape=torch.Size([1, 3, 1232, 2880]), dtype=torch.float16, device=device(type='cuda', index=0), shape_ranges=[], has_batch_dim=True), InputTensorSpec(shape=torch.Size([1, 3, 1232, 2880]), dtype=torch.float16, device=device(type='cuda', index=0), shape_ranges=[], has_batch_dim=True)]
Information: Timing cache is used!
Information: TRT INetwork construction elapsed time: 0:00:00.000999
Information: Build TRT engine elapsed time: 0:00:00.222227
Information: Lowering submodule _run_on_acc_2 elapsed time 0:00:00.226829
Error: fwrite() call failed when writing frame: 0, plane: 0, errno: 22
Output 33 frames in 0.81 seconds (40.53 fps)

Encoding works fine with trt disabled.

Issue is unrelated to weighted resize, also happens without it.

Cu Selur

[Dan64]

I'm able to encode the movie using the same script in vsViewer.exe...

[Selur]

I know, but that doesn't help at all.
https://github.com/HolyWu/vs-realesrgan/issues/21

Cu Selur

[Selur]

As soon as trt is enabled it crashes.
(Independent of the model, fp16, num_streams or the encoder used)
Probably an issue with one of the dlls in the torch_dependencies folder.

[Dan64]

I tried to run the encoding manually, and I get the following error



If I remove with an Editor the 2 non-printable characters (I had to remove all the spaces) the encoding worked.

Maybe vsViewer is able to remove all the non printable characters. These characters are visible in Notepadd++ if I select the Encoding ANSI (actually the file is using a UTF-8-BOM encoding).

Dan

[Selur]

I removed those characters too, in Hybrid, but I still got crashes.

[Selur]

I'll send you a link to my current dev (~15min) which should fix the non-printable character, maybe that works for you.

Cu Selur

[Dan64]

I had the same problem. I was not able to really remove these characters in Notepad++. The only way to not get the crash is to change the script as following

clip = core.std.Merge(clipa=clipSmoothed,clipb=clip,weight=0.4)