Failed to parse script: Python exception: operator(): no valid optimization profile found

[Yuvraj Seegolam]

Apologies for creating another issue, but I'm not understanding why my realesrgan engine is only working with landscape aspect ratio videos.

I compiled the engine using this command:

trtexec --fp16 --onnx=model.onnx --minShapes=input:1x3x8x8 --optShapes=input:1x3x720x1280 --maxShapes=input:1x3x720x1280 --saveEngine=model.engine --tacticSources=+CUDNN,-CUBLAS,-CUBLAS_LT --skipInference

I thought it would upscale any videos (portrait/landscape) with dynamic widths and heights upto 1280px, but it is not working as expected.

Aspect ratio 16:9 works as expected upto 1280px with any dimensions
Aspect ratio 9:16 doesn't work above 360x640 (w x h)
Aspect ratio: 1:1 doesn't work above 360x360

What command should I use to make the engine work with every combination of dimensions upto 1280px?

Thanks for your help. Keep up the good work!

[Yuvraj Seegolam]

This works for portrait videos, but not landscape:

trtexec --fp16 --onnx=model.onnx --minShapes=input:1x3x8x8 --optShapes=input:1x3x1280x720 --maxShapes=input:1x3x1280x720 --saveEngine=model.engine --tacticSources=+CUDNN,-CUBLAS,-CUBLAS_LT --skipInference

[mafiosnik]

@yuvraj108c Portrait videos on mobile sometimes arent really vertical, it's just metadata that rotates the image. That's probably the issue.

[Yuvraj Seegolam]

@yuvraj108c Portrait videos on mobile sometimes arent really vertical, it's just metadata that rotates the image. That's probably the issue.

I was just wrote portrait/landscape instead of aspect ratio, i'm not actually using videos taken from mobile phones.
I'm using videos on my pc with aspect ratio 9:16 & 16:9 (e.g 960x540 or 576x1024)

Currently, I built 2 different engines using the 2 commands above for these 2 aspect ratios and switch dynamically based on the video, not sure what is the issue.

[mafiosnik]

This works for portrait videos, but not landscape:

trtexec --fp16 --onnx=model.onnx --minShapes=input:1x3x8x8 --optShapes=input:1x3x1280x720 --maxShapes=input:1x3x1280x720 --saveEngine=model.engine --tacticSources=+CUDNN,-CUBLAS,-CUBLAS_LT --skipInference

Well yeah, of course it won't lol

Height & width are flipped in shape input, so if you try doing 960x540 with that for example it won't work because 960 > 720.

If you have enough VRAM, just build large enough engine so everything fits (eg. 1x3x1080x1920), then you won't need to switch as long as the resolution is under 1920x1080.

[sudo]

It should work with every shape that is within the bounds. Created a cugan 2x engine with:

trtexec --fp16 --bf16 --onnx=cugan_pro-denoise3x-up2x_op18_fp16_clamp_colorfix.onnx --minShapes=input:1x3x8x8 --optShapes=input:1x3x1280x720 --maxShapes=input:1x3x1280x720 --saveEngine=cugan_pro-denoise3x-up2x_op18_fp16_clamp_colorfix_8x8_720p.engine --tacticSources=+CUDNN,-CUBLAS,-CUBLAS_LT --skipInference

To test Aspect ratio 9:16 doesn't work above 360x640 (w x h) I tested vs.core.resize.Bicubic(clip, width=396, height=704, format=vs.RGBH, matrix_in_s="709") and it works.

To test Aspect ratio: 1:1 doesn't work above 360x360 I tested vs.core.resize.Bicubic(clip, width=720, height=720, format=vs.RGBH, matrix_in_s="709") and it works.

Both work because both stay below 720px.

width=960, height=540 throws vapoursynth.Error: operator(): no valid optimization profile found because the max width in the engine was 720. vs.core.resize.Bicubic(clip, width=540, height=960, format=vs.RGBH, matrix_in_s="709") works fine.

Everything seems to work as intended. Like @mafiosnik777 already said, check for rotation metadata and ensure you are within the engine bounds.

[Yuvraj Seegolam]

Thanks for clarifying! It now makes sense!