AmusementClub
Support for FDAT models
I would like to know if vs-mlrt has support for FDAT models or fp16 models?
I tried using the following model with TRT in 720x480, but I got some errors and when I did, the image looked weird, black and white with lots of cracks. 2x_animefilm_light_161k_fp16_static_720x480.onnx
src = core.ffms2.Source(r'C:\anime_720x480.mkv') clipes = core.resize.Spline64(src,720, 480, format=vs.RGBS, matrix_in_s="709") clipes = core.trt.Model(clipes, engine_path=r"C:\2x_animefilm_light_161k_fp16_static_720x480.engine") clipes = core.resize.Spline64(clipes, 1280, 720, format=vs.YUV420P10, matrix_in_s="470bg", matrix_s="709") clipe.set_output()
I used the following code to generate the engine.
--fp16 --inputIOFormats=fp32:chw --outputIOFormats=fp32:chw --tacticSources=+CUDNN,+CUBLAS --builderOptimizationLevel=2 --skipInference
if it works directly with onnx, that would be good too. I tested it with TRT because it technically works with VideoJanai, which references vs-mlrt for this.
I'd appreciate it if someone with more experience could test it and point out where I'm going wrong.
Thank you.
(Sorry for my late response.)
These kinds of precision issue is common for these transformer-based models. There is no simple solution AFAIK, and you may have to disable --fp16.
Hi, that's so kind, thanks for your help.
Actually, removing --fp16 made the model work, I didn't imagine that could be it, since the trained model is fp16.
Although the model seems slow anyway. I don't know if it's a consequence of this change, as I use other models in fp32 and they are stupidly fast even when combined with other filters.
AnimeJanai or AniSD can often reach 20fps with x264 and other filters, for example. I understand that compact and SPAN are fast, but the idea behind using fp16 (according to the model's creators) is that, in theory, it would be faster than fp32.
Anyway, I appreciate the help. It seems to have worked without the errors I experienced previously. Ty~
Which version of vs-mlrt (or TensorRT in particular) and what GPU are you using? With TensorRT 10.3, I cannot reproduce the issue.
After your --fp16 tip the problem was solved.
I'm using the latest version of vsmlrt, CPU i7 13700kf and GPU 3060TI 8GiB.
what happened before was something like this image.