Closed TNTwise closed 8 months ago
Thanks for doing this. I assume you used 4.13.2. Would you mind to also include the ensemble model and the 4.13 lite model? I was quite busy and didn't do it myself yet. It is a bit hard to verify a model by just looking at the param file, but if output looks good it is probably correct. I will merge once I test it later myself.
Just made the ensemble model. https://github.com/TNTwise/Rife-Vulkan-Models/releases/tag/4.13-ensemble 4.13-lite didnt export correctly, so I'll try again.
I notice that the size of models exported by others (i.e. https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/blob/master/models/rife-v4.8_ensembleFalse/flownet.bin) is larger than styler00dollar's. What caused it?
Exporting method I'm assuming. Also could be because I used convertmodel.com, instead of simplifying onnx and converting it myself.
because I used convertmodel.com
Try to enable fp16 option.
All the models I exported myself are in fp16 and I would prefer if all are. 4.13 is ~10.4mb and 4.13 lite is ~5.19mb.
Sorry, I'm changing it now.
I notice that the size of models exported by others (i.e. https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/blob/master/models/rife-v4.8_ensembleFalse/flownet.bin) is larger than styler00dollar's. What caused it?
I didn't even notice this. Looks like fp32. I can just replace the bin file later.
converted using fp16, thanks for telling me.
Tested myself and compared to TensorRT a bit, looks good visually, merging.
Sorry if this is converted incorrectly, I followed the tutorial https://github.com/styler00dollar/VSGAN-tensorrt-docker/issues/37. Param file looks different than other recent conversions, but it works fine on my end.
Edit: Converted 4.12 as well, and the output matches the model here, so I can assume 4.13 is also converted correctly. https://github.com/TNTwise/Rife-Vulkan-Models/releases/tag/4.12