search

Theo-SAMINADIN-td / NukeDepthCrafter

DepthCrafter for Nuke allows you to generate temporally consistent Depth sequences inside Nuke
41 stars 1 forks source link

Windows - Triton not available #5

Closed arlo-ml closed 1 month ago

arlo-ml commented 1 month ago

thank you for sharing your plugin, it works like a charm! In order to install the proper version for Windows 11, I had to split the Pytorch/Cuda and xformers commands: pip install torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118 pip install xformers==0.0.20

I only have a few questions:

(-) I could not find any Triton implementation for Windows: WARNING:xformers:A matching Triton is not available, some optimizations will not be enabled. Error caught was: No module named 'triton' Can you suggest a workaround?

(-) I successfully used .mov .mp4 and .jpg source videos. Can you confirm .exr is not supported as a source file? [ERROR:0@0.004] global cap_ffmpeg_impl.hpp:1237 open Could not find decoder for codec_id=178 [ERROR:0@0.005] global cap_ffmpegimpl.hpp:1286 open VIDEOIO/FFMPEG: Failed to initialize VideoCapture [ WARN:0@850.947] global loadsave.cpp:248 cv::findDecoder imread('D:/DepthTest/birago2_test/clip01.exr/birago2_clip01.0025.exr'): can't open/read file: check file path/integrity

(-) All the datasets work fine, except the "open" dataset. When I use it, Nuke goes to a stall, and I have to force quit the software

depthtarget_nuke

Theo-SAMINADIN-td commented 1 month ago

Hi @arlo-ml! Glad that you like the tool :)

  1. Thanks for this question, as I didn't take the time to look for a solution, it was a great opportunity to look for one!

    I've found a workaround by downloading a precompiled Windows version of Triton : https://huggingface.co/madbuda/triton-windows-builds/tree/main

    As I work with Python 3.10 my Triton version is triton-2.1.0-cp310-cp310-win_amd64.whl

    Once downloaded, simply pip install "path/to/triton-version-win_amd64.whl". Should do the work!

  2. You are right, for now .EXR is not supported as input. I have just modified the code so now, you have a list (non-exhaustive) of supported input format. You can still try other formats by modifying the list here

  3. Depending on your CPU and VRAM, the "Open" Dataset option could be too heavy. As you can see here those datasets are simply pre-configured resolution, that DepthCrafter will use to process the input. In the case of "Open", it will process the input with its original resolution. So, if your input is 4K. It will process 4K. Which will — in more cases — be to heavy. If you still want to use the "Open" option, try to select the Sequential mode in the CPU Offload Options. It will be slower but will also save a lot of memory. I have just updated the code, so Nuke shouldn't freeze anymore. The process is now running on a child thread, so everything should run smoothly!

Hope this help :)

arlo-ml commented 1 month ago

Thank you, solved!