DAINAPP
patreon
News on multiprocessing and V0.41
🕑 Added 2020-07-21 19:13:06 +0000 UTCFirst about 0.41:
It will be coming out very soon, it just need to some testing and it will be ready to release. The latest version of pytorch (Dain-App V0.39 and V0.40) appear to contain a lot of new bugs despise being a bigger version of Pytorch.
So i did try to update pytorch again to try to solve the new bugs, it got even worse? Pytorch don't like much of being packaged with a binary, but its the only solution to have a self contained executable as far I know. So there will be a slight downgrade on pytorch on the next build, let see if it fix some bugs.
Along with it there was some bug fixes with splits, plus even more memory improvement.
About multiprocessing:
After struggling a lot, finally I had some success on feeding multiple frames to the GPU at the same time! Is not that hard to feed multiple frames together, the hard part is improving the render time by doing it.
This is the result of my test today:
Without multiprocessing + half precision
With multiprocessing + Half Precision
From 00:27 sec. to 00:14 sec.
Not bad, I think there is still room for improvement, plus this is the first step to work with multiple processors.
Multiprocessing will not be ready for 0.41 since I pretty much broke every single thing by doing this work, broken alpha, broken depth, broken experimental, broken scene skip. But I just need time to fix all again.
Keep in mind that your GPU need to be able to store multiple frames for this to work, so its best for pixel-art and small resolution files.
Model Training:
This month I got myself Colab Pro and spent pretty much all days on it by testing model training in Dain-App. So very soon I will release a colab file V2 with improvements if someone want to train it in the future.
I improved the memory usage, to use a little more batch.
Improved the triplets dataset creator. This is useful even to others projects I think.
Will also release all my tests to help figure out what help and what don't help in the training.
Comments
GRisk
The color shift could be because of the YUV to RGB conversion, are you using mp4 or gif as input? Also, can you elaborate more about the audio output, so i can try to improve it?
Sergey Tokarev
Great. Wondering if it will improve audio output, because so far I wasn't able to get a proper result with this option turned on. Also, I was curious if CRF option works as it should, I can't really see a difference (pixel art only for now). Image quality is acceptable, but I'm still noticing a minor color shift in the output. Testing it on Win7, though.
