A fork of the SuperSloMo repository, modified in various ways to experiment with the capabilities of the neural net.
Source: https://github.com/avinashpaliwal/Super-SloMo
(Special mention of this, which the above builds on: https://github.com/TheFairBear/Super-SlowMo)
This fork is just a personal fork, which features experiments with different parameters in the neural net and dateset creation, to attempt to create nice interpolations. I personally do not know how to write or much about changing networks so this is a sandbox to try things out.
- Dataset is made to keep resolution of 720p
- Images are loaded at full resolution, and random crop is almost full image. (Sadly the only values i managed to match nicely)
- The skip connections should be going to the first convolutional layer, as per the original paper the model was derived from, unlike the project i forked this from, which sends the skip connection to the secondary convolutional layers. Actual gain is unknown, but it seemed like a change i was able to do.
- Learning rate decrease is lowered but for no other reason than to experiment.
- Replaced scheduler, using ReduceLRonPlateau.
- Reformatted stuff to ensure less warning on each run.
- Highlighting text in the console (causes a pause once python code prints i think) will now cause a clear of cache after a train/validation session, leaving an effective pause of the program without stopping training session with full GPU Memory use, letting you play games or use the GPU for other things that need some GPU Memory available.
If i get any interesting results, i might update the repository with them. I used the adobe dataset as the original repository, but with a few clips from youtube. Personally might film some 240fps videos with mobile to create data samples myself. It seems the dataset provided lacks variation to provide nice flow to some relatively simple videos. (Based on my limited knowledge and few attempts obviously, i may be speaking out of my ass on all of this for all i know...)
Please see the source repository for installation and use. Do note that many default values are changed in this fork. If you got more than 6 GB memory, i suggest increasing the train and validation batch sizes.