MoviNet PyTorch-TF

[thefatbandit]

Ishan | Rohit

• [x] Build the dataloader for the official implementation
• [x] Optional: Check out if the PyTorch & TF models are the same.

[IshanManchanda]

TF inference pipeline is ready aside from the RAM issue, expecting that to be resolved with the upgraded Colab. Worst case we'll stream the data from disk or read it in parts. Tested inference last night on a video from K600 and it produced a confident, accurate result. PyTorch model doesn't provide meaningful output for the K600 input, it seems as if the weights aren't being loaded for some reason. As discussed in the WA group that we'll be going ahead with the TF implementation, the torch issue can be left.

Can probably close this and open another issue later if the RAM problem persists.

[IshanManchanda]

Update: Full dataset inference worked on Kaggle, will be sending a PR with the code factored into functions soon