Soft flow mini repository

Includes code to run some soft flow things. This git repository automatically uploads to weights and biases at the soft_flow project (can be changed in overfit.yaml).

To train a model, do

python main.py --name name_of_run --mode {online,dryrun,offline,disabled}

Note: pytorch lightning will try to use all the gpus on the machine.

Files

• main.py: training loop, shouldn't need to be modified
• overfit_soft_learner.py: code for the soft flow model. This version uses superresolution, warping a 1024x512 source frame (frame 3) to a 256x128 target frame (frame 2), learning flows frame2 to frame3 (i think). Includes brief explanations of the logged data
• sintel_superres.py: dataloading. Note: the path to the Sintel dataset needs to be changed
• roaming_images.py: roamingimages dataset, toy dataset. Not integrated with the current soft flow learner
• overfit.yaml: configuration file
• download_learner.py: some minimal code to download trained models (models are currently saved every 5000 steps), and then you can run arbitrary code on them
• soft_losses.py, soft_utils.py: losses and utilities for soft flow