FAQ

Where is the main training loop?

The main loop is shared between different trainers:

• https://github.com/NeedsMorePie/interpolator/blob/master/train/trainer.py

The details are specified in the subclasses:

• https://github.com/NeedsMorePie/interpolator/blob/master/train/pwcnet/trainer.py
• https://github.com/NeedsMorePie/interpolator/blob/master/train/context_interp/trainer.py

Where is the PWC Net model defined?

The top level definition is in this file:

• https://github.com/NeedsMorePie/interpolator/blob/master/pwcnet/model.py

The various components (i.e. context and estimator networks) are in subdirectories.

Where is the PWC Net loss defined?

The total loss (with regularization) is constructed in the main model file:

• https://github.com/NeedsMorePie/interpolator/blob/master/pwcnet/model.py

The implementation details are in a separate file:

• https://github.com/NeedsMorePie/interpolator/blob/master/pwcnet/losses/loss.py

Where are the dataset classes?

https://github.com/NeedsMorePie/interpolator/tree/master/data

How do you add a new Tensorflow CUDA op?

Take the warp op for example: https://github.com/NeedsMorePie/interpolator/tree/master/pwcnet/warp/native

There are 3 files in that folder, 2 source files and 1 cmake. If you read through these files it should be fairly straight forward. The root cmake file will automatically glob/find these subdirectory cmake files and add them to the build rules.

To use these ops in Python, create a wrapper like so: https://github.com/NeedsMorePie/interpolator/blob/master/pwcnet/warp/warp.py