CONTRIBUTING.md

Requested Contributions

Please make all contributions compliant with PEP 8.

• The preprocessing method cats_vs_dogs.src.pytorch_impl.src.preprocessing.ImageDataPipeline.preprocess_classes relies on incrementing a flag step to indicate when the while loop should be terminated. At first glance, it may seem like an easy fix—just use a for loop instead of the while loop! However, it's not that simple. Doing something like for _ in range(steps) causes an entire run through each class to be considered one step. This is incorrect—each run through an image is what I'm considering a step to be (i.e. a gradient update).
  • The code I wrote achieves this vision, but it's quite ugly. If someone could refactor it or figure out a different way to approach the issue, that would be great!
  • Note that this is effectively duplicated in the TensorFlow implementation, but the PyTorch version is the "primary" implementation that I'm focused on. Focus your contributions in that one, and then I'll see to updating the TensorFlow one.
• Figure out a way to get TensorFlow's SavedModel to work as a replacement for tf.train.Saver for saving and loading not just the model, but also all relevant variables and operations (e.g. objective function, summary ops, etc.).
  • Ensure that you're working with the TensorFlow implementation, not the PyTorch one!
• A general refactoring of the TensorFlow version is desirable. The TensorFlow library, while powerful, is very complicated and sometimes very unintuitive. My implementation was based off of my own understanding of the library, but it may not fit how the authors of TensorFlow see it. I encourage you to go through the code and refactor it where you see fit! This is very much appreciated!