Write module to run denoising inference

[danielsf]

This ticket is being written to refine the work in #374.

In order to better control the code being used for denoising, we want to write our own wrapper around tensorflow to accomplish the inference step in denoising (the act of actually applying a model to a noisy movie to produce a denoised movie).

• The module should conform to Pika's coding standards.
• The module should be well tested.
• The module should use our typical argschema-driven CLI.
• The module should run efficiently on CPU nodes (it may have the option of utilizing GPU ndoes). As a baseline, we have been able to denoise a 20 GB movie on a CPU node in about four hours using this branch of deepinterpolation https://github.com/danielsf/deepinterpolation/tree/staging/ophys_etl

The core_inferrence class in deepinterpolation can be used as a guide

https://github.com/danielsf/deepinterpolation/blob/staging/ophys_etl/deepinterpolation/inferrence_collection.py#L264

Components that will need to be implemented

• A class to read in, store, and iterate over data (should be implemented in Write data generator for denoising #471)
• A way to define the network
• A way to load the model weights into the network
• A way to call tensorflow's predict method on the data provided by the data iterator
• Save the output movie to disk

Tasks

• Write a module to apply a denoising model to a noisy movie and produce a denoised movie.

Validation

Take a sample of the noisy movies from
/allen/programs/mindscope/workgroups/surround/motion_correction_labeling_2022

And the corresponding models from
/allen/programs/mindscope/workgroups/surround/denoising_labeling_2022/bespoke_models

And reproduce the corresponding denoised movies from
/allen/programs/mindscope/workgroups/surround/denoising_labeling_2022/denoised_movies