System. I set aside 10 equally-sized contiguous regions for validation. I spin up an object-oriented system that allows me to rapidly prototype models. I train these models remotely on a Tesla GPU. The final model is in experiments/refine_6/train.py. It consists of several inception blocks and employs batch normalization, dataset augmentation, and dropout.

This explanation is very brief - please reach out to me if you have any questions.

Results. My best model yielded an MSE of 9.15 on my validation set.

How to run. To run my models follow these steps:

1. Put train.mp4, test.mp4, train.txt in the data subfolder.

2. Run these commands in the project directory.
pip install -r requirements.txt
python preprocess.py
python -m experiments.[experiment_subfolder].train [epochs] [batch_size] [samples_per_epoch]

A concrete example of the last line above is:
python -m experiments.refine_6.train 1000 10 2500