mpekala23/h264_beds

H264-Beds

A framework for producing online video-embeddings directly on H-264 encodings.

The most important files are likely: embedding.py, small_gg.py, small_gg.ipynb, downstream.ipynb, codec.py (in that order). The rest are mainly failed experiments, plotting, or data.

Overview of Files

autoencoders - Folder containining saved weights for trained models.

embeddor_runs - Saved embeddings from running the embeddors on a single video.

frames - Frames as images.

cheap.ipynb - Scratch work for using translations to compute embeddings more efficiently. Not necessary to review.

codec.py - Our simple implementation of H-264. When run, looks in frames and produces simulated H-264 data in the coded_vids folder.

dictionary.py - Old work experimenting with using dictionary learning instead of autoencoders.

dictlearn.ipynb - Old work experiment with using dictionary learning instead of autoencoders.

downstream.ipynb - Our simple downstream anomoly detection task.

embedding.py - Defines our three embeddors. Notably contains implementations for the naive embeddor and the invariant based embeddor. Probably one of the most relevant/interesting files.

plot_embeddors.ipyb - For plotting saved embeddor runs.

small_gg.ipynb - Notebook that trains the autoencoder.

small_gg.py - File defining the autoencoder architecture.