film-poster-genres

Predicting film genres from their posters with Tensorflow.js

The example code & slide deck for a talk I gave on getting started with AI. A link to the unlisted YouTube video is available upon request (because it contains my face, and this is a public repo) - see my website for ways to get in touch.

This is the post-talk-improvements branch, which contains a number of adjustments and changes to this codebase. These changes improve the model, but make it more complex to understand and hence are kept in a separate branch to allow regular users to look at the simpler / easier to understand codebase.

Dataset

The dataset used with this demo can be found here:

https://nextcloud.starbeamrainbowlabs.com/index.php/s/wmCdL4tX3HHGAFD

The format is as follows:

+ dataset_dir/
	+ categories.txt
	+ train/
		+ 1234,Comedy,Fantasy.jpg
		+ 4321,Western,Short,Drama.jpg
		+ .....
	+ validate/
		+ 6789,Drama,Mystery,Thriller.jpg
		+ 9876,History,Documentary,Animation.jpg
		+ .....

The filenames of the images take the following format: ID,GENRE_1,GENRE_2,GENRE_N.jpg.

The categories.txt file should contain 1 category name per line (the order matters, as the line numbers - starting from 0 - are used as the numerical ids when training the model).

System / User Requirements

• Node.js
• NPM (installed by default with Node.js)
• A relatively decent CPU
• Basic knowledge of the command-line / terminal

Installation

First, clone this git repo:

git clone https://git.starbeamrainbowlabs.com/Demos/film-poster-genres.git
cd film-poster-genres

Then, install the dependencies:

npm install

Usage

Training

To train a new model:

node src/index.mjs train --input path/to/dataset_dir --output path/to/output_dir

The output directory will look like this:

 + output_dir/
	+ metrics.stream.json
	+ checkpoints/
		+ 0/
		+ 1/
		+ 2/
		+ 3/

Predicting

To make a prediction using an existing model:

node src/index.mjs predict --input path/to/image.jpg --ai-model path/to/checkpoint_dir/

The result will be written to the standard output. Extra debugging data is written to the standard error, but this can be ignored.

Contributing

Contributions are very welcome! Git patches are preferred - I can move this repo to GitHub if that makes it easier. Please mention in your contribution that you release your work under the MPL-2.0 (see below).

Licence

This code is released under the Mozilla Public License 2.0. The full license text is included in the LICENSE file in this repository. Tldr legal have a great summary of the license if you're interested.