Card Classifier

Buzzfeed style app that outputs the card color for Magic The Gathering from an input image.

The experiments fit classification models with pre-trained models in keras (VGG16). The models are each "One vs. All" for each mana class (Green, White, Blue, Black, Red, None). The app makes no consideration of the input's likelihood of being a magic card. This approach assumes it already is a magic card, just of an uncertain mana class (Like a Buzzfeed Quiz that already assumes you're a Harry Potter character).

Note: To predict whether or not an input image is a magic card one could create a classifier with all of these images in the positive class and something like COCO or ImageNet images in the negative class. I might do this later.

Procedure

Optional: Set up environment on Ubuntu 18.04 server (e.g. AWS) with:
- wget https://repo.continuum.io/archive/Anaconda3-2018.12-Linux-x86_64.sh (Download Conda)
- bash Anaconda3-2018.12-Linux-x86_64.sh (Accept defaults in prompts)
- source .bashrc for conda commands
- conda create -n card-classifier python=3.5
- tmux new -s card-classifier
- conda activate card-classifier
- pip install --upgrade pip
- sudo apt-get install libsm6 libxrender1 libfontconfig1 (for open-cv compatibility)
- pip install setuptools --upgrade
cd card-classifier
pip install -e .
aws configure (enter AWS keys in prompt, email me for a pair)
Optional:
- Download datasets and results sets with cc_download --aws (add --windows if on a windows machine)
- Skip datasets or results sets with cc_download --aws --skipdata --skipresults
- Add a --dryrun flag to outline the job without running it

Data

Download raw data locally with cc_download
- The images are extracted from this API: https://mtgjson.com/. Consider donating to support open source projects.
Curate data to crop down to artwork with cc_curate_images
Optional: Run cc_count_cards to count the cards in each category

Train

Configure experiments in card_classifier/experiments.json
Run cc_run_experiments --version v1
Optional: Run cc_run_experiments --version v0 --debug for a minimal working model. Or --overwrite to restart experiments.

Diagnostics

The experiments output a diagnostics report and sample cards to results/{model_type}/{version}/{color}
Running cc_diagnostics --model_type VGG --version v1 will run a report for all models to results/{model_type}/{version}/all.
- Note: I unwittingly deleted the training logs when I shut down my ec2 instances. So if you want those you have to go to the individual colors reports.

Predictions

Run cc_predictions --version v1 --input_path ./data/card_classifier/cc_samples --display_output to generate sample predictions
Note: samples images should reside in a subdirectory in the input_path