5th Place Solution of Kaggle Fast or Slow Competition

This is knshnb's solution for "Google - Fast or Slow? Predict AI Model Runtime"

Run

Dataset Preparetion

1. Download the competition dataset under input/.

$ ls -F input
npz_all/  pb/  sample_submission.csv

2. Run the following command and make sure npy files are generated under input/npz_all/npz/layout-config/.

$ python -m scripts.preprocess_layout_data

Training and Inference

Layout model:

python -m scripts.train --save_model --config_path config/default.yaml --exp_name layout-exp

Tile model:

python -m scripts.train --save_model --config_path config/tile.yaml --exp_name tile-exp

By running the above commands, the trained models and inference results will be saved under result/layout-exp/0/ and result/tile-exp/0/ respectively.

Submission

By specifying the directories generated in the previous step, you can make a submission file for the competition.

Example:

python -m scripts.make_submission \
--layout_model_dirs result/layout-exp/0 \
--tile_model_dirs result/tile-exp/0

Links

• For an overview of our key ideas and detailed explanation, please also refer to 5th Place Solution: GNN with Invariant Dimension Features