2nd_level

Transformer second level model

This model is based on Set Transformer

Usage

1. Define data paths

Change first level predictions, sample submission, model and final submission paths in the config.py file if needed (although, the default config should work as well). First level predictions must be in .npz format.

Example:

# Paths
PRED_PATHS = ['../1st_level/submissions/sub_140_xxl_xception41_516_test.npz',
              '../1st_level/submissions/test_preds/sub_140_xxl_xception41_bs128_524_test.npz',
              '../1st_level/submisisons/test_preds/sub_140_xxl_xception41_wp_518_test.npz',
              '../1st_level/submisisons/test_preds/sub_140_xxl_xception65_368_test.npz',
              '../1st_level/submisisons/test_preds/sub_140_xxl_xception71_370_test.npz',
              '../1st_level/submisisons/test_preds/sub_140_xxl_enet_b5_368_test.npz']
MODE_16_PATH = '../1st_level/submissions/sub_140_xxl_xception41_16modes_244_test.npz'
SAMPLE_SUB_PATH = '../1st_level/submissions/multi_mode_sample_submission.csv'
MODEL_PATH = './'
INFER_SAVE_PATH = './'

2. Training

We provide our pretrained final model in this repository -- transformer.bin. It scores 10.227 at Public LB and 9.404 at Private LB.

If you want to retrain the model from scratch, run python train.py.
The model will be saved in MODEL_PATH as transformer.bin.

3. Inference

Run python infer.py. The final submission will be saved in INFER_SAVE_PATH as submission.csv.