Inclusive Images Challenge

Kaggle competition https://www.kaggle.com/c/inclusive-images-challenge. Solution implemented in PyTorch, (19th place).

How to run

• Setup data directory - To run, data directory needs to be setup as follows:

data_dir
└─── train_images
└─── test_images
└─── labels
    │─── class-descriptions.csv
    │─── classes-trainable.csv
    │    ...

• train_images with the training data (portion of the Open Images dataset) is only required to replicate training process.

• test_images - can be stage-1 or stage-2.

• To generate predictions on stage-1 or stage-2 for submission (with provided weights)

python main.py --data-path <path_to_data_dir> --mode adapt --use-ensemble true --n-models 5

• The model ensemble weights are loaded from experiments, and predictions over the test set are generated, and the ensemble then bootstraps itself over the predictions for 2 epochs (emperically chosen to give best results), then the generated file is used for submission.

• To train, create a text file with paths to training images. This is split into train/val. The write_trainval_image_list method in the IncImagesDataset class can be used for this.

• The script only supports single model training. Single model checkpoint files are combined to form an ensemble which is then applied to the test set.