This repository hosts our code for the iMaterialist Fashion Challenge.

Install Python 3.6 or higher and all dependencies necessary in a virtual environment. Then install this project with the editable option:

$ pip install -e .

Download the Kaggle data files into the data directory, and run the following scripts (includes deletion of broken images):

$ python scripts/download_images.py ./data/train.json ./data/train
$ python scripts/download_images.py ./data/validation.json ./data/validation
$ python scripts/download_images.py ./data/test.json ./data/test
$ python scripts/delete_broken_images.py

$ python scripts/create_labels.py ./data/ ./data/

All scripts should be ran from the project root directory, e.g.:

$ python scripts/multilabel_nn.py

This is necessary to make the relative imports and path strings work properly.

Install Google Cloud SDK for your OS: https://cloud.google.com/sdk/docs/

Run the following command, replacing job_name with a nice name for your session and module-name pointing to the script you want to run (eg. scripts.inceptionv3):

$ gcloud --project mlip-team-hunter2 ml-engine jobs submit training job_name --stream-logs --runtime-version 1.4 --job-dir gs://hunter2-project/job --package-path scripts --module-name scripts.simple_xception --region europe-west1 --config ./config.yaml -- --save-filename test --gcp