- Image classification with 340 labels.
- Evaluation metric is top-1 accuracy.
- QuickDraw dataset (Link) is used.
- Timestamped vectors are converted to 1ch images with 32x32 resolution.
- Dataset is randomly subsampled.
- You need to modify
./invoke.yamland./project/work/settings.yaml. - For using spreadsheet api, you need to put credential json in
./project/work/reporter/*.json. (Reference)
- docker (with GPU support)
- invoke
Almost all commands are executed by invoke (http://www.pyinvoke.org/). Command details are defined in tasks.py.
inv build-local
inv run-local
inv attach-local
inv stop-local
For removing unused container or image, use docker container prune docker image prune
inv push-remote
inv train-ai-platform --batch-size 128 --learning-rate 1e-2
cd project/data_preparation
python quickdraw.py
You need to put kaggle original dataset in quickdraw_original folder.
data_preparation
├── quickdraw.py
└── quickdraw_original
├── sample_submission.csv
├── test_raw.csv
├── test_simplified.csv
├── train_raw [340 entries]
└── train_simplified [340 entries]
.
├── README.md
├── ai_platform.py
├── fig
│ ├── env.png
│ └── image_sample.png
├── image
│ ├── local
│ │ └── Dockerfile
│ └── remote
│ └── Dockerfile
├── invoke.yaml
├── make_command.py
├── project
│ ├── data_preparation
│ │ ├── quickdraw.py
│ │ └── quickdraw_original
│ ├── input
│ │ └── quickdraw
│ │ ├── label_name.csv
│ │ ├── test_images.pkl.gz
│ │ ├── test_labels.pkl.gz
│ │ ├── train_images.pkl.gz
│ │ └── train_labels.pkl.gz
│ ├── notebook
│ ├── output
│ └── work
│ ├── common.py
│ ├── dataset
│ │ ├── base_dataset.py
│ │ └── distillation_dataset.py
│ ├── main.py
│ ├── model.py
│ ├── reporter
│ │ ├── put_credential_json_here.txt
│ │ └── reporter.py
│ ├── runner
│ │ ├── base_runner.py
│ │ ├── distillation_runner.py
│ │ └── mixup_runner.py
│ └── settings.yaml
├── requirements.txt
└── tasks.py