This is a RES-Net based Neural Network model that is challenging a kaggle contest, Facial Keypoints Detection.
The objective of this task is to predict keypoint positions on face images. Datasets are available on the kaggle website.
As a Deep Learning framework, I used Pytorch version 1.3.
If you want to training the model on your own environment,
$ python3 train.py If you want to use my pretrained model to fine-tuning, --load option is available.
$ python3 train.py --loadWhen you would like to generate kaggle submission file,
$ python3 inference.pyThen type the model file name (generally saved as save.pt)
save model file : [YOUR MODEL FILE NAME].pt
The score update has been canceled because the image size has been changed to (226, 226) from (96, 96)
| Kaggle Private Score | Kaggle Public Score | Training epoch |
|---|---|---|
| 2.99794 | 3.08839 | choose best epoch after 1000 epochs training |
| 2.97505 | 3.07060 | 500 epochs |
| 3.10676 | 3.21217 | 300 epochs |
- log system
- Improve usability
- Add comments to the functions
- Generate Results image function