- Modified based on the Code of quark0.
- The paper is called DARTS: Differentiable Architecture Search
- This repository is used for learning Darts. I summary the procedures of Darts in search stage and val stage (please see the figure below).
- I modify the original code and run it on ChestX-Ray14 Toy Set, which is sampled by myself (as it's too slow to train Darts on large datasets).
- CheXNet is selected as baseline.
- python 3.7.3
- pytorch 1.1.0
- Select order (first or second) in search_train.py.
SECOND_ORDER = True - Run search_train.py, then genotype.txt will be generated in search_XX_order/.
- Select order and Run val_train.py, then weights .pth will be saved in weights/.
- Select order and Run test.py.
- The Metric is AUC.
- The Optimizer can be changed to get better performance maybe (when I use Adam, Darts achieves about 0.70 average AUC and 6.3 MB. On the other hand, ChestX-Ray14 dataset is sensitive to learning rate. Setting different LR (0.001/0.0001) of Adam can get different results).
- The second order version of Darts haven't been tested, as it's extremely slow.
| Method | CheXNet | Darts (first order) |
|---|---|---|
| Parameter | 6.97MB | 7.34MB |
| Atelectasis | 0.685 | 0.667 |
| Cardiomegaly | 0.779 | 0.769 |
| Effusion | 0.788 | 0.742 |
| Infiltration | 0.669 | 0.648 |
| Mass | 0.734 | 0.654 |
| Nodule | 0.655 | 0.588 |
| Pneumonia | 0.681 | 0.578 |
| Pneumothorax | 0.797 | 0.709 |
| Consolidation | 0.715 | 0.677 |
| Edema | 0.783 | 0.758 |
| Emphysema | 0.767 | 0.647 |
| Fibrosis | 0.689 | 0.638 |
| Pleural_Thickening | 0.682 | 0.642 |
| Hernia | 0.876 | 0.814 |
| Average | 0.736 | 0.681 |
None