The captcha we about to crack is very simple, for example:
Commit Hash | Accuracy | Remark |
---|---|---|
f15750 |
50.5208% | Batch size 64, 20k. |
f15750 |
55.4688% | Batch size 128, 20k. |
not save |
48.4375% | Disable random translate, 20k. |
bf6b08 |
45.3125% | Use VGG16, 10k. |
The loss don't drop, always ~2.7. :(
- Construct the model.
- Crawl the captcha dataset.
- Label the dataset.
- Make it runnable.
- Finish the model training.
- Construct API for the trained model.
- Data augmentation: add salt & pepper noise.
I referred xmcp/elective-dataset-2021spring when crafting this project.