kaggle keypoints detection tutorial by keras with tensorflow
- AWS GPU g2.2xlarge
- CUDA 8.0
- cuDNN 5.0
- TensorFlow 0.1.2
- Keras 0.0.1
- Python 3.5.2 (pyenv)
| model | description | epoch | time | train loss | val loss |
|---|---|---|---|---|---|
| model1 | NN | 100 | 33sec | 0.00229 | 0.00333 |
| model2 | CNN | 1,000 | 51min | 0.00054 | 0.00168 |
| model3 | model2 + flip | 3,000 | 155min | 0.00011 | 0.00160 |
| model4 | model3 + lr | 1,000 | 52min | 0.00031 | 0.00158 |
| model5 | model4 + dropout | 1,000 | 58min | 0.00169 | 0.00158 |
| model6 | model5 + epoch/fc | 5,000 | 341min | 0.00066 | 0.00097 |
| model7 | model6 + early stop | 5,000 (3578:early stop) | 246min | 0.00066 | 0.00094 |
| model8 | train each facial key points using model7 weights | 5,000 | about 2-3 days | ? | ? |