Remove fc layers, replace pool5 with 14x14 average pooling and softmax layer, then fine tune on CUB200.
- modify generate_prototxt.py to specify the path of caffe, lmdb
- ./fine_tune.sh
- lmdb: 256*256 cub200 dataset(#train: 5994, #val:5794).
- adopt random cropping to generate 224*224 images when training and use center cropping when test.
- training issues:
- learning rate of softmax is 10x larger;
- batch size:32, 188 iters/epoch, train 21 epoch;
- use 2 gpu cards;
- learning rate change 10^-3~10^-5 (decay 10x every 7 epoch);
| epoch | val acc | epoch | val acc | epoch | val acc |
|---|---|---|---|---|---|
| 1 | 0.4294 | 2 | 0.5299 | 3 | 0.5614 |
| 4 | 0.5740 | 5 | 0.6051 | 6 | 0.5564 |
| 7 | 0.5732 | 8 | 0.7083 | 9 | 0.7130 |
| 10 | 0.7154 | 11 | 0.7106 | 12 | 0.7125 |
| 13 | 0.7126 | 14 | 0.7140 | 15 | 0.7145 |
| 16 | 0.7149 | 17 | 0.7147 | 18 | 0.7145 |
| 19 | 0.7132 | 20 | 0.7119 | 21 | 0.7118 |
This strategy is much better than train softmax first follwed with fine tuning, which only achieving 0.70176 accuracy!