fcn （三）training cityscapes 學習紀錄

-
這一篇是想訓練cityscapes的資料庫，

紀錄訓練過程的一些小碰撞

1. 首先是 copy 了 voc-fcn32s/ 的檔案，
.
├── infer.py
├── net.py
├── voc_layers.py ---> new add cityscapes_layers.py
├── score.py
├── snapshot   ---> cteate snapshot/train
├── solve.py
├── solver.prototxt
├── surgery.py
├── train.prototxt
└── val.prototxt

2. datasets



6.
Result 512x512
fcn32s mean acc. and mean IU 太低
>>> 2019-12-11 22:45:07.762072 Begin seg tests
>>> 2019-12-11 22:45:41.055221 Iteration 100000 loss 126417.949921875
>>> 2019-12-11 22:45:41.055304 Iteration 100000 overall accuracy 0.602877947321908
>>> 2019-12-11 22:45:41.055342 Iteration 100000 mean accuracy 0.09731878508947552
>>> 2019-12-11 22:45:41.055503 Iteration 100000 mean IU 0.061253671561096214
>>> 2019-12-11 22:45:41.055591 Iteration 100000 fwavacc 0.3756002265727415

real    312m59.278s --> 5 hours
user    251m9.625s
sys    60m51.121s


Result 1024*2048
fcn32s iter 200, origin image size 效果好一些哈
>>> 2019-12-12 13:49:12.992267 Begin seg tests
>>> 2019-12-12 13:54:41.148831 Iteration 100000 loss 983594.96875
>>> 2019-12-12 13:54:41.149053 Iteration 100000 overall accuracy 0.8349910559649044
>>> 2019-12-12 13:54:41.149228 Iteration 100000 mean accuracy 0.2978698254521429
>>> 2019-12-12 13:54:41.149530 Iteration 100000 mean IU 0.24129410974675303
>>> 2019-12-12 13:54:41.149755 Iteration 100000 fwavacc 0.7225745036399619

real    1205m50.615s  -->20 hours
user    1215m26.764s
sys    242m45.794s

用了之前train voc_fcn32s 的係數 精準度有提昇一些
I1215 01:10:01.700105  2337 sgd_solver.cpp:284] Snapshotting solver state to binary proto file snapshot/train/solver_iter_200000.solverstate
>>> 2019-12-15 01:10:02.649641 Begin seg tests
>>> 2019-12-15 01:15:35.265890 Iteration 200000 loss 861699.7655
>>> 2019-12-15 01:15:35.266108 Iteration 200000 overall accuracy 0.8491470003094094
>>> 2019-12-15 01:15:35.266284 Iteration 200000 mean accuracy 0.3588967739276047
>>> 2019-12-15 01:15:35.266571 Iteration 200000 mean IU 0.2900807413231396
>>> 2019-12-15 01:15:35.266798 Iteration 200000 fwavacc 0.7471066540397774

real    2282m12.953s  -->  38 hours
user    2064m43.217s
sys     452m21.950s









留言

張貼留言