fcn voc-fcn32s training 手起刀落

-


這一篇是想訓練fcn.berkleyvision.org/voc-fcn32s,

但是照著網路教學，瞎跑都不能得到低的 loss 跟 高 mean accuracy

藉此想po,紀錄想明白自己的盲點，照著做不會成功的，還沒成功過哈，（成功了11/29）

如果好心人經過可以指點一二，真心感激不盡～

1. github : fcn. berkelyvision
    • https://github.com/shelhamer/fcn.berkeleyvision.org.git
2. dataset
    • ├── data
      │   ├── pascal
      │   │   ├── classes.txt
      │   │   ├── README.md
      │   │   ├── seg11valid.txt
      README.md 有提供下載以下兩個資料庫的網址
  • test dataset: 
    • VOCtrainval_11-May-2012
       
  • training dataset : 
    • benchmark.tgz   (我是從官網下載，有一天連進去就載下來了)
3.  copy 以下檔案到voc-fcn32s/
  • fcn.berkeleyvison
    • ├── infer.py
      ├── score.py
      ├── surgery.py
      ├── vis.py
      ├── voc_layers.py
  •  voc-fcn8s
    • ├── voc-fcn8s
      │   ├── deploy.prototxt 
4. 修改每個檔案的紀錄, voc-fcn32s/
  • ├── caffemodel-url
    ├── solver.prototxt
    ├── solve.py
    ├── net.py     ---> 執行net.py 會產生train.prototxt, deploy.prototxt.
    ├── train.prototxt
    └── val.prototxt
  •  ├── deploy.prototxt
  • ├── infer.py
    ├── score.py
    ├── surgery.py
    ├── vis.py
    ├── voc_layers.py
 
4.1
Solve.py
*seg11valid.txt 是test 檔案的 list，改不改都可以，只是我把它拉到voc-fcn32s. 這邊的seg11valid.txt 跟 (VOC2012/ImageSets/segmentaton/val.txt) 一樣。
*seg11valid.txt 是test 檔案的 list，改不改都可以，只是我把它拉到voc-fcn32s.
這邊的seg11valid.txt 跟 (VOC2012/ImageSets/segmentaton/val.txt) 一樣。
























 
4.2 Solver.prototxt


4.3
net.py


4.4
 deploy.protxt		 Copy from voc-fen8s的layout, 所以根據這邊的train.prototxt, 把多的刪掉。


自己從train.prototxt 改過來也行，就是把drop 跟 loss拿掉 ，還有一個我忘了，先這樣。


5. Result:

>>> 2019-11-04 21:12:46.816136 Begin seg tests
>>> 2019-11-04 21:13:38.519890 Iteration 100000 loss 14785045.613111412
>>> 2019-11-04 21:13:38.519979 Iteration 100000 overall accuracy 0.7519151040535762
>>> 2019-11-04 21:13:38.520038 Iteration 100000 mean accuracy 0.047619047619047616
>>> 2019-11-04 21:13:38.520189 Iteration 100000 mean IU 0.035805481145408394
>>> 2019-11-04 21:13:38.520283 Iteration 100000 fwavacc 0.5653763237039003



real    373m22.532s
user    291m2.600s
sys    71m55.303s


6. caffemodel 用過 vgg16-fcn.caffemodel 和 vgg_ilsvrc_16_layers_deploy.prototxt
├── ilsvrc-nets
│   ├── README.md
│   ├── vgg16-fcn.caffemodel
│   └── vgg_ilsvrc_16_layers_deploy.prototxt

6.1 Result
>>> 2019-10-28 23:09:09.807761 Begin seg tests
>>> 2019-10-28 23:10:03.737844 Iteration 100000 loss 204808.4506119438
>>> 2019-10-28 23:10:03.737958 Iteration 100000 overall accuracy 0.7519151040535762
>>> 2019-10-28 23:10:03.738027 Iteration 100000 mean accuracy 0.047619047619047616
>>> 2019-10-28 23:10:03.738209 Iteration 100000 mean IU 0.035805481145408394
>>> 2019-10-28 23:10:03.738317 Iteration 100000 fwavacc 0.5653763237039003

--------------------------------------------------------------------------------------------------------
2019/11/19 result
 voc-fcn32s fixed learning rate 
>>> 2019-11-19 11:17:32.511732 Begin seg tests
>>> 2019-11-19 11:18:40.850477 Iteration 400000 loss 110037.21422411047
>>> 2019-11-19 11:18:40.883480 Iteration 400000 overall accuracy 0.7997448944020139
>>> 2019-11-19 11:18:40.884994 Iteration 400000 mean accuracy 0.39614343211991015
>>> 2019-11-19 11:18:40.886317 Iteration 400000 mean IU 0.3090150660161627
>>> 2019-11-19 11:18:40.886498 Iteration 400000 fwavacc 0.6698493403057357

real    1477m49.550s
user    1067m49.814s
sys     262m44.040s


WARNING: Logging before InitGoogleLogging() is written to STDERR
I1118 10:41:06.658674  3580 upgrade_proto.cpp:1113] snapshot_prefix was a directory and is replaced to snapshot/train_mean/solver
I1118 10:41:06.658700  3580 solver.cpp:45] Initializing solver from parameters:
train_net: "train.prototxt"
test_net: "val.prototxt"
test_iter: 5105
test_interval: 999999999
base_lr: 1e-14
display: 20
max_iter: 100000
lr_policy: "fixed"
momentum: 0.99
weight_decay: 0.0005
snapshot: 4000
snapshot_prefix: "snapshot/train_mean/solver"
test_initialization: false
average_loss: 20
iter_size: 1
I1118 10:41:06.680038  3580 solver.cpp:92] Creating training net from train_net file: train.prototxt
I1118 10:41:06.714169  3580 net.cpp:53] Initializing net from parameters:
state {
  phase: TRAIN
}
layer {
  name: "data"
  type: "Python"
  top: "data"
  top: "label"
  python_param {
    module: "voc_layers"
    layer: "SBDDSegDataLayer"
    param_str: "{\'sbdd_dir\': \'/home/zoey/fcn.berkeleyvision.org/data/pascal/benchmark_RELEASE/dataset\', \'seed\': 1337, \'split\': \'train\', \'mean\': (72.5249, 82.9668, 73.1944)}"

--------------------------------------------------------------------------------------------------------
2019/11/26 fcn16s result
>>> 2019-11-26 08:47:00.978160 Begin seg tests
>>> 2019-11-26 08:47:54.618272 Iteration 400000 loss 52454.30979355522
>>> 2019-11-26 08:47:54.618347 Iteration 400000 overall accuracy 0.8988595064042302
>>> 2019-11-26 08:47:54.618378 Iteration 400000 mean accuracy 0.7665972274429271
>>> 2019-11-26 08:47:54.618493 Iteration 400000 mean IU 0.610513842937149
>>> 2019-11-26 08:47:54.618590 Iteration 400000 fwavacc 0.8290560953274788
 
real    1393m40.017s
user    1057m28.505s
sys     263m40.896s

 train_net: "train.prototxt"
test_net: "val.prototxt"
test_iter: 5105
# make test net, but don't invoke it from the solver itself
test_interval: 999999999
display: 20
average_loss: 20
lr_policy: "fixed"
# lr for unnormalized softmax
base_lr: 1e-12
# high momentum
momentum: 0.99
# no gradient accumulation
iter_size: 1
max_iter: 100000
weight_decay: 0.0005
snapshot: 16000
snapshot_prefix: "snapshot/train"
test_initialization: false
----------------------------------------------------------------------------------------------------
 2019/11/27 fcn8s result

ary proto file snapshot/train/solver_iter_400000.solverstate
>>> 2019-11-27 09:56:14.725841 Begin seg tests
>>> 2019-11-27 09:57:08.472657 Iteration 400000 loss 52320.597529286926
>>> 2019-11-27 09:57:08.472755 Iteration 400000 overall accuracy 0.8992662322465721
>>> 2019-11-27 09:57:08.472821 Iteration 400000 mean accuracy 0.7735758690259322
>>> 2019-11-27 09:57:08.473015 Iteration 400000 mean IU 0.6134128305071153
>>> 2019-11-27 09:57:08.473153 Iteration 400000 fwavacc 0.8305588046072755

real    1400m26.232s
user    1063m17.779s
sys     263m38.287s

--------------------------------------------------------------------------------------------------
 2019/11/27 fcn8s-atonce result

>>> 2019-11-29 03:44:02.948359 Begin seg tests
>>> 2019-11-29 03:44:57.677574 Iteration 600000 loss 73545.70225838992
>>> 2019-11-29 03:44:57.677653 Iteration 600000 overall accuracy 0.9367826556673325
>>> 2019-11-29 03:44:57.677707 Iteration 600000 mean accuracy 0.8225041223763434
>>> 2019-11-29 03:44:57.677886 Iteration 600000 mean IU 0.7184923129601997
>>> 2019-11-29 03:44:57.678008 Iteration 600000 fwavacc 0.885896417975581

real    2112m24.997s
user    1614m12.777s
sys    395m18.724s


 


[參考]
  • FCN学习:Semantic Segmentation 
    •  介紹了Fcn 裡面 cnn layer 的參數跟影像大小的算法。
    • infer.py，不用自己設置，現在新的voc_palette = vis.make_palette(21)
      就可以直接生成21個顏色，跟論文的順序是一樣的。
  • 手把手在自己的数据集cityscapes上训练FCN 
    • 最佳训练的流程是 FCN-32s -> FCN-16s -> FCN 8s，FCN-32s我们用VGG16-fcn从头训练，收敛后Copy weights训练FCN-16s ..至 FCN-8s. 如果一步到FCN-8s也是可以的
    • solve.py 的修改，是參考這裡的。
      vgg_weights = '../ilsvrc-nets/vgg16-fcn.caffemodel'
vgg_proto = '../ilsvrc-nets/VGG_ILSVRC_16_layers_deploy.prototxt'vgg_net = caffe.Net(vgg_proto, vgg_weights, caffe.TRAIN)
surgery.transplant(solver.net, vgg_net)
del vgg_net
  • FCN网络的训练——以SIFT-Flow 数据集为例 
    • 一般情况下不要直接修改test.prototxt和trainval.prototxt，而是执行net.py这个脚本，执行完成后也不要将test.prototxt和trainval.prototxt中的fc6和fc7替换为其他名称.
    • deploy.prototxt这个文件，可以从test.prototxt或者trainval.prototxt复制，然后删除最后一层loss层
    • 這網誌告訴我們4.1 solve.py , 注意是传入的测试集对应的test.txt，不是训练集对应的train.txt。

留言

張貼留言