覚え書きブログ

読者です 読者をやめる 読者になる 読者になる

Deep Learning覚え書き(独自データでFasater R-CNNを学習)

hirotaka-hachiya.hatenablog.com
前回導入したFaster R-CNNを、つくばチャレンジの看板検出用に学習してみた。

以下の「独自のデータでの学習」を参考にした。
https://www.cs.gunma-u.ac.jp/~nagai/wiki/index.php?py-faster-rcnn%20%B3%D0%A4%A8%BD%F1%A4%AD

1)独自の学習データを準備
data/kanban/results/set1/2007/Main
data/kanban/set1/Annotations/*.xml
data/kanban/set1/ImageSets/Main/{trainval,test}.txt
data/kanban/set1/PNGImages/*.png

PNGImagesには、以下のようにUnityで自動生成した看板(kanban)画像と、看板なし(other)画像を200枚ずつ用意した。

>ls data/kanban/set1/PNGImages/
kanban_00000001.png  kanban_00000134.png  other_00000068.png
kanban_00000002.png  kanban_00000135.png  other_00000069.png
kanban_00000003.png  kanban_00000136.png  other_00000070.png
...
> eog data/kanban/set1/PNGImages/kanban_00000003.png

f:id:hirotaka_hachiya:20170518173505p:plain

> eog data/kanban/set1/PNGImages/other_00000003.png

f:id:hirotaka_hachiya:20170518173636p:plain

また、Annotationsには、xmlファイルで、kanbanとotherのラベルとbounding boxの座標を以下のように記載しているものを用意した。

cat data/kanban/set1/Annotations/kanban_00000003.xml
<annotation>
<filename>kanban_00000003.png</filename>
	<object>
		<name>kanban</name>
		<pose>Unspecified</pose>
		<truncated>1</truncated>
		<difficult>0</difficult>
		<bndbox>
			<xmin>183</xmin>
			<ymin>108</ymin>
			<xmax>293</xmax>
			<ymax>277</ymax>
		</bndbox>
	</object>
</annotation>

また、ImageSets/Main/には、以下のようにランダムに選択した画像名の一覧を含む、trainval.txtとtest.txtを置いた。

> cat data/kanban/set1/ImageSets/Main/trainval.txt 
other_00000018
other_00000167
other_00000038
other_00000160
...
> cat data/kanban/set1/ImageSets/Main/test.txt 
other_00000063
kanban_00000107
kanban_00000063

ちなみに、trainval.txtとtest.txtは以下のスクリプトで自動生成した。

import numpy as np
import os
import os.path

sourcedir = 'data/kanban/set1/PNGImages/'
targetdir = 'data/kanban/set1/ImageSets/Main'
trainfile = 'trainval.txt'
testfile = 'test.txt'
trainRatio = 0.8

# get file list
filelist = os.listdir(sourcedir)
filelist = np.array(filelist)

# random permutation
numfile = filelist.shape[0]
randindex = np.random.permutation(numfile)

# number of files
numtrain = np.floor(numfile*trainRatio)

# open target files
trainpath=os.path.join(targetdir,trainfile)
testpath=os.path.join(targetdir,testfile)
ftrain=open(trainpath,'w')
ftest=open(testpath,'w')
cnt = 1

# write to target files
for index in randindex:
        splits = filelist[index].split(".")

        if cnt <= numtrain:
                ftrain.write(splits[0]+"\n")
        else:
                ftest.write(splits[0]+"\n")
        cnt=cnt+1

ftrain.close()
ftest.close()

2)lib/datasets/kanban.pyを作成
lib/datasets/pascal_voc.pyからコピーし、看板検出用に修正

> cp lib/datasets/pascal_voc.py lib/datasets/kanban.py
> vi lib/datasets/kanban.py
...
class kanban(imdb):
    # setname added by hachiya
    def __init__(self, image_set, setname, year, devkit_path=None):
        imdb.__init__(self, image_set + '_' + setname)
        self._year = year
        # added by hachiya
        self._setname = setname
        self._image_set = image_set
        self._devkit_path = self._get_default_path() if devkit_path is None \
                            else devkit_path
        self._data_path = os.path.join(self._devkit_path, self._setname)
        # modified by hachiya
        self._classes = ('__background__', # always index 0
                         'other', 'kanban'
                         )
        self._class_to_ind = dict(zip(self.classes, xrange(self.num_classes)))
        # modified by hachiya from jpg to png
        self._image_ext = '.png'
...
    def image_path_from_index(self, index):
        """
        Construct an image path from the image's "index" identifier.
        """
        # modified by hachiya to load images from PNGImages
        image_path = os.path.join(self._data_path, 'PNGImages',
                                  index + self._image_ext)
...
    def _get_default_path(self):
        """
        Return the default path where PASCAL VOC is expected to be installed.
        """
        # modified by hachiya, root dir of kanban data
        return os.path.join(cfg.DATA_DIR, 'kanban')

    def gt_roidb(self):
        """
        Return the database of ground-truth regions of interest.

        This function loads/saves from/to a cache file to speed up future calls.
        """
        # modified by hachiya
        cache_file = os.path.join(self.cache_path, 'kanban_' + self.name + '_gt_roidb.pkl')
...
    def _get_voc_results_file_template(self):
        # VOCdevkit/results/VOC2007/Main/<comp_id>_det_test_aeroplane.txt
        filename = self._get_comp_id() + '_det_' + self._image_set + '_{:s}.txt'
        path = os.path.join(
            self._devkit_path,
            'results',
            # modified by hachiya
            self._setname,
            self._year,
            'Main',
            filename)
...
    def _do_python_eval(self, output_dir = 'output'):
        annopath = os.path.join(
            self._devkit_path,
            # modified by hachiya
            self._setname,
            'Annotations',
            '{:s}.xml')
        imagesetfile = os.path.join(
            self._devkit_path,
            # modified by hachiya
            self._setname,
            'ImageSets',
            'Main',
            self._image_set + '.txt')

3)lib/datasets/factory.pyに追加

> vi lib/datasets/factory.py
...
# added by hachiya
from datasets.kanban import kanban

# Set up kanban 
for setname in ['set1','set2','set3']:
    for split in ['trainval','test']:
        name = 'kanban_{}_{}'.format(setname, split)
        __sets[name] = (lambda split=split, setname=setname: kanban(split, setname, '2007'))

4)experiments/scripts/faster_rcnn_end2end.sh に追加

> vi experiments/scripts/faster_rcnn_end2end.sh
...
# added by hachiya
  kanban)
    TRAIN_IMDB="kanban_set1_trainval"
    TEST_IMDB="kanban_set1_test"
    PT_DIR="kanban"
    ITERS=70000
    ;;

5)run.shを作成

vi run.sh
#!/bin/bash

GPU=0

#NET=ZF
NET=VGG_CNN_M_1024
#NET=VGG16

DATASET=kanban
#DATASET=pascal_voc
#DATASET=pascal_voc_2012
#DATASET=coco

EXPDIR=hachiya

HOST=`hostname`
(time ./experiments/scripts/faster_rcnn_end2end.sh $GPU $NET $DATASET --set EXP_DIR $EXPDIR) 2>&1

6)kanban用のモデルファイルを用意
Pascal_VOCはクラス数が21だったのに対し、今回は3なので、train.prototxtとtest.txtのnum_classesとnum_outputを以下のように修正

> cp -rp models/pascal_voc models/kanban
> vi models/kanban/VGG_CNN_M_1024/faster_rcnn_end2end/train.prototxt
name: "VGG_CNN_M_1024"
layer {
  name: 'input-data'
  type: 'Python'
  top: 'data'
  top: 'im_info'
  top: 'gt_boxes'
  python_param {
    module: 'roi_data_layer.layer'
    layer: 'RoIDataLayer'
    param_str: "'num_classes': 3" #used to be 21
  }
}
...
layer {
  name: 'roi-data'
  type: 'Python'
  bottom: 'rpn_rois'
  bottom: 'gt_boxes'
  top: 'rois'
  top: 'labels'
  top: 'bbox_targets'
  top: 'bbox_inside_weights'
  top: 'bbox_outside_weights'
  python_param {
    module: 'rpn.proposal_target_layer'
    layer: 'ProposalTargetLayer'
    param_str: "'num_classes': 3" #used to be 21
  }
}
...
  inner_product_param {
    num_output: 3  #used to be 21
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "bbox_pred"
  type: "InnerProduct"
  bottom: "fc7"
  top: "bbox_pred"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 12  #used to be 84
    weight_filler {
      type: "gaussian"
      std: 0.001
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

> vi models/kanban/VGG_CNN_M_1024/faster_rcnn_end2end/train.prototxt
...
  inner_product_param {
    num_output: 3 # used to be 21
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "bbox_pred"
  type: "InnerProduct"
  bottom: "fc7"
  top: "bbox_pred"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 12 # used to be 84
    weight_filler {
      type: "gaussian"
      std: 0.001
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
...

また、solver.txtを編集

train_net: "models/kanban/VGG_CNN_M_1024/faster_rcnn_end2end/train.prototxt"
base_lr: 0.001
lr_policy: "step"
gamma: 0.1
stepsize: 50000
display: 20
average_loss: 100
momentum: 0.9
weight_decay: 0.0005
# We disable standard caffe solver snapshotting and implement our own snapshot
# function
snapshot: 0
# We still use the snapshot prefix, though
snapshot_prefix: "vgg_cnn_m_1024_faster_rcnn"

7)lib/fast_rcnn/config.pyの編集

> vi lib/fast_rcnn/config.py
...
# Model directory
# modified by hachiya
#__C.MODELS_DIR = osp.abspath(osp.join(__C.ROOT_DIR, 'models', 'pascal_voc'))
__C.MODELS_DIR = osp.abspath(osp.join(__C.ROOT_DIR, 'models', 'kanban'))
...

8)run.shを実行

> ./run.sh
...
Evaluating detections
Writing other VOC results file
Writing kanban VOC results file
VOC07 metric? Yes
Reading annotation for 1/80
Saving cached annotations to /home/hachiya/Works/DeepNet/py-faster-rcnn/data/kanban/annotations_cache/annots.pkl
AP for other = 0.0524
AP for kanban = 1.0000
Mean AP = 0.5262
~~~~~~~~
Results:
0.052
1.000
0.526
~~~~~~~~

--------------------------------------------------------------
Results computed with the **unofficial** Python eval code.
Results should be very close to the official MATLAB eval code.
Recompute with `./tools/reval.py --matlab ...` for your paper.
-- Thanks, The Management
--------------------------------------------------------------

real	0m7.559s
user	0m6.611s
sys	0m0.937s

real	120m27.721s
user	107m52.652s
sys	12m19.880s

学習は約2時間で終わった。テストのAverage Precisionの結果は、kanbanが100%なのに対し、otherは5%だった。。。otherは単なる背景画像だから学習できないので、次はもともとFaster R-CNNで用意している__background__とkanbanの2クラスで学習した方してみる。

9)tools/demo_kanban.pyを作成する

> cp tools/demo.py tools/demo_kanban.py
> vi tools/demo_kanban.py
...
CLASSES = ('__background__',
           'other', 'kanban')
NETS = {'vgg16': ('VGG_CNN_M_1024',
                  'VGG16_faster_rcnn_final_kanban.caffemodel'),
        'zf': ('ZF',
                  'ZF_faster_rcnn_final.caffemodel')}
...
if __name__ == '__main__':
    cfg.TEST.HAS_RPN = True  # Use RPN for proposals

    args = parse_args()

    #prototxt = os.path.join(cfg.MODELS_DIR, NETS[args.demo_net][0],
    #                        'faster_rcnn_alt_opt', 'faster_rcnn_test.pt')
    prototxt = os.path.join(cfg.MODELS_DIR, NETS[args.demo_net][0],
                            'faster_rcnn_end2end', 'test.prototxt')
...

10)学習したモデルを所定の場所にコピーし、demo_kanban.pyを実行

> cp output/hachiya/trainval_set1/vgg_cnn_m_1024_faster_rcnn_iter_70000.caffemodel data/faster_rcnn_models/VGG16_faster_rcnn_final_kanban.caffemodel
> ./tools/demo_kanban.py > /tmp/output.txt 2>&1

f:id:hirotaka_hachiya:20170519170320p:plain
f:id:hirotaka_hachiya:20170519170334p:plain
f:id:hirotaka_hachiya:20170519170346p:plain
f:id:hirotaka_hachiya:20170519175440p:plain
f:id:hirotaka_hachiya:20170519175447p:plain
f:id:hirotaka_hachiya:20170519175449p:plain
f:id:hirotaka_hachiya:20170519175749p:plain

※faster_rcnn_end2endの代わりに、models/kanban/VGG_CNN_M_1024/faster_rcnn_alt_optをモデルとして用いた場合は、下記のようなクラス数が21から3に変わったことによるblobの出力のサイズの違いに関するエラーがでた。models/kanban/VGG_CNN_M_1024/以下のptファイルの修正では対応できなかった。

I0517 20:29:52.864132 13909 net.cpp:380] loss_bbox -> loss_bbox
F0517 20:29:52.864150 13909 smooth_L1_loss_layer.cpp:28] Check failed: bottom[0]->channels() == bottom[1]->channels() (84 vs. 12)