YOLO v1各種コマンド - KentaKawamata/Learning-data-tools GitHub Wiki
重みの認識精度を確認する
./darknet yolo recall cfg/tiny-yolo_2class.cfg results/tiny-yolo_2class_final.weights
なおここでは,voc.2007.testをdata/ディレクトリに用意する必要がある.
voc.2007.testの内容は事前にprocess.pyで作成したtest.textをコピー.
結果
0: Convolutional Layer: 448 x 448 x 3 image, 16 filters -> 448 x 448 x 16 image
1: Maxpool Layer: 448 x 448 x 16 image, 2 size, 2 stride
2: Convolutional Layer: 224 x 224 x 16 image, 32 filters -> 224 x 224 x 32 image
3: Maxpool Layer: 224 x 224 x 32 image, 2 size, 2 stride
4: Convolutional Layer: 112 x 112 x 32 image, 64 filters -> 112 x 112 x 64 image
5: Maxpool Layer: 112 x 112 x 64 image, 2 size, 2 stride
6: Convolutional Layer: 56 x 56 x 64 image, 128 filters -> 56 x 56 x 128 image
7: Maxpool Layer: 56 x 56 x 128 image, 2 size, 2 stride
8: Convolutional Layer: 28 x 28 x 128 image, 256 filters -> 28 x 28 x 256 image
9: Maxpool Layer: 28 x 28 x 256 image, 2 size, 2 stride
10: Convolutional Layer: 14 x 14 x 256 image, 512 filters -> 14 x 14 x 512 image
11: Maxpool Layer: 14 x 14 x 512 image, 2 size, 2 stride
12: Convolutional Layer: 7 x 7 x 512 image, 1024 filters -> 7 x 7 x 1024 image
13: Convolutional Layer: 7 x 7 x 1024 image, 256 filters -> 7 x 7 x 256 image
14: Connected Layer: 12544 inputs, 588 outputs
15: Detection Layer
forced: Using default '0'
Loading weights from results/tiny-yolo_2class_final.weights...Done!
Learning Rate: 0.0005, Momentum: 0.9, Decay: 0.0005
0 3 3 RPs/Img: 40.00 IOU: 88.34% Recall:100.00%
1 6 6 RPs/Img: 42.50 IOU: 84.32% Recall:100.00%
2 8 9 RPs/Img: 42.67 IOU: 76.69% Recall:88.89%
3 11 12 RPs/Img: 45.00 IOU: 75.65% Recall:91.67%
4 14 15 RPs/Img: 47.20 IOU: 77.61% Recall:93.33%
5 16 18 RPs/Img: 47.67 IOU: 74.44% Recall:88.89%
6 19 21 RPs/Img: 46.86 IOU: 74.67% Recall:90.48%
7 21 23 RPs/Img: 45.88 IOU: 75.22% Recall:91.30%
8 24 26 RPs/Img: 45.67 IOU: 76.47% Recall:92.31%
9 25 27 RPs/Img: 46.20 IOU: 77.02% Recall:92.59%
10 28 30 RPs/Img: 46.09 IOU: 77.45% Recall:93.33%
11 31 33 RPs/Img: 45.42 IOU: 77.20% Recall:93.94%
12 34 36 RPs/Img: 45.77 IOU: 78.07% Recall:94.44%
13 37 39 RPs/Img: 45.57 IOU: 78.14% Recall:94.87%
14 40 42 RPs/Img: 45.13 IOU: 78.40% Recall:95.24%
15 43 45 RPs/Img: 45.12 IOU: 77.62% Recall:95.56%
16 46 48 RPs/Img: 45.29 IOU: 77.84% Recall:95.83%
17 49 51 RPs/Img: 44.50 IOU: 78.26% Recall:96.08%
18 52 54 RPs/Img: 44.79 IOU: 78.39% Recall:96.30%
19 54 56 RPs/Img: 44.95 IOU: 78.55% Recall:96.43%
20 57 59 RPs/Img: 44.76 IOU: 79.13% Recall:96.61%
21 60 62 RPs/Img: 44.68 IOU: 78.80% Recall:96.77%
22 63 65 RPs/Img: 44.52 IOU: 78.33% Recall:96.92%
23 66 68 RPs/Img: 43.92 IOU: 77.96% Recall:97.06%
24 69 71 RPs/Img: 44.32 IOU: 78.50% Recall:97.18%
25 72 74 RPs/Img: 44.08 IOU: 79.04% Recall:97.30%
26 75 77 RPs/Img: 44.26 IOU: 79.46% Recall:97.40%
27 77 79 RPs/Img: 44.14 IOU: 79.48% Recall:97.47%
28 80 82 RPs/Img: 44.21 IOU: 79.80% Recall:97.56%
29 83 85 RPs/Img: 44.50 IOU: 79.87% Recall:97.65%
30 86 88 RPs/Img: 44.58 IOU: 79.67% Recall:97.73%
31 86 89 RPs/Img: 44.75 IOU: 79.08% Recall:96.63%
32 87 90 RPs/Img: 44.79 IOU: 79.26% Recall:96.67%
RPs/Img
IOU(Intersection Over Union)
学習の成功を評価する評価関数には報酬関数にはIoU(Intersection-over-Union)を用いる. IoUは, boxであるb(box)に対して, 目的となる領域g(ground truth box)がどれだけ含まれているかを見る.
IoUを用いて, 状態sにおいて行動aを行って状態s′に遷移する時の報酬関数Rは, 以下のように定義されます.
ある状態のIoUから次の状態へのIoUの差. またこの値は正負がbinaryで制御される. sign() ... 符号関数
閾値τを超えていれば+η, 無ければ−ηの報酬. η,τは経験則によるものが大きく, τは過大だとなかなか達成できない.
Recall
認識精度