盤メーカーとIoTそしてAI: 学習フェイズ

E-proさんから頂いたサーモグラフィーデータで学習を始めています。
動画から画像を取り出したり、その動画にアノテーションの為の座標とラベルを作るソフトのインストールと学習をスタートするまでの作業で一週間掛かりました。
と言っても日常の業務の合間に少し進めてはストップの繰り返しで歳のせいか途中までの作業を思い出すのも一苦労です。

さて、データを切り出してアノテーションしましたがデータ自体が１００ちょっとと学習には少なすぎますがその後の手順の確認に作業を進めて学習をスタートしました。

AlexeyAB さんのDarknetの場合はhttps://github.com/AlexeyAB/darknetが全てと言う事でその指示にしたがってパラメータを入力しました。
1: 45.906712, 45.906712 ~~avg loss, 0.000000~~ rate, 1460.403615 seconds, 64 images
Loaded: 0.000022 seconds
Region Avg IOU: 0.090963, Class: 0.332260, Obj: 0.500453, No Obj: 0.500137, Avg Recall: 0.023810, count: 42
Region Avg IOU: 0.063928, Class: 0.332210, Obj: 0.500530, No Obj: 0.500144, Avg Recall: 0.000000, count: 29
Region Avg IOU: 0.181828, Class: 0.332264, Obj: 0.500599, No Obj: 0.500149, Avg Recall: 0.171429, count: 70
Region Avg IOU: 0.062633, Class: 0.332089, Obj: 0.500546, No Obj: 0.500141, Avg Recall: 0.000000, count: 46
Region Avg IOU: 0.109374, Class: 0.332049, Obj: 0.500692, No Obj: 0.500140, Avg Recall: 0.038462, count: 26
Region Avg IOU: 0.147027, Class: 0.332460, Obj: 0.500515, No Obj: 0.500137, Avg Recall: 0.081967, count: 61
Region Avg IOU: 0.115006, Class: 0.332235, Obj: 0.500552, No Obj: 0.500143, Avg Recall: 0.035714, count: 28
Region Avg IOU: 0.357765, Class: 0.333455, Obj: 0.500809, No Obj: 0.500143, Avg Recall: 0.108108, count: 37

~~アベレージロスが0.000000~~と出ています。間違えてました。~~0.000000~~ rateって何だろう？
作者によれば、
Note: If during training you see nan values for avg (loss) field - then training goes wrong, but if nan is in some other lines - then training goes well.
nan＝０なので学習に失敗しています。

原因は今の時点ではわかりません。アノテーションの作業自体にエラーは出ていないのでデータの量と質に問題があるのかも？
今日はたまたま社外の方からAIの外注は高額な理由を聞かれましたが、この試行錯誤の作業をシステムのエラーかデータの不備かの判断をこの最新を理解するエンジニアに依頼すれば高額にならざるおえないと説明しました。

もちろん既知の技術で問題を解決する作業が簡単と言うわけではなくディープラーニング自体の情報が少なく手探りで進めるしかないのが現状です。ただその向こうには他社に真似の出来ない結果が待っていると言う事です。

出来るだけ他の作業を止めてこの作業に集中しなければいけないようです。

１１月２６日すべて復旧して再開。
Murouのファイルのラベルを確認
次にlabelimg/data の中にあるpredefined_class.txtを修正

labelimgフォルダの中でプログラムを実行します。

　python3 labelimg.py

darknet/data の中に先ほど作ったjpgとtxtの入ったごちゃ混ぜデータをディレクトリごと移動またはコピー。

　その中にprocess.pyの名称で、以下の内容のファイルを作成ディレクトリ要注意

path_data = 'data/Murou/obj/'

　一応　darknet/ 内に必要ファイルをダウンロードします。このファイルは学習を収束させるために必要なファイルだそうです。　　

　　YOLO V3.0の場合は

　　　　wget https://pjreddie.com/media/files/darknet53.conv.74

　　　　darknet53.conv.74がダウンロードされます。

　darknet/cfgの中にobj.namesファイルを作成

　　obj.names　：先ほどのpredefined_class.txtと内容が同じファイル。

　　　クラス名を改行しながら書きます。今回はHotspot、Panel-Err、String-Err

obj.data :必要ファイルのリンクを書き込んだファイル。以下のような感じ。

　　　　　　class 必ず指定。今回の場合3

　　　　　　train　ディレクトリとファイル名を指定　先ほどのtrain.txt

　　　　　　valid　ディレクトリとファイル名を指定　先ほどのtest.txt

　　　　　　names クラスファイルの場所とファイル名　obj.names

　　　　　　backup 学習済みデータの保存先　このままでOK

classes= 3

train = data/Murou/train.txt

valid = data/Murou/test.txt

names = cfg/obj.names

backup = backup/

cfgファイルの設定

Create file yolo-obj.cfg with the same content as in yolov3.cfg (or copy yolov3.cfg to yolo-obj.cfg) and:
cfgファイルは新たに作らないと下記デモで同じcfgファイルを使わうので。。。
./darknet detector demo cfg/coco.data cfg/yolov3.cfg yolov3.weights

change line batch to batch=64

change line subdivisions to subdivisions=16

一般的には416x416。精度を上げるには608x608または832x832とする。ただしyolov3の場合608x608で学習させると、私の環境ではメモリーオーバーで止まる。今回618x618の場合は　subdivisions=16 とした。

20行目の　max_bachesを30000くらいにする

22行目　steps=4000,4500　くらいにする

603,610,689,696,776,783行に書いてある

filters=255

classes=80

の所をクラスが３なら

filters=24

classes=3

に変更します　filtersは　（クラス数+5)X3 の数値です

いよいよ学習

./darknet detector train cfg/obj.data cfg/yolo-obj.cfg darknet53.conv.74 -map

コマンド間違って８：２０頃止めてしまいました。

Ctrl+cでコピーコマンドでログを取ろうとしましたがCtrl+cはquit(停止）と同じでしたね（笑）

(next mAP calculation at 14700 iterations)
Last accuracy mAP@0.5 = 99.35 %, best = 99.43 %
14681: 0.225036, 0.257150 avg loss, 0.000010 rate, 6.993637 seconds, 939584 images
Loaded: 0.000019 seconds
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.912137, GIOU: 0.910393), Class: 0.999881, Obj: 0.983405, No Obj: 0.000766, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.830349, GIOU: 0.826423), Class: 0.995696, Obj: 0.897695, No Obj: 0.000465, .5R: 1.000000, .75R: 1.000000, count: 5
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000005, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000002, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.880829, GIOU: 0.880195), Class: 0.999061, Obj: 0.854198, No Obj: 0.000727, .5R: 1.000000, .75R: 1.000000, count: 7
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.800322, GIOU: 0.795880), Class: 0.998760, Obj: 0.395001, No Obj: 0.000165, .5R: 1.000000, .75R: 0.714286, count: 7
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.761692, GIOU: 0.743397), Class: 0.992795, Obj: 0.088460, No Obj: 0.000442, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.892330, GIOU: 0.891681), Class: 0.999716, Obj: 0.970838, No Obj: 0.001251, .5R: 1.000000, .75R: 1.000000, count: 10
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.781882, GIOU: 0.780087), Class: 0.949109, Obj: 0.497191, No Obj: 0.000057, .5R: 1.000000, .75R: 0.333333, count: 3
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.857719, GIOU: 0.857054), Class: 0.999225, Obj: 0.844893, No Obj: 0.000485, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.802336, GIOU: 0.798664), Class: 0.998920, Obj: 0.955014, No Obj: 0.000785, .5R: 1.000000, .75R: 0.750000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.838987, GIOU: 0.834024), Class: 0.998663, Obj: 0.400468, No Obj: 0.000035, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000282, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.869823, GIOU: 0.867091), Class: 0.999491, Obj: 0.947843, No Obj: 0.000634, .5R: 1.000000, .75R: 1.000000, count: 5
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000002, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.913126, GIOU: 0.911567), Class: 0.999862, Obj: 0.995943, No Obj: 0.000754, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.804345, GIOU: 0.794554), Class: 0.999741, Obj: 0.833474, No Obj: 0.000557, .5R: 1.000000, .75R: 0.666667, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.828259, GIOU: 0.822738), Class: 0.998422, Obj: 0.600344, No Obj: 0.000076, .5R: 1.000000, .75R: 1.000000, count: 3
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.873014, GIOU: 0.870102), Class: 0.999591, Obj: 0.998685, No Obj: 0.001034, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.888265, GIOU: 0.887465), Class: 0.999761, Obj: 0.987195, No Obj: 0.000591, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.792944, GIOU: 0.787508), Class: 0.999825, Obj: 0.907702, No Obj: 0.000157, .5R: 1.000000, .75R: 0.500000, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.900194, GIOU: 0.899056), Class: 0.999739, Obj: 0.997777, No Obj: 0.000506, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.880843, GIOU: 0.879901), Class: 0.999923, Obj: 0.999096, No Obj: 0.000403, .5R: 1.000000, .75R: 1.000000, count: 5
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.834652, GIOU: 0.832130), Class: 0.999865, Obj: 0.769075, No Obj: 0.000301, .5R: 1.000000, .75R: 0.909091, count: 11
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.847607, GIOU: 0.845608), Class: 0.998550, Obj: 0.997036, No Obj: 0.000357, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.828542, GIOU: 0.826807), Class: 0.999834, Obj: 0.938351, No Obj: 0.000534, .5R: 1.000000, .75R: 0.833333, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.807382, GIOU: 0.807382), Class: 0.999645, Obj: 0.446545, No Obj: 0.000044, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000001, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.864837, GIOU: 0.862551), Class: 0.999919, Obj: 0.996243, No Obj: 0.000655, .5R: 1.000000, .75R: 1.000000, count: 5
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.812558, GIOU: 0.807358), Class: 0.999930, Obj: 0.933584, No Obj: 0.000038, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000002, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.897845, GIOU: 0.897455), Class: 0.999608, Obj: 0.984354, No Obj: 0.000419, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.815775, GIOU: 0.806992), Class: 0.999730, Obj: 0.619597, No Obj: 0.000044, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000002, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.862937, GIOU: 0.860239), Class: 0.999918, Obj: 0.991235, No Obj: 0.000616, .5R: 1.000000, .75R: 1.000000, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.841489, GIOU: 0.834102), Class: 0.999776, Obj: 0.703621, No Obj: 0.000030, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.827845, GIOU: 0.826595), Class: 0.999646, Obj: 0.941644, No Obj: 0.001461, .5R: 1.000000, .75R: 1.000000, count: 3
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.832839, GIOU: 0.831670), Class: 0.999925, Obj: 0.993587, No Obj: 0.000677, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.799025, GIOU: 0.795008), Class: 0.913052, Obj: 0.587756, No Obj: 0.000132, .5R: 1.000000, .75R: 0.833333, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.775533, GIOU: 0.769263), Class: 0.999165, Obj: 0.476534, No Obj: 0.000147, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.867817, GIOU: 0.863557), Class: 0.999006, Obj: 0.668676, No Obj: 0.000477, .5R: 1.000000, .75R: 1.000000, count: 3
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.623888, GIOU: 0.616664), Class: 0.546459, Obj: 0.391367, No Obj: 0.000020, .5R: 1.000000, .75R: 0.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.870889, GIOU: 0.869707), Class: 0.999214, Obj: 0.914650, No Obj: 0.001233, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.862204, GIOU: 0.860642), Class: 0.999140, Obj: 0.719939, No Obj: 0.000630, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.805423, GIOU: 0.799391), Class: 0.994189, Obj: 0.102547, No Obj: 0.000026, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000002, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.862386, GIOU: 0.859989), Class: 0.999880, Obj: 0.742697, No Obj: 0.000445, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.791871, GIOU: 0.787106), Class: 0.999460, Obj: 0.744960, No Obj: 0.000040, .5R: 1.000000, .75R: 0.500000, count: 2

(next mAP calculation at 14700 iterations)
Last accuracy mAP@0.5 = 99.35 %, best = 99.43 %
14682: 0.278860, 0.259321 avg loss, 0.000010 rate, 7.174890 seconds, 939648 images
Loaded: 0.000033 seconds
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000002, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.810703, GIOU: 0.806056), Class: 0.999030, Obj: 0.744211, No Obj: 0.000674, .5R: 1.000000, .75R: 0.750000, count: 8
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.764298, GIOU: 0.760145), Class: 0.998906, Obj: 0.575124, No Obj: 0.000132, .5R: 1.000000, .75R: 0.600000, count: 5
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000003, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.868377, GIOU: 0.867648), Class: 0.999904, Obj: 0.975032, No Obj: 0.000677, .5R: 1.000000, .75R: 1.000000, count: 7
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.781184, GIOU: 0.778921), Class: 0.999764, Obj: 0.854608, No Obj: 0.000164, .5R: 1.000000, .75R: 0.666667, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.917590, GIOU: 0.916560), Class: 0.999819, Obj: 0.994249, No Obj: 0.000950, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.840999, GIOU: 0.834302), Class: 0.999719, Obj: 0.951601, No Obj: 0.000755, .5R: 1.000000, .75R: 0.714286, count: 7
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.827223, GIOU: 0.823592), Class: 0.999732, Obj: 0.644256, No Obj: 0.000074, .5R: 1.000000, .75R: 0.666667, count: 3
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000233, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.838078, GIOU: 0.837860), Class: 0.999717, Obj: 0.971552, No Obj: 0.000375, .5R: 1.000000, .75R: 1.000000, count: 3
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.840294, GIOU: 0.835282), Class: 0.997959, Obj: 0.265230, No Obj: 0.000010, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000047, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.883361, GIOU: 0.881718), Class: 0.999836, Obj: 0.991195, No Obj: 0.000858, .5R: 1.000000, .75R: 1.000000, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.826935, GIOU: 0.820828), Class: 0.999917, Obj: 0.748244, No Obj: 0.000063, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.816095, GIOU: 0.808041), Class: 0.999662, Obj: 0.978934, No Obj: 0.001594, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.788542, GIOU: 0.782187), Class: 0.999452, Obj: 0.979136, No Obj: 0.000612, .5R: 1.000000, .75R: 0.833333, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.864566, GIOU: 0.859319), Class: 0.999375, Obj: 0.407669, No Obj: 0.000028, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000001, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.840736, GIOU: 0.838519), Class: 0.999748, Obj: 0.875304, No Obj: 0.000537, .5R: 1.000000, .75R: 0.800000, count: 5
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: -nan, GIOU: -nan), Class: -nan, Obj: -nan, No Obj: 0.000001, .5R: -nan, .75R: -nan, count: 0
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.827050, GIOU: 0.824066), Class: 0.998728, Obj: 0.460583, No Obj: 0.000486, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.898753, GIOU: 0.897707), Class: 0.999815, Obj: 0.871375, No Obj: 0.000473, .5R: 1.000000, .75R: 1.000000, count: 4
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.546495, GIOU: 0.494828), Class: 0.977138, Obj: 0.350278, No Obj: 0.000027, .5R: 0.500000, .75R: 0.500000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.878498, GIOU: 0.875214), Class: 0.999045, Obj: 0.600723, No Obj: 0.000620, .5R: 1.000000, .75R: 1.000000, count: 2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.869378, GIOU: 0.867205), Class: 0.999579, Obj: 0.939749, No Obj: 0.000715, .5R: 1.000000, .75R: 1.000000, count: 6
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.817488, GIOU: 0.814568), Class: 0.999738, Obj: 0.601861, No Obj: 0.000276, .5R: 1.000000, .75R: 1.000000, count: 9
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 82 Avg (IOU: 0.889346, GIOU: 0.886728), Class: 0.999476, Obj: 0.954856, No Obj: 0.000708, .5R: 1.000000, .75R: 1.000000, count: 1
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 94 Avg (IOU: 0.858460, GIOU: 0.854488), Class: 0.999810, Obj: 0.926675, No Obj: 0.000816, .5R: 1.000000, .75R: 1.000000, count: 8
v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 106 Avg (IOU: 0.831838, GIOU: 0.831819), Class: 0.999879, Obj: 0.682484, No Obj: 0.000056, .5R: 1.000000, .75R: 1.000000, count: 1
^C

wiwao@wiwao-desktop:~/darknet$

これで一旦終了。

一応学習再開の方法を試す。。。

wiwao@wiwao-desktop:~/darknet$ ./darknet detector train cfg/obj.data cfg/yolo-obj.cfg darknet53.conv.74 -map

Prepare additional network for mAP calculation...

で同じコマンドを使ってしまったので途中で気がついてストップ。。。

なのでyolo-obj_1000.weightsは間違ったところで上書きされていると思います。

なので使えない。

./darknet detector train cfg/obj.data cfg/yolo-obj.cfg backup/yolo-obj_14000.weights

で再スタート

このプロセス作った人は天才ですね。

作業を理解するだけでも大変です。

これで学習フェイズ終了

盤メーカーとIoTそしてAI

2019年10月19日土曜日

学習フェイズ

0 件のコメント:

コメントを投稿