yoloV3引數理解及註釋
阿新 • • 發佈：2019-01-27
不對的地方求各位糾正
[net] # Testing #batch=1 #subdivisions=1 # Training batch=64 一批訓練樣本的樣本數量，每batch個樣本更新一次引數 subdivisions=64 batch/subdivisions作為一次性送入訓練器的樣本數量 如果記憶體不夠大，將batch分割為subdivisions個子batch 上面這兩個引數如果電腦記憶體小，則把batch改小一點，batch越大，訓練效果越好 subdivisions越大，可以減輕顯示卡壓力 width=416 height=416 channels=3 以上三個引數為輸入影象的引數資訊 width和height影響網路對輸入影象的解析度， 從而影響precision，只可以設定成32的倍數 momentum=0.9 DeepLearning1中最優化方法中的動量引數，這個值影響著梯度下降到最優值得速度 decay=0.0005 權重衰減正則項，防止過擬合 angle=0 通過旋轉角度來生成更多訓練樣本 saturation = 1.5 通過調整飽和度來生成更多訓練樣本 exposure = 1.5 通過調整曝光量來生成更多訓練樣本 hue=.1 通過調整色調來生成更多訓練樣本 learning_rate=0.001 學習率決定著權值更新的速度，設定得太大會使結果超過最優值，太小會使下降速度過慢。 如果僅靠人為干預調整引數，需要不斷修改學習率。剛開始訓練時可以將學習率設定的高一點， 而一定輪數之後，將其減小 在訓練過程中，一般根據訓練輪數設定動態變化的學習率。 剛開始訓練時：學習率以 0.01 ~ 0.001 為宜。 一定輪數過後：逐漸減緩。 接近訓練結束：學習速率的衰減應該在100倍以上。 學習率的調整參考https://blog.csdn.net/qq_33485434/article/details/80452941 burn_in=1000 在迭代次數小於burn_in時，其學習率的更新有一種方式，大於burn_in時，才採用policy的更新方式 max_batches = 500200 訓練達到max_batches後停止學習 policy=steps 這個是學習率調整的策略，有policy：constant, steps, exp, poly, step, sig, RANDOM，constant等方式 參考https://nanfei.ink/2018/01/23/YOLOv2%E8%B0%83%E5%8F%82%E6%80%BB%E7%BB%93/#more steps=40000,45000 下面這兩個引數steps和scale是設定學習率的變化，比如迭代到40000次時，學習率衰減十倍。 scales=.1,.1 45000次迭代時，學習率又會在前一個學習率的基礎上衰減十倍 [convolutional] batch_normalize=1 ？ filters=32 輸出特徵圖的數量 size=3 卷積核的尺寸 stride=1 做卷積運算的步長 pad=1 如果pad為0,padding由 padding引數指定。 如果pad為1，padding大小為size/2，padding應該是對輸入影象左邊緣拓展的畫素數量 activation=leaky 啟用函式的型別 # Downsample [convolutional] batch_normalize=1 filters=64 size=3 stride=2 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=32 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=64 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear # Downsample [convolutional] batch_normalize=1 filters=128 size=3 stride=2 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=64 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=128 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=64 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=128 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear # Downsample [convolutional] batch_normalize=1 filters=256 size=3 stride=2 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear # Downsample [convolutional] batch_normalize=1 filters=512 size=3 stride=2 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear # Downsample [convolutional] batch_normalize=1 filters=1024 size=3 stride=2 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=512 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=1024 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=512 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=1024 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=512 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=1024 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear [convolutional] batch_normalize=1 filters=512 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 filters=1024 size=3 stride=1 pad=1 activation=leaky [shortcut] from=-3 activation=linear ###################### [convolutional] batch_normalize=1 filters=512 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=1024 activation=leaky [convolutional] batch_normalize=1 filters=512 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=1024 activation=leaky [convolutional] batch_normalize=1 filters=512 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=1024 activation=leaky [convolutional] size=1 stride=1 pad=1 filters=18 每一個[region/yolo]層前的最後一個卷積層中的 filters=num(yolo層個數)*(classes+5) 5的意義是5個座標，論文中的tx,ty,tw,th,to activation=linear [yolo] 在yoloV2中yolo層叫region層 mask = 6,7,8 anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 anchors是可以事先通過cmd指令計算出來的，是和圖片數量，width,height以及cluster(應該就是下面的num的值， 即想要使用的anchors的數量)相關的預選框，可以手工挑選，也可以通過k means 從訓練樣本中學出 classes=1 num=9 jitter=.3 ignore_thresh = .5 truth_thresh = 1 random=1 random設定成1，可以增加檢測精度precision [route] layers = -4 [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [upsample] stride=2 [route] layers = -1, 61 [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=512 activation=leaky [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=512 activation=leaky [convolutional] batch_normalize=1 filters=256 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=512 activation=leaky [convolutional] size=1 stride=1 pad=1 filters=18 activation=linear [yolo] mask = 3,4,5 anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 classes=1 num=9 每個grid cell預測幾個box,和anchors的數量一致。當想要使用更多anchors時需要調大num， 且如果調大num後訓練時Obj趨近0的話可以嘗試調大object_scale jitter=.3 利用資料抖動產生更多資料，YOLOv2中使用的是crop，filp，以及net層的angle，flip是隨機的， jitter就是crop的引數，tiny-yolo-voc.cfg中jitter=.3，就是在0~0.3中進行crop ignore_thresh = .5 決定是否需要計算IOU誤差的引數，大於thresh，IOU誤差不會夾在cost function中 truth_thresh = 1 random=1 如果為1，每次迭代圖片大小隨機從320到608，步長為32，如果為0，每次訓練大小與輸入大小一致 [route] layers = -4 [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [upsample] stride=2 [route] layers = -1, 36 [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=256 activation=leaky [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=256 activation=leaky [convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky [convolutional] batch_normalize=1 size=3 stride=1 pad=1 filters=256 activation=leaky [convolutional] size=1 stride=1 pad=1 filters=18 activation=linear [yolo] mask = 0,1,2 anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 classes=1 num=9 jitter=.3 ignore_thresh = .5 truth_thresh = 1 random=1 官網介紹中的關鍵點： 圖片樣本的獲取可以自己找，也可以下載相關資料集，coco,voc資料集中可能也含有相關能用的樣本 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. When should I stop training: When you see that average loss 0.xxxxxx avg no longer decreases at many iterations then you should stop training. Once training is stopped, you should take some of last .weights-files from darknet\build\darknet\x64\backup and choose the best of them. Overfitting - is case when you can detect objects on images from training-dataset, but can't detect objects on any others images. You should get weights from Early Stopping Point. IoU (intersect of union) - average instersect of union of objects and detections for a certain threshold = 0.24 How to improve object detection: Before training: set flag random=1 in your .cfg-file - it will increase precision by training Yolo for different resolutions. increase network resolution in your .cfg-file (height=608, width=608 or any value multiple of 32) - it will increase precision. recalculate anchors for your dataset for width and height from cfg-file: darknet.exe detector calc_anchors data/obj.data -num_of_clusters 9 -width 416 -height 416 then set the same 9 anchors in each of 3 [yolo]-layers in your cfg-file 設定錨點 desirable that your training dataset include images with objects at diffrent: scales, rotations, lightings, from different sides, on different backgrounds 樣本特點儘量多樣化，亮度，旋轉，背景，目標位置，尺寸 desirable that your training dataset include images with non-labeled objects that you do not want to detect - negative samples without bounded box (empty .txt files) 可以新增沒有標註框的圖片和其空的txt檔案，作為negative資料 for training with a large number of objects in each image, add the parameter max=200 or higher value in the last layer [region] in your cfg-file to speedup training (with decreasing detection accuracy) do Fine-Tuning instead of Transfer-Learning, set param stopbackward=1 in one of the penultimate convolutional layers before the 1-st [yolo]-layer, for example here: https://github.com/AlexeyAB/darknet/blob/0039fd26786ab5f71d5af725fc18b3f521e7acfd/cfg/yolov3.cfg#L598 可以在第一個[yolo]層之前的倒數第二個[convolutional]層末尾新增 stopbackward=1,以此提升訓練速度 After training - for detection: Increase network-resolution by set in your .cfg-file (height=608 and width=608) or (height=832 and width=832) or (any value multiple of 32) - this increases the precision and makes it possible to detect small objects, you do not need to train the network again, just use .weights-file already trained for 416x416 resolution. 即使在用416*416訓練完之後，也可以在cfg檔案中設定較大的width和height，增加網路對影象的解析度，從而更可能檢測出影象中的小目標，而不需要重新訓練 if error Out of memory occurs then in .cfg-file you should increase subdivisions=16, 32 or 64 Out of memory的錯誤需要通過增大subdivisions來解決官網介紹中的關鍵點： 圖片樣本的獲取可以自己找，也可以下載相關資料集，coco,voc資料集中可能也含有相關能用的樣本 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. When should I stop training: When you see that average loss 0.xxxxxx avg no longer decreases at many iterations then you should stop training. Once training is stopped, you should take some of last .weights-files from darknet\build\darknet\x64\backup and choose the best of them. Overfitting - is case when you can detect objects on images from training-dataset, but can't detect objects on any others images. You should get weights from Early Stopping Point. IoU (intersect of union) - average instersect of union of objects and detections for a certain threshold = 0.24 How to improve object detection: Before training: set flag random=1 in your .cfg-file - it will increase precision by training Yolo for different resolutions. increase network resolution in your .cfg-file (height=608, width=608 or any value multiple of 32) - it will increase precision. recalculate anchors for your dataset for width and height from cfg-file: darknet.exe detector calc_anchors data/obj.data -num_of_clusters 9 -width 416 -height 416 then set the same 9 anchors in each of 3 [yolo]-layers in your cfg-file 設定錨點 desirable that your training dataset include images with objects at diffrent: scales, rotations, lightings, from different sides, on different backgrounds 樣本特點儘量多樣化，亮度，旋轉，背景，目標位置，尺寸 desirable that your training dataset include images with non-labeled objects that you do not want to detect - negative samples without bounded box (empty .txt files) 可以新增沒有標註框的圖片和其空的txt檔案，作為negative資料 for training with a large number of objects in each image, add the parameter max=200 or higher value in the last layer [region] in your cfg-file to speedup training (with decreasing detection accuracy) do Fine-Tuning instead of Transfer-Learning, set param stopbackward=1 in one of the penultimate convolutional layers before the 1-st [yolo]-layer, for example here: https://github.com/AlexeyAB/darknet/blob/0039fd26786ab5f71d5af725fc18b3f521e7acfd/cfg/yolov3.cfg#L598 可以在第一個[yolo]層之前的倒數第二個[convolutional]層末尾新增 stopbackward=1,以此提升訓練速度 After training - for detection: Increase network-resolution by set in your .cfg-file (height=608 and width=608) or (height=832 and width=832) or (any value multiple of 32) - this increases the precision and makes it possible to detect small objects, you do not need to train the network again, just use .weights-file already trained for 416x416 resolution. 即使在用416*416訓練完之後，也可以在cfg檔案中設定較大的width和height，增加網路對影象的解析度，從而更可能檢測出影象中的小目標，而不需要重新訓練 if error Out of memory occurs then in .cfg-file you should increase subdivisions=16, 32 or 64 Out of memory的錯誤需要通過增大subdivisions來解決
參考文章：yolo引數
[net] batch=64 一批次用的影象數量，一般來說，在視訊記憶體允許範圍內儘量的大會收斂更精細，不過也可能導致收斂到區域性最優解 subdivisions=4 劃分子集 width=416 height=416 網路輸入尺寸 channels=3 影象通道數 momentum=0.9 SGD方法的一個缺點是其更新方向完全依賴於當前batch計算出的梯度，因而十分不穩定。Momentum演算法借用了物理中的動量概念，它模擬的是物體運動時的慣性，即更新的時候在一定程度上保留之前更新的方向，同時利用當前batch的梯度微調最終的更新方向。這樣一來，可以在一定程度上增加穩定性，從而學習地更快，並且還有一定擺脫區域性最優的能力：
vt=γ⋅vt−1+α⋅▽ΘJ(Θ)vt=γ⋅vt−1+α⋅▽ΘJ(Θ)
Θ=Θ−vtΘ=Θ−vt
Momentum演算法會觀察歷史梯度vt−1，若當前梯度的方向與歷史梯度一致（表明當前樣本不太可能為異常點），則會增強這個方向的梯度，若當前梯度與歷史梯方向不一致，則梯度會衰減。一種形象的解釋是：我們把一個球推下山，球在下坡時積聚動量，在途中變得越來越快，γ可視為空氣阻力，若球的方向發生變化，則動量會衰減。 對於權重來說公式如下： 
這裡的m就是我們可以調整的引數，一般取值有0.5,0.9,0.99，當然,也可以讓α的值隨著時間而變化,一開始小點,後來再加大。 
特點: 前後梯度方向一致時,能夠加速學習 前後梯度方向不一致時,能夠抑制震盪 decay=0.0005 在實際應用中，為了避免網路的過擬合，必須對價值函式（Cost function）加入一些正則項，在SGD中加入正則項對這個Cost function進行規範化： 
上面這個公式基本思想就是減小不重要的引數對最後結果的影響，網路中有用的權重則不會收到Weight decay影響。 可以看到，此項越大則防止過擬合的能力越強。 angle=0 資料擴充時圖片旋轉的角度 saturation = 1.5 飽和度範圍 exposure = 1.5 曝光度範圍 hue=.1 色調變化範圍 learning_rate=0.001 初始學習率 max_batches = 200000 訓練達到max_batches後停止學習 policy=steps 調整學習率的policy，有如下policy：constant, steps, exp, poly, step, sig, RANDOM constant 保持學習率為常量，caffe裡為fixed steps 比較好理解，按照steps來改變學習率 
steps=-1,400,100000,150000 scales=.1,10,.1,.1 steps和scales是對應的 有人認為，使用自己的資料集訓練時很大概概率需要這個warm up的trick，當然這個也要看具體的資料集 Warmup method: Constant warmup: 在train前幾個epoch（一般前5 epochs）時採用較小的constant learning rate，但是對於大的learning rate，constant warmup不能很好的初始化網路。 gradual warmup： 在 training的前幾個 epoch，逐漸將learning rate由小到大的提高，讓training在開始的時候健康的收斂。 重要參考：https://arxiv.org/abs/1706.02677 exp gamma= 返回base_lr*gamma^iter,iter為當前迭代次數，gamma設定為0.98 
poly power=4 max_batches=800000 
對學習率進行多項式衰減。圖中power為0.9 sig 學習率進行sigmod函式衰減 gamma= 0.05 step=200 效果如圖所示 
step 返回net.learning_rate*pow(net.scale, batch_num/net.step)
[convolutional] batch_normalize=1 是否做BN filters=32 輸出多少個特徵圖 size=3 卷積核的尺寸 stride=1 做卷積運算的步長 pad=1 如果pad為0,padding由padding引數指定。如果pad為1，padding大小為size/2 activation=leaky 啟用函式有以下幾種：logistic，loggy，relu，elu，relie，plse，hardtan，lhtan，linear，ramp，leaky，tanh，stair 一般來說，現在用leaky的會比較多 
[maxpool] size=2 池化層尺寸 stride=2 池化步長
特別：最後一個卷積層 [convolutional] size=1 stride=1 pad=1 filters=125 region前最後一個卷積層的filters數是特定的，計算公式為filter=num*(classes+5)，5的意義是5個座標，論文中的tx,ty,tw,th,to activation=linear
[region] anchors = 1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52 預測框的初始寬高，第一個是w，第二個是h，總數量是num*2 聚類的指令碼放在github中。 bias_match=1 只是在選擇與groundturth_box寬、高最相近的anchors，然後在選定的anchors基礎之上進行寬、高的偏差調節。最終的預測框寬、高不一定與anchors一致。 classes=20 類別數 coords=4 BoundingBox的tx,ty,tw,th，tx與ty是相對於左上角的gird，同時是當前grid的比例，tw與th是寬度與高度取對數 num=5 每個grid預測的BoundingBox個數 softmax=1 jitter=.2 利用資料抖動產生更多資料，YOLOv2中使用的是crop，filp，以及net層的angle，flip是隨機的，crop就是jitter的引數，tiny-yolo-voc.cfg中jitter=.2，就是在0~0.2中進行crop rescore=1 決定使用哪種方式計算IOU的誤差，為1時，使用當前best iou計算，為0時，使用1計算 object_scale=5 noobject_scale=1 class_scale=1 coord_scale=1 loss的係數 absolute=1 thresh = .6 決定是否需要計算IOU誤差的引數，大於thresh，IOU誤差不會夾在cost function中 random=1 如果為1每次迭代圖片大小隨機從320到608，步長為32，如果為0，每次訓練大小與輸入大小一致