1.導(dǎo)入需要的包和基本配置

importargparse#解析命令行參數(shù)模塊
importjson#字典列表和JSON字符串之間的相互解析模塊
importos#與操作系統(tǒng)進(jìn)行交互的模塊包含文件路徑操作和解析
importsys#sys系統(tǒng)模塊包含了與Python解釋器和它的環(huán)境有關(guān)的函數(shù)
frompathlibimportPath#Path將str轉(zhuǎn)換為Path對(duì)象使字符串路徑易于操作的模塊

importnumpyasnp#NumPy（NumericalPython）是Python的一種開源的數(shù)值計(jì)算擴(kuò)展
importoneflowasflow#OneFlow深度學(xué)習(xí)框架
fromtqdmimporttqdm#進(jìn)度條模塊

frommodels.commonimportDetectMultiBackend#下面都是one-yolov5定義的模塊，在本系列的其它文章都有涉及
fromutils.callbacksimportCallbacks
fromutils.dataloadersimportcreate_dataloader
fromutils.generalimport(
LOGGER,
check_dataset,
check_img_size,
check_requirements,
check_yaml,
coco80_to_coco91_class,
colorstr,
increment_path,
non_max_suppression,
print_args,
scale_coords,
xywh2xyxy,
xyxy2xywh,
)
fromutils.metricsimportConfusionMatrix,ap_per_class,box_iou
fromutils.oneflow_utilsimportselect_device,time_sync
fromutils.plotsimportoutput_to_target,plot_images,plot_val_study

2.opt參數(shù)詳解

3.main函數(shù)

根據(jù)解析的opt參數(shù)，調(diào)用run函數(shù)

defmain(opt):
#檢測(cè)requirements文件中需要的包是否安裝好了
check_requirements(requirements=ROOT/"requirements.txt",exclude=("tensorboard","thop"))

ifopt.taskin("train","val","test"):#runnormally
ifopt.conf_thres>0.001:#更多請(qǐng)見https://github.com/ultralytics/yolov5/issues/1466
LOGGER.info(f"WARNING:confidencethreshold{opt.conf_thres}>0.001producesinvalidresults")
run(**vars(opt))

else:
weights=opt.weightsifisinstance(opt.weights,list)else[opt.weights]
opt.half=True#FP16forfastestresults
ifopt.task=="speed":#speedbenchmarks
#pythonval.py--taskspeed--datacoco.yaml
#--batch1--weightsyolov5n/yolov5s/...
opt.conf_thres,opt.iou_thres,opt.save_json=0.25,0.45,False
foropt.weightsinweights:
run(**vars(opt),plots=False)

elifopt.task=="study":#speedvsmAPbenchmarks
#pythonval.py--taskstudy--datacoco.yaml
#--iou0.7--weightsyolov5n/yolov5s/...
foropt.weightsinweights:
f=f"study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt"
x,y=(
list(range(256,1536+128,128)),
[],
)#xaxis(imagesizes),yaxis
#"study":模型在各個(gè)尺度下的指標(biāo)并可視化，
#上面list(range(256,1536+128,128)),代表img-size的各個(gè)尺度,具體代碼如下：
foropt.imgszinx:#img-size
LOGGER.info(f"
Running{f}--imgsz{opt.imgsz}...")
r,_,t=run(**vars(opt),plots=False)
y.append(r+t)#resultsandtimes
np.savetxt(f,y,fmt="%10.4g")#save
os.system("zip-rstudy.zipstudy_*.txt")
#可視化各個(gè)指標(biāo)
plot_val_study(x=x)#plot

3. run函數(shù)

3.1 載入?yún)?shù)

#不參與反向傳播
@flow.no_grad()
defrun(
data,#數(shù)據(jù)集配置文件地址包含數(shù)據(jù)集的路徑、類別個(gè)數(shù)、類名、下載地址等信息train.py時(shí)傳入data_dict
weights=None,#模型的權(quán)重文件地址運(yùn)行train.py=None運(yùn)行test.py=默認(rèn)weights/yolov5s
batch_size=32,#前向傳播的批次大小運(yùn)行test.py傳入默認(rèn)32運(yùn)行train.py則傳入batch_size//WORLD_SIZE*2
imgsz=640,#輸入網(wǎng)絡(luò)的圖片分辨率運(yùn)行test.py傳入默認(rèn)640運(yùn)行train.py則傳入imgsz_test
conf_thres=0.001,#object置信度閾值默認(rèn)0.001
iou_thres=0.6,#進(jìn)行NMS時(shí)IOU的閾值默認(rèn)0.6
task="val",#設(shè)置測(cè)試的類型有train,val,test,speedorstudy幾種默認(rèn)val
device="",#執(zhí)行val.py所在的設(shè)備cudadevice,i.e.0or0,1,2,3orcpu
workers=8,#dataloader中的最大worker數(shù)（線程個(gè)數(shù)）
single_cls=False,#數(shù)據(jù)集是否只有一個(gè)類別默認(rèn)False
augment=False,#測(cè)試時(shí)增強(qiáng)，詳細(xì)請(qǐng)看我們的教程：https://start.oneflow.org/oneflow-yolo-doc/tutorials/03_chapter/TTA.html
verbose=False,#是否打印出每個(gè)類別的mAP運(yùn)行test.py傳入默認(rèn)Fasle運(yùn)行train.py則傳入nc

	

	3.2 Initialize/load model and set device（初始化/加載模型以及設(shè)置設(shè)備）

	
iftraining:#通過train.py調(diào)用的run函數(shù)
device,of,engine=(
next(model.parameters()).device,
True,
False,
)#getmodeldevice,OneFlowmodel
half&=device.type!="cpu"#halfprecisiononlysupportedonCUDA
model.half()ifhalfelsemodel.float()
else:#直接通過val.py調(diào)用run函數(shù)
device=select_device(device,batch_size=batch_size)

#Directories生成save_dir文件路徑run/val/expn
save_dir=increment_path(Path(project)/name,exist_ok=exist_ok)#incrementrun
(save_dir/"labels"ifsave_txtelsesave_dir).mkdir(parents=True,exist_ok=True)#makedir

#加載模型只在運(yùn)行val.py才需要自己加載model
model=DetectMultiBackend(weights,device=device,dnn=dnn,data=data,fp16=half)

stride,of,engine=model.stride,model.of,model.engine
#檢測(cè)輸入圖片的分辨率imgsz是否能被stride整除
imgsz=check_img_size(imgsz,s=stride)#checkimagesize
half=model.fp16#FP16supportedonlimitedbackendswithCUDA
ifengine:
batch_size=model.batch_size
else:
device=model.device
ifnotof:
batch_size=1#export.pymodelsdefaulttobatch-size1
LOGGER.info(f"Forcing--batch-size1inference(1,3,{imgsz},{imgsz})fornon-OneFlowmodels")

#Data
data=check_dataset(data)#check


	

	3.3 Configure

	
#配置
model.eval()#啟動(dòng)模型驗(yàn)證模式
cuda=device.type!="cpu"
is_coco=isinstance(data.get("val"),str)anddata["val"].endswith(f"coco{os.sep}val2017.txt")#通過COCO數(shù)據(jù)集的文件夾組織結(jié)構(gòu)判斷當(dāng)前數(shù)據(jù)集是否為COCO數(shù)據(jù)集
nc=1ifsingle_clselseint(data["nc"])#numberofclasses
#設(shè)置iou閾值從0.5-0.95取10個(gè)(0.05間隔)iouvectorformAP@0.5:0.95
#iouv:[0.50000,0.55000,0.60000,0.65000,0.70000,0.75000,0.80000,0.85000,0.90000,0.95000]
iouv=flow.linspace(0.5,0.95,10,device=device)#iouvectorformAP@0.5:0.95
niou=iouv.numel()#示例mAP@0.5:0.95iou閾值個(gè)數(shù)=10個(gè)，計(jì)算mAP的詳細(xì)教程可以在https://start.oneflow.org/oneflow-yolo-doc/tutorials/05_chapter/map_analysis.html這里查看


	

	3.4 Dataloader

	通過 train.py 調(diào)用 run 函數(shù)會(huì)傳入一個(gè) Dataloader，而通過 val.py 需要加載測(cè)試數(shù)據(jù)集

	
#Dataloader
#如果不是訓(xùn)練(執(zhí)行val.py腳本調(diào)用run函數(shù))就調(diào)用create_dataloader生成dataloader
#如果是訓(xùn)練(執(zhí)行train.py調(diào)用run函數(shù))就不需要生成dataloader可以直接從參數(shù)中傳過來testloader
ifnottraining:#加載val數(shù)據(jù)集
ifofandnotsingle_cls:#check--weightsaretrainedon--data
ncm=model.model.nc
assertncm==nc,(
f"{weights}({ncm}classes)trainedondifferent--datathanwhatyoupassed({nc}"f"classes).Passcorrectcombinationof"f"--weightsand--datathataretrainedtogether."
)
model.warmup(imgsz=(1ifofelsebatch_size,3,imgsz,imgsz))#warmup
pad=0.0iftaskin("speed","benchmark")else0.5
rect=Falseiftask=="benchmark"elseof#squareinferenceforbenchmarks
task=taskiftaskin("train","val","test")else"val"#pathtotrain/val/testimages
#創(chuàng)建dataloader這里的rect默認(rèn)為True矩形推理用于測(cè)試集在不影響mAP的情況下可以大大提升推理速度
dataloader=create_dataloader(
data[task],
imgsz,
batch_size,
stride,
single_cls,
pad=pad,
rect=rect,
workers=workers,
prefix=colorstr(f"{task}:"),
)[0]


	

	3.5 初始化

	
#初始化驗(yàn)證的圖片的數(shù)量
seen=0
#初始化混淆矩陣
confusion_matrix=ConfusionMatrix(nc=nc)

#獲取數(shù)據(jù)集所有目標(biāo)類別的類名
names=dict(enumerate(model.namesifhasattr(model,"names")elsemodel.module.names))

#coco80_to_coco91_class:converts80-index(val2014)to91-index(paper)
#https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/
class_map=coco80_to_coco91_class()ifis_cocoelselist(range(1000))
#設(shè)置進(jìn)度條模塊顯示信息
s=("%20s"+"%11s"*6)%(
"Class",
"Images",
"Labels",
"P",
"R",
"mAP@.5",
"mAP@.5:.95",
)
#初始化時(shí)間dt[t0（預(yù)處理的時(shí)間）,t1（推理的時(shí)間）,t2（后處理的時(shí)間）]和p,r,f1,mp,mr,map50,map指標(biāo)
dt,p,r,f1,mp,mr,map50,map=(
[0.0,0.0,0.0],
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
)
#初始化驗(yàn)證集的損失
loss=flow.zeros(3,device=device)
#初始化json文件中的字典，統(tǒng)計(jì)信息，ap，ap_class
jdict,stats,ap,ap_class=[],[],[],[]
callbacks.run("on_val_start")
#初始化tqdm進(jìn)度條模塊
pbar=tqdm(dataloader,desc=s,bar_format="{l_bar}{bar:10}{r_bar}{bar:-10b}")

示例輸出
val:data=data/coco.yaml,weights=['yolov5x'],batch_size=32,imgsz=640,conf_thres=0.001,iou_thres=0.6,task=val,
device=,workers=8,single_cls=False,augment=False,verbose=False,save_txt=False,save_hybrid=False,
save_conf=False,save_json=True,project=runs/val,name=exp,exist_ok=False,half=True,dnn=False
YOLOv5v1.0-8-g94ec5c4Python-3.8.13oneflow-0.8.1.dev20221018+cu112
Fusinglayers...
Modelsummary:322layers,86705005parameters,571965gradients
val:Scanning'/data/dataset/fengwen/coco/val2017.cache'imagesandlabels...4952found,48missing,0empty,0corrupt:100%|████████
ClassImagesLabelsPRmAP@.5mAP@.5:.95:100%|██████████|157/157[01:55<00:00,??1.36it/
?????????????????all???????5000??????36335??????0.743??????0.627??????0.685??????0.503
Speed:?0.1ms?pre-process,?7.5ms?inference,?2.1ms?NMS?per?image?at?shape?(32,?3,?640,?640)??#?<---?baseline?speed

Evaluating?pycocotools?mAP...?saving?runs/val/exp3/yolov5x_predictions.json...

...
Average?Precision??(AP)?@[?IoU=0.50:0.95?|?area=???all?|?maxDets=100?]?=?0.505?#?<---?baseline?mAP
?Average?Precision??(AP)?@[?IoU=0.50??????|?area=???all?|?maxDets=100?]?=?0.689
?Average?Precision??(AP)?@[?IoU=0.75??????|?area=???all?|?maxDets=100?]?=?0.545
?Average?Precision??(AP)?@[?IoU=0.50:0.95?|?area=?small?|?maxDets=100?]?=?0.339
?Average?Precision??(AP)?@[?IoU=0.50:0.95?|?area=medium?|?maxDets=100?]?=?0.557
?Average?Precision??(AP)?@[?IoU=0.50:0.95?|?area=?large?|?maxDets=100?]?=?0.650
?Average?Recall?????(AR)?@[?IoU=0.50:0.95?|?area=???all?|?maxDets=??1?]?=?0.382
?Average?Recall?????(AR)?@[?IoU=0.50:0.95?|?area=???all?|?maxDets=?10?]?=?0.628
?Average?Recall?????(AR)?@[?IoU=0.50:0.95?|?area=???all?|?maxDets=100?]?=?0.677??#?<---?baseline?mAR
?Average?Recall?????(AR)?@[?IoU=0.50:0.95?|?area=?small?|?maxDets=100?]?=?0.523
?Average?Recall?????(AR)?@[?IoU=0.50:0.95?|?area=medium?|?maxDets=100?]?=?0.730
?Average?Recall?????(AR)?@[?IoU=0.50:0.95?|?area=?large?|?maxDets=100?]?=?0.826


	

	3.6 開始驗(yàn)證

	
forbatch_i,(im,targets,paths,shapes)inenumerate(pbar):
"""https://github.com/Oneflow-Inc/one-yolov5/blob/bf8c66e011fcf5b8885068074ffc6b56c113a20c/utils/dataloaders.py#L735
im:flow.from_numpy(img);
targets:labels_out
paths:self.im_files[index]
shapes:shapes
"""


	

	3.6.1 驗(yàn)證開始前的預(yù)處理

	
callbacks.run("on_val_batch_start")
t1=time_sync()
ifcuda:
im=im.to(device)
targets=targets.to(device)
im=im.half()ifhalfelseim.float()#uint8tofp16/32
im/=255#0-255to0.0-1.0
nb,_,height,width=im.shape#batchsize,channels,height,width
t2=time_sync()
dt[0]+=t2-t1


	

	3.6.2 推理

	
#Inference
out,train_out=model(im)iftrainingelsemodel(im,augment=augment,val=True)#輸出為：推理結(jié)果、損失值
dt[1]+=time_sync()-t2


	

	3.6.3 計(jì)算損失

	
#Loss
"""
分類損失(cls_loss)：該損失用于判斷模型是否能夠準(zhǔn)確地識(shí)別出圖像中的對(duì)象，并將其分類到正確的類別中。

置信度損失(obj_loss)：該損失用于衡量模型預(yù)測(cè)的框（即包含對(duì)象的矩形）與真實(shí)框之間的差異。

邊界框損失(box_loss)：該損失用于衡量模型預(yù)測(cè)的邊界框與真實(shí)邊界框之間的差異，這有助于確保模型能夠準(zhǔn)確地定位對(duì)象。
"""
ifcompute_loss:
loss+=compute_loss([x.float()forxintrain_out],targets)[1]#box,obj,cls


	

	3.6.4 Run NMS

	
#NMS
#將真實(shí)框target的xywh(因?yàn)閠arget是在labelimg中做了歸一化的)映射到真實(shí)的圖像(test)尺寸
targets[:,2:]*=flow.tensor((width,height,width,height),device=device)#topixels
#在NMS之前將數(shù)據(jù)集標(biāo)簽targets添加到模型預(yù)測(cè)中，這允許在數(shù)據(jù)集中自動(dòng)標(biāo)記(forautolabelling)其它對(duì)象(在pred中混入gt)并且mAP反映了新的混合標(biāo)簽
#targets:[num_target,img_index+class_index+xywh]=[31,6]
#lb:{list:bs}第一張圖片的target[17,5]第二張[1,5]第三張[7,5]第四張[6,5]
lb=[targets[targets[:,0]==i,1:]foriinrange(nb)]ifsave_hybridelse[]#forautolabelling
t3=time_sync()
"""non_max_suppression(非最大值抑制)
Non-MaximumSuppression(NMS)oninferenceresultstorejectoverlappingboundingboxes
該算法的原理：
先假設(shè)有6個(gè)矩形框，根據(jù)分類器的類別分類概率大小排序，假設(shè)從小到大屬于車輛(被檢測(cè)的目標(biāo))的概率分別為：A、B、C、D、E、F
（1）從最大概率矩形框F開始，分別判斷A~E與F的重疊度IOU是否大于某個(gè)指定的閥值；
（2）假設(shè)B、D與F的重疊度大于指定的閥值，則丟棄B、D，并標(biāo)記第一個(gè)矩形框F，是我們要保留的
（3）從剩下的矩形框A、C、E中，選擇最大概率，假設(shè)為E，然后判斷A、C與E的重疊度是否大于指定的閥值，
假如大于就丟棄A、C，并標(biāo)記E，是我們保留下來的第二個(gè)矩形框
一直重復(fù)上述過程，找到所有被保留的矩形框
Returns:
listofdetections,on(n,6)tensorperimage[xyxy,conf,cls]
"""
out=non_max_suppression(out,conf_thres,iou_thres,labels=lb,multi_label=True,agnostic=single_cls)
#獲取NMS時(shí)間
dt[2]+=time_sync()-t3


	

	3.6.5 統(tǒng)計(jì)每張圖片的真實(shí)框、預(yù)測(cè)框信息

	
#為每張圖片做統(tǒng)計(jì)，寫入預(yù)測(cè)信息到txt文件，生成json文件字典，統(tǒng)計(jì)tp等
#out:list{bs}[300,6][42,6][300,6][300,6][:,image_index+class+xywh]
forsi,predinenumerate(out):
#獲取第si張圖片的gt標(biāo)簽信息包括class,x,y,w,htarget[:,0]為標(biāo)簽屬于哪張圖片的編號(hào)
labels=targets[targets[:,0]==si,1:]#[:,class+xywh]
nl,npr=labels.shape[0],pred.shape[0]#numberoflabels,predictions
path,shape=Path(paths[si]),shapes[si][0]
correct=flow.zeros(npr,niou,dtype=flow.bool,device=device)#init
seen+=1#統(tǒng)計(jì)測(cè)試圖片數(shù)量+1

ifnpr==0:#如果預(yù)測(cè)為空，則添加空的信息到stats里
ifnl:
stats.append((correct,*flow.zeros((2,0),device=device),labels[:,0]))
ifplots:
confusion_matrix.process_batch(detections=None,labels=labels[:,0])
continue
#Predictions
ifsingle_cls:
pred[:,5]=0
predn=pred.clone()
#將預(yù)測(cè)坐標(biāo)映射到原圖img中
scale_coords(im[si].shape[1:],predn[:,:4],shape,shapes[si][1])#native-spacepred

#Evaluate
ifnl:
tbox=xywh2xyxy(labels[:,1:5])#targetboxes
scale_coords(im[si].shape[1:],tbox,shape,shapes[si][1])#native-spacelabels
labelsn=flow.cat((labels[:,0:1],tbox),1)#native-spacelabels
correct=process_batch(predn,labelsn,iouv)
ifplots:
confusion_matrix.process_batch(predn,labelsn)
stats.append((correct,pred[:,4],pred[:,5],labels[:,0]))#(correct,conf,pcls,tcls)

#Save/log
#保存預(yù)測(cè)信息到txt文件runsvalexp7labelsimage_name.txt
ifsave_txt:
save_one_txt(
predn,
save_conf,
shape,
file=save_dir/"labels"/f"{path.stem}.txt",
)
ifsave_json:
save_one_json(predn,jdict,path,class_map)#appendtoCOCO-JSONdictionary
callbacks.run("on_val_image_end",pred,predn,path,names,im[si])


	

	3.6.6 畫出前三個(gè)batch圖片的 gt 和 pred 框

	gt : 真實(shí)框，Ground truth box, 是人工標(biāo)注的位置，存放在標(biāo)注文件中

	pred : 預(yù)測(cè)框，Prediction box， 是由目標(biāo)檢測(cè)模型計(jì)算輸出的框

	
#Plotimages
ifplotsandbatch_i

	

	3.7 計(jì)算指標(biāo)

	指標(biāo)名字在代碼中體現(xiàn)

	
#Computemetrics
stats=[flow.cat(x,0).cpu().numpy()forxinzip(*stats)]#tonumpy
iflen(stats)andstats[0].any():
tp,fp,p,r,f1,ap,ap_class=ap_per_class(*stats,plot=plots,save_dir=save_dir,names=names)
ap50,ap=ap[:,0],ap.mean(1)#AP@0.5,AP@0.5:0.95
mp,mr,map50,map=p.mean(),r.mean(),ap50.mean(),ap.mean()
nt=np.bincount(stats[3].astype(int),minlength=nc)#numberoftargetsperclass


	

	3.8 打印日志

	
#Printresultsperclass
if(verboseor(nc1andlen(stats):
fori,cinenumerate(ap_class):
LOGGER.info(pf%(names[c],seen,nt[c],p[i],r[i],ap50[i],ap[i]))

#Printspeeds
t=tuple(x/seen*1e3forxindt)#speedsperimage
ifnottraining:
shape=(batch_size,3,imgsz,imgsz)
LOGGER.info(f"Speed:%.1fmspre-process,%.1fmsinference,%.1fmsNMSperimageatshape{shape}"%t)


	

	3.9 保存驗(yàn)證結(jié)果

	
#Plots
ifplots:
confusion_matrix.plot(save_dir=save_dir,names=list(names.values()))
callbacks.run("on_val_end")

#SaveJSON
ifsave_jsonandlen(jdict):
w=Path(weights[0]ifisinstance(weights,list)elseweights).stemifweightsisnotNoneelse""#weights
anno_json=str(Path(data.get("path","../coco"))/"annotations/instances_val2017.json")#annotationsjson
pred_json=str(save_dir/f"{w}_predictions.json")#predictionsjson
LOGGER.info(f"
EvaluatingpycocotoolsmAP...saving{pred_json}...")
withopen(pred_json,"w")asf:
json.dump(jdict,f)
#try-catch，會(huì)有哪些error
"""
pycocotools介紹:
https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
嘗試:
使用pycocotools工具計(jì)算loss
COCOAPI-http://cocodataset.org/
失敗error:
直接打印拋出的異常
1.可能沒有安裝pycocotools，但是網(wǎng)絡(luò)有問題，無法實(shí)現(xiàn)自動(dòng)下載。
2.pycocotools包版本有問題
"""
try:#https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
check_requirements(["pycocotools"])
frompycocotools.cocoimportCOCO
frompycocotools.cocoevalimportCOCOeval

anno=COCO(anno_json)#initannotationsapi
pred=anno.loadRes(pred_json)#initpredictionsapi
eval=COCOeval(anno,pred,"bbox")
ifis_coco:
eval.params.imgIds=[int(Path(x).stem)forxindataloader.dataset.im_files]#imageIDstoevaluate
eval.evaluate()
eval.accumulate()
eval.summarize()
map,map50=eval.stats[:2]#updateresults(mAP@0.5:0.95,mAP@0.5)
exceptExceptionase:
LOGGER.info(f"pycocotoolsunabletorun:{e}")


	

	3.10 返回結(jié)果

	
#Returnresults
model.float()#fortraining
ifnottraining:
s=f"
{len(list(save_dir.glob('labels/*.txt')))}labelssavedto{save_dir/'labels'}"ifsave_txtelse""
LOGGER.info(f"Resultssavedto{colorstr('bold',save_dir)}{s}")
maps=np.zeros(nc)+map
fori,cinenumerate(ap_class):
maps[c]=ap[i]
return(mp,mr,map50,map,*(loss.cpu()/len(dataloader)).tolist()),maps,t


	

	




	審核編輯：劉清