一覽YOLOv5中的評價(jià)方式
加入技術(shù)交流群
代碼倉庫地址：https://github.com/Oneflow-Inc/one-yolov5歡迎star one-yolov5項(xiàng)目 獲取最新的動(dòng)態(tài)。
源碼解讀：https://github.com/Oneflow-Inc/one-yolov5/blob/main/val.py 。文章里面的超鏈接可能被公眾號吃掉，可以直接到我們的文檔網(wǎng)站閱讀獲得更好的體驗(yàn)：https://start.oneflow.org/oneflow-yolo-doc/source_code_interpretation/val_py.htmlUltralytics YOLOv5 官方給的介紹:
Validate a model's accuracy on COCO val or test-dev datasets. Models are downloaded automatically from the latest YOLOv5 release. To show results by class use the --verbose flag. Note that pycocotools metrics may be ~1% better than the equivalent repo metrics, as is visible below, due to slight differences in mAP computation.1.導(dǎo)入需要的包和基本配置
import argparse # 解析命令行參數(shù)模塊
import json # 字典列表和JSON字符串之間的相互解析模塊
import os # 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
import sys # 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)
from pathlib import Path # Path將str轉(zhuǎn)換為Path對象 使字符串路徑易于操作的模塊
import numpy as np # NumPy（Numerical Python）是Python的一種開源的數(shù)值計(jì)算擴(kuò)展
import oneflow as flow # OneFlow 深度學(xué)習(xí)框架
from tqdm import tqdm # 進(jìn)度條模塊
from models.common import DetectMultiBackend # 下面都是 one-yolov5 定義的模塊，在本系列的其它文章都有涉及
from utils.callbacks import Callbacks
from utils.dataloaders import create_dataloader
from utils.general import (
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,
)
from utils.metrics import ConfusionMatrix, ap_per_class, box_iou
from utils.oneflow_utils import select_device, time_sync
from utils.plots import output_to_target, plot_images, plot_val_study
2.opt參數(shù)詳解| 參數(shù) | 解析 | |
| data | dataset.yaml path | 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 |
| weights | model weights path(s) | 模型的權(quán)重文件地址 weights/yolov5s |
| batch-size | batch size | 計(jì)算樣本的批次大小 默認(rèn)32 |
| imgsz | inference size (pixels) | 輸入網(wǎng)絡(luò)的圖片分辨率 默認(rèn)640 |
| conf-thres | confidence threshold | object置信度閾值 默認(rèn)0.001 |
| iou-thres | NMS IoU threshold | 進(jìn)行NMS時(shí)IOU的閾值 默認(rèn)0.6 |
| task | train, val, test, speed or study | 設(shè)置測試的類型 有train, val, test, speed or study幾種 默認(rèn)val |
| device | cuda device, i.e. 0 or 0,1,2,3 or cpu | 測試的設(shè)備 |
| workers | max dataloader workers (per RANK in DDP mode) | 加載數(shù)據(jù)使用的 dataloader workers |
| single-cls | treat as single-class dataset | 數(shù)據(jù)集是否只用一個(gè)類別 默認(rèn)False |
| augment | augmented inference | 測試是否使用TTA Test Time Augment 默認(rèn)False |
| verbose | report mAP by class | 是否打印出每個(gè)類別的mAP 默認(rèn)False |
| save-hybrid | save label+prediction hybrid results to *.txt | 保存label+prediction 雜交結(jié)果到對應(yīng).txt 默認(rèn)False |
| save-conf | save confidences in --save-txt labels | |
| save-json | save a COCO-JSON results file | 是否按照coco的json格式保存結(jié)果 默認(rèn)False |
| project | save to project/name | 測試保存的源文件 默認(rèn)runs/val |
| name | save to project/name | 測試保存的文件地址名 默認(rèn)exp 保存在runs/val/exp下 |
| exist-ok | existing project/name ok, do not increment | 是否保存在當(dāng)前文件，不新增 默認(rèn)False |
| half | use FP16 half-precision inference | 是否使用半精度推理 默認(rèn)False |
| dnn | use OpenCV DNN for ONNX inference | 是否使用 OpenCV DNN 對 ONNX 模型推理 |
根據(jù)解析的opt參數(shù)，調(diào)用run函數(shù)
def main(opt):
# 檢測requirements文件中需要的包是否安裝好了
check_requirements(requirements=ROOT / "requirements.txt", exclude=("tensorboard", "thop"))
if opt.task in ("train", "val", "test"): # run normally
if opt.conf_thres > 0.001: # 更多請見 https://github.com/ultralytics/yolov5/issues/1466
LOGGER.info(f"WARNING: confidence threshold {opt.conf_thres} > 0.001 produces invalid results")
run(**vars(opt))
else:
weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
opt.half = True # FP16 for fastest results
if opt.task == "speed": # speed benchmarks
# python val.py --task speed --data coco.yaml
# --batch 1 --weights yolov5n/ yolov5s/ ...
opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
for opt.weights in weights:
run(**vars(opt), plots=False)
elif opt.task == "study": # speed vs mAP benchmarks
# python val.py --task study --data coco.yaml
# --iou 0.7 --weights yolov5n/ yolov5s/...
for opt.weights in weights:
f = f"study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt"
x, y = (
list(range(256, 1536 + 128, 128)),
[],
) # x axis (image sizes), y axis
# "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：
for opt.imgsz in x: # img-size
LOGGER.info(f"\nRunning {f} --imgsz {opt.imgsz}...")
r, _, t = run(**vars(opt), plots=False)
y.append(r + t) # results and times
np.savetxt(f, y, fmt="%10.4g") # save
os.system("zip -r study.zip study_*.txt")
# 可視化各個(gè)指標(biāo)
plot_val_study(x=x) # plot
3. run函數(shù)https://github.com/Oneflow-Inc/one-yolov5/blob/bf8c66e011fcf5b8885068074ffc6b56c113a20c/val.py#L112-L3833.1 載入?yún)?shù)
# 不參與反向傳播
@flow.no_grad()
def run(
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)置測試的類型 有train, val, test, speed or study幾種 默認(rèn)val
device="", # 執(zhí)行 val.py 所在的設(shè)備 cuda device, i.e. 0 or 0,1,2,3 or cpu
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, # 測試時(shí)增強(qiáng)，詳細(xì)請看我們的教程：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 < 50 and final_epoch
save_txt=False, # 是否以txt文件的形式保存模型預(yù)測框的坐標(biāo) 默認(rèn)True
save_hybrid=False, # 是否save label+prediction hybrid results to *.txt 默認(rèn)False
save_conf=False, # 是否保存預(yù)測每個(gè)目標(biāo)的置信度到預(yù)測txt文件中 默認(rèn)True
save_json=False, # 是否按照coco的json格式保存預(yù)測框，并且使用cocoapi做評估（需要同樣coco的json格式的標(biāo)簽）,
#運(yùn)行test.py傳入默認(rèn)Fasle 運(yùn)行train.py則傳入is_coco and final_epoch(一般也是False)
project=ROOT / "runs/val", # 驗(yàn)證結(jié)果保存的根目錄 默認(rèn)是 runs/val
name="exp", # 驗(yàn)證結(jié)果保存的目錄 默認(rèn)是exp 最終: runs/val/exp
exist_ok=False, # 如果文件存在就increment name，不存在就新建 默認(rèn)False(默認(rèn)文件都是不存在的)
half=True, # 使用 FP16 的半精度推理
dnn=False, # 在 ONNX 推理時(shí)使用 OpenCV DNN 后段端
model=None, # 如果執(zhí)行val.py就為None 如果執(zhí)行train.py就會(huì)傳入( model=attempt_load(f, device).half() )
dataloader=None, # 數(shù)據(jù)加載器 如果執(zhí)行val.py就為None 如果執(zhí)行train.py就會(huì)傳入testloader
save_dir=Path(""), # 文件保存路徑 如果執(zhí)行val.py就為‘’ , 如果執(zhí)行train.py就會(huì)傳入save_dir(runs/train/expn)
plots=True, # 是否可視化 運(yùn)行val.py傳入，默認(rèn)True
callbacks=Callbacks(),
compute_loss=None, # 損失函數(shù) 運(yùn)行val.py傳入默認(rèn)None 運(yùn)行train.py則傳入compute_loss(train)
):
3.2 Initialize/load model and set device（初始化/加載模型以及設(shè)置設(shè)備） if training: # 通過 train.py 調(diào)用的run函數(shù)
device, of, engine = (
next(model.parameters()).device,
True,
False,
) # get model device, OneFlow model
half &= device.type != "cpu" # half precision only supported on CUDA
model.half() if half else model.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) # increment run
(save_dir / "labels" if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir
# 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
# 檢測輸入圖片的分辨率 imgsz 是否能被 stride 整除
imgsz = check_img_size(imgsz, s=stride) # check image size
half = model.fp16 # FP16 supported on limited backends with CUDA
if engine:
batch_size = model.batch_size
else:
device = model.device
if not of:
batch_size = 1 # export.py models default to batch-size 1
LOGGER.info(f"Forcing --batch-size 1 inference (1,3,{imgsz},{imgsz}) for non-OneFlow models")
# 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) and data["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 = 1 if single_cls else int(data["nc"]) # number of classes
# 設(shè)置iou閾值 從0.5-0.95取10個(gè)(0.05間隔) iou vector for mAP@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) # iou vector for mAP@0.5:0.95
niou = iouv.numel() # 示例 mAP@0.5:0.95 iou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
if not training: # 加載val數(shù)據(jù)集
if of and not single_cls: # check --weights are trained on --data
ncm = model.model.nc
assert ncm == nc, (
f"{weights} ({ncm} classes) trained on different --data than what you passed ({nc} " f"classes). Pass correct combination of" f" --weights and --data that are trained together."
)
model.warmup(imgsz=(1 if of else batch_size, 3, imgsz, imgsz)) # warmup
pad = 0.0 if task in ("speed", "benchmark") else 0.5
rect = False if task == "benchmark" else of # square inference for benchmarks
task = task if task in ("train", "val", "test") else "val" # path to train/val/test images
# 創(chuàng)建dataloader 這里的rect默認(rèn)為True 矩形推理用于測試集 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.names if hasattr(model, "names") else model.module.names))
# coco80_to_coco91_class : converts 80-index (val2014) to 91-index (paper)
# https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/
class_map = coco80_to_coco91_class() if is_coco else list(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
YOLOv5*博客內(nèi)容為網(wǎng)友個(gè)人發(fā)布，僅代表博主個(gè)人觀點(diǎn)，如有侵權(quán)請聯(lián)系工作人員刪除。
