操作系统

Ubuntu（Linux）

程序语言

Python（3.11.4）

第三方依赖

torch, torchvision, matplotlib等

xml标注文件转txt

import os

import xml.etree.ElementTree as ET

def convert_xml_to_txt(xml_folder, txt_folder, class_list):

    if not os.path.exists(txt_folder):

        os.makedirs(txt_folder)

    for xml_file in os.listdir(xml_folder):

        if xml_file.endswith('.xml'):

            tree = ET.parse(os.path.join(xml_folder, xml_file))

            root = tree.getroot()

            txt_filename = os.path.splitext(xml_file)[0] + '.txt'

            with open(os.path.join(txt_folder, txt_filename), 'w') as txt_file:

                for obj in root.findall('object'):

                    cls = obj.find('name').text

                    if cls not in class_list:

                        continue

                    cls_id = class_list.index(cls)

                    xmlbox = obj.find('bndbox')

                    xmin = int(xmlbox.find('xmin').text)

                    ymin = int(xmlbox.find('ymin').text)

                    xmax = int(xmlbox.find('xmax').text)

                    ymax = int(xmlbox.find('ymax').text)

                    img_width = int(root.find('size').find('width').text)

                    img_height = int(root.find('size').find('height').text)

                    x_center = ((xmin + xmax) / 2) / img_width

                    y_center = ((ymin + ymax) / 2) / img_height

                    width = (xmax - xmin) / img_width

                    height = (ymax - ymin) / img_height

                    txt_file.write(f"{cls_id} {x_center:.6f} {y_center:.6f} {width:.6f} {height:.6f}\n")

# 定义您的类别列表

classes = ['Leconte', 'Boerner', 'linnaeus', 'armandi', 'coleoptera']

# 转换 XML 文件

xmlpath = r"/home/ubuntu/yolov3/dataset/labels-xml/val"

txtpath = r"/home/ubuntu/yolov3/dataset/labels/val"

convert_xml_to_txt(xmlpath, txtpath, classes)

Data.yaml（数据加载）

# YOLOv5 🚀 by Ultralytics, AGPL-3.0 license

# PASCAL VOC dataset http://host.robots.ox.ac.uk/pascal/VOC by University of Oxford

# Example usage: python train.py --data VOC.yaml

# parent

# ├── yolov5

# └── datasets

#     └── VOC  ← downloads here (2.8 GB)

# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]

path: ../datasets/VOC

train: # train images (relative to 'path')  16551 images

  - /home/ubuntu/yolov3/dataset/images/train

val: # val images (relative to 'path')  4952 images

  - /home/ubuntu/yolov3/dataset/images/val

test: # test images (optional)

# Classes

names:

  0: Leconte

  1: Boerner

  2: linnaeus

  3: armandi

  4: coleoptera

Yolov3-tiny.yaml（网络结构加载）

# YOLOv3 🚀 by Ultralytics, AGPL-3.0 license

# Parameters

nc: 5 # number of classes

depth_multiple: 1.0 # model depth multiple

width_multiple: 1.0 # layer channel multiple

anchors:

  - [10, 14, 23, 27, 37, 58] # P4/16

  - [81, 82, 135, 169, 344, 319] # P5/32

# YOLOv3-tiny backbone

backbone:

  # [from, number, module, args]

  [

    [-1, 1, Conv, [16, 3, 1]], # 0

    [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 1-P1/2

    [-1, 1, Conv, [32, 3, 1]],

    [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 3-P2/4

    [-1, 1, Conv, [64, 3, 1]],

    [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 5-P3/8

    [-1, 1, Conv, [128, 3, 1]],

    [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 7-P4/16

    [-1, 1, Conv, [256, 3, 1]],

    [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 9-P5/32

    [-1, 1, Conv, [512, 3, 1]],

    [-1, 1, nn.ZeroPad2d, [[0, 1, 0, 1]]], # 11

    [-1, 1, nn.MaxPool2d, [2, 1, 0]], # 12

  ]

# YOLOv3-tiny head

head: [

    [-1, 1, Conv, [1024, 3, 1]],

    [-1, 1, Conv, [256, 1, 1]],

    [-1, 1, Conv, [512, 3, 1]], # 15 (P5/32-large)

    [-2, 1, Conv, [128, 1, 1]],

    [-1, 1, nn.Upsample, [None, 2, "nearest"]],

    [[-1, 8], 1, Concat, [1]], # cat backbone P4

    [-1, 1, Conv, [256, 3, 1]], # 19 (P4/16-medium)

    [[19, 15], 1, Detect, [nc, anchors]], # Detect(P4, P5)

  ]

Train.py（模型训练）

# YOLOv3 🚀 by Ultralytics, AGPL-3.0 license

"""

Train a YOLOv3 model on a custom dataset. Models and datasets download automatically from the latest YOLOv3 release.

Usage - Single-GPU training:

    $ python train.py --data coco128.yaml --weights yolov5s.pt --img 640  # from pretrained (recommended)

    $ python train.py --data coco128.yaml --weights '' --cfg yolov5s.yaml --img 640  # from scratch

Usage - Multi-GPU DDP training:

    $ python -m torch.distributed.run --nproc_per_node 4 --master_port 1 train.py --data coco128.yaml --weights yolov5s.pt --img 640 --device 0,1,2,3

Models:     https://github.com/ultralytics/yolov5/tree/master/models

Datasets:   https://github.com/ultralytics/yolov5/tree/master/data

Tutorial:   https://docs.ultralytics.com/yolov5/tutorials/train_custom_data

"""

import argparse

import math

import os

import random

import subprocess

import sys

import time

from copy import deepcopy

from datetime import datetime

from pathlib import Path

try:

    import comet_ml  # must be imported before torch (if installed)

except ImportError:

    comet_ml = None

import numpy as np

import torch

import torch.distributed as dist

import torch.nn as nn

import yaml

from torch.optim import lr_scheduler

from tqdm import tqdm

FILE = Path(__file__).resolve()

ROOT = FILE.parents[0]  # YOLOv3 root directory

if str(ROOT) not in sys.path:

    sys.path.append(str(ROOT))  # add ROOT to PATH

ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative

import val as validate  # for end-of-epoch mAP

from models.experimental import attempt_load

from models.yolo import Model

from utils.autoanchor import check_anchors

from utils.autobatch import check_train_batch_size

from utils.callbacks import Callbacks

from utils.dataloaders import create_dataloader

from utils.downloads import attempt_download, is_url

from utils.general import (

    LOGGER,

    TQDM_BAR_FORMAT,

    check_amp,

    check_dataset,

    check_file,

    check_git_info,

    check_git_status,

    check_img_size,

    check_requirements,

    check_suffix,

    check_yaml,

    colorstr,

    get_latest_run,

    increment_path,

    init_seeds,

    intersect_dicts,

    labels_to_class_weights,

    labels_to_image_weights,

    methods,

    one_cycle,

    print_args,

    print_mutation,

    strip_optimizer,

    yaml_save,

)

from utils.loggers import Loggers

from utils.loggers.comet.comet_utils import check_comet_resume

from utils.loss import ComputeLoss

from utils.metrics import fitness

from utils.plots import plot_evolve

from utils.torch_utils import (

    EarlyStopping,

    ModelEMA,

    de_parallel,

    select_device,

    smart_DDP,

    smart_optimizer,

    smart_resume,

    torch_distributed_zero_first,

)

LOCAL_RANK = int(os.getenv("LOCAL_RANK", -1))  # https://pytorch.org/docs/stable/elastic/run.html

RANK = int(os.getenv("RANK", -1))

WORLD_SIZE = int(os.getenv("WORLD_SIZE", 1))

GIT_INFO = check_git_info()

def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary

    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = (

        Path(opt.save_dir),

        opt.epochs,

        opt.batch_size,

        opt.weights,

        opt.single_cls,

        opt.evolve,

        opt.data,

        opt.cfg,

        opt.resume,

        opt.noval,

        opt.nosave,

        opt.workers,

        opt.freeze,

    )

    callbacks.run("on_pretrain_routine_start")

    # Directories

    w = save_dir / "weights"  # weights dir

    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir

    last, best = w / "last.pt", w / "best.pt"

    # Hyperparameters

    if isinstance(hyp, str):

        with open(hyp, errors="ignore") as f:

            hyp = yaml.safe_load(f)  # load hyps dict

    LOGGER.info(colorstr("hyperparameters: ") + ", ".join(f"{k}={v}" for k, v in hyp.items()))

    opt.hyp = hyp.copy()  # for saving hyps to checkpoints

    # Save run settings

    if not evolve:

        yaml_save(save_dir / "hyp.yaml", hyp)

        yaml_save(save_dir / "opt.yaml", vars(opt))

    # Loggers

    data_dict = None

    if RANK in {-1, 0}:

        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance

        # Register actions

        for k in methods(loggers):

            callbacks.register_action(k, callback=getattr(loggers, k))

        # Process custom dataset artifact link

        data_dict = loggers.remote_dataset

        if resume:  # If resuming runs from remote artifact

            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size

    # Config

    plots = not evolve and not opt.noplots  # create plots

    cuda = device.type != "cpu"

    init_seeds(opt.seed + 1 + RANK, deterministic=True)

    with torch_distributed_zero_first(LOCAL_RANK):

        data_dict = data_dict or check_dataset(data)  # check if None

    train_path, val_path = data_dict["train"], data_dict["val"]

    nc = 1 if single_cls else int(data_dict["nc"])  # number of classes

    names = {0: "item"} if single_cls and len(data_dict["names"]) != 1 else data_dict["names"]  # class names

    is_coco = isinstance(val_path, str) and val_path.endswith("coco/val2017.txt")  # COCO dataset

    # Model

    check_suffix(weights, ".pt")  # check weights

    pretrained = weights.endswith(".pt")

    if pretrained:

        with torch_distributed_zero_first(LOCAL_RANK):

            weights = attempt_download(weights)  # download if not found locally

        ckpt = torch.load(weights, map_location="cpu")  # load checkpoint to CPU to avoid CUDA memory leak

        model = Model(cfg or ckpt["model"].yaml, ch=3, nc=nc, anchors=hyp.get("anchors")).to(device)  # create

        exclude = ["anchor"] if (cfg or hyp.get("anchors")) and not resume else []  # exclude keys

        csd = ckpt["model"].float().state_dict()  # checkpoint state_dict as FP32

        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect

        model.load_state_dict(csd, strict=False)  # load

        LOGGER.info(f"Transferred {len(csd)}/{len(model.state_dict())} items from {weights}")  # report

    else:

        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get("anchors")).to(device)  # create

    amp = check_amp(model)  # check AMP

    # Freeze

    freeze = [f"model.{x}." for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze

    for k, v in model.named_parameters():

        v.requires_grad = True  # train all layers

        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)

        if any(x in k for x in freeze):

            LOGGER.info(f"freezing {k}")

            v.requires_grad = False

    # Image size

    gs = max(int(model.stride.max()), 32)  # grid size (max stride)

    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple

    # Batch size

    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size

        batch_size = check_train_batch_size(model, imgsz, amp)

        loggers.on_params_update({"batch_size": batch_size})

    # Optimizer

    nbs = 64  # nominal batch size

    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing

    hyp["weight_decay"] *= batch_size * accumulate / nbs  # scale weight_decay

    optimizer = smart_optimizer(model, opt.optimizer, hyp["lr0"], hyp["momentum"], hyp["weight_decay"])

    # Scheduler

    if opt.cos_lr:

        lf = one_cycle(1, hyp["lrf"], epochs)  # cosine 1->hyp['lrf']

    else:

        lf = lambda x: (1 - x / epochs) * (1.0 - hyp["lrf"]) + hyp["lrf"]  # linear

    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)

    # EMA

    ema = ModelEMA(model) if RANK in {-1, 0} else None

    # Resume

    best_fitness, start_epoch = 0.0, 0

    if pretrained:

        if resume:

            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)

        del ckpt, csd

    # DP mode

    if cuda and RANK == -1 and torch.cuda.device_count() > 1:

        LOGGER.warning(

            "WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n"

            "See Multi-GPU Tutorial at https://docs.ultralytics.com/yolov5/tutorials/multi_gpu_training to get started."

        )

        model = torch.nn.DataParallel(model)

    # SyncBatchNorm

    if opt.sync_bn and cuda and RANK != -1:

        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)

        LOGGER.info("Using SyncBatchNorm()")

    # Trainloader

    train_loader, dataset = create_dataloader(

        train_path,

        imgsz,

        batch_size // WORLD_SIZE,

        gs,

        single_cls,

        hyp=hyp,

        augment=True,

        cache=None if opt.cache == "val" else opt.cache,

        rect=opt.rect,

        rank=LOCAL_RANK,

        workers=workers,

        image_weights=opt.image_weights,

        quad=opt.quad,

        prefix=colorstr("train: "),

        shuffle=True,

        seed=opt.seed,

    )

    labels = np.concatenate(dataset.labels, 0)

    mlc = int(labels[:, 0].max())  # max label class

    assert mlc < nc, f"Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}"

    # Process 0

    if RANK in {-1, 0}:

        val_loader = create_dataloader(

            val_path,

            imgsz,

            batch_size // WORLD_SIZE * 2,

            gs,

            single_cls,

            hyp=hyp,

            cache=None if noval else opt.cache,

            rect=True,

            rank=-1,

            workers=workers * 2,

            pad=0.5,

            prefix=colorstr("val: "),

        )[0]

        if not resume:

            if not opt.noautoanchor:

                check_anchors(dataset, model=model, thr=hyp["anchor_t"], imgsz=imgsz)  # run AutoAnchor

            model.half().float()  # pre-reduce anchor precision

        callbacks.run("on_pretrain_routine_end", labels, names)

    # DDP mode

    if cuda and RANK != -1:

        model = smart_DDP(model)

    # Model attributes

    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)

    hyp["box"] *= 3 / nl  # scale to layers

    hyp["cls"] *= nc / 80 * 3 / nl  # scale to classes and layers

    hyp["obj"] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers

    hyp["label_smoothing"] = opt.label_smoothing

    model.nc = nc  # attach number of classes to model

    model.hyp = hyp  # attach hyperparameters to model

    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights

    model.names = names

    # Start training

    t0 = time.time()

    nb = len(train_loader)  # number of batches

    nw = max(round(hyp["warmup_epochs"] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)

    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training

    last_opt_step = -1

    maps = np.zeros(nc)  # mAP per class

    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)

    scheduler.last_epoch = start_epoch - 1  # do not move

    scaler = torch.cuda.amp.GradScaler(enabled=amp)

    stopper, stop = EarlyStopping(patience=opt.patience), False

    compute_loss = ComputeLoss(model)  # init loss class

    callbacks.run("on_train_start")

    LOGGER.info(

        f'Image sizes {imgsz} train, {imgsz} val\n'

        f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'

        f"Logging results to {colorstr('bold', save_dir)}\n"

        f'Starting training for {epochs} epochs...'

    )

    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------

        callbacks.run("on_train_epoch_start")

        model.train()

        # Update image weights (optional, single-GPU only)

        if opt.image_weights:

            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights

            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights

            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx

        # Update mosaic border (optional)

        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)

        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders

        mloss = torch.zeros(3, device=device)  # mean losses

        if RANK != -1:

            train_loader.sampler.set_epoch(epoch)

        pbar = enumerate(train_loader)

        LOGGER.info(("\n" + "%11s" * 7) % ("Epoch", "GPU_mem", "box_loss", "obj_loss", "cls_loss", "Instances", "Size"))

        if RANK in {-1, 0}:

            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar

        optimizer.zero_grad()

        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------

            callbacks.run("on_train_batch_start")

            ni = i + nb * epoch  # number integrated batches (since train start)

            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0

            # Warmup

            if ni <= nw:

                xi = [0, nw]  # x interp

                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)

                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())

                for j, x in enumerate(optimizer.param_groups):

                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0

                    x["lr"] = np.interp(ni, xi, [hyp["warmup_bias_lr"] if j == 0 else 0.0, x["initial_lr"] * lf(epoch)])

                    if "momentum" in x:

                        x["momentum"] = np.interp(ni, xi, [hyp["warmup_momentum"], hyp["momentum"]])

            # Multi-scale

            if opt.multi_scale:

                sz = random.randrange(int(imgsz * 0.5), int(imgsz * 1.5) + gs) // gs * gs  # size

                sf = sz / max(imgs.shape[2:])  # scale factor

                if sf != 1:

                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)

                    imgs = nn.functional.interpolate(imgs, size=ns, mode="bilinear", align_corners=False)

            # Forward

            with torch.cuda.amp.autocast(amp):

                pred = model(imgs)  # forward

                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size

                if RANK != -1:

                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode

                if opt.quad:

                    loss *= 4.0

            # Backward

            scaler.scale(loss).backward()

            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html

            if ni - last_opt_step >= accumulate:

                scaler.unscale_(optimizer)  # unscale gradients

                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients

                scaler.step(optimizer)  # optimizer.step

                scaler.update()

                optimizer.zero_grad()

                if ema:

                    ema.update(model)

                last_opt_step = ni

            # Log

            if RANK in {-1, 0}:

                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses

                mem = f"{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G"  # (GB)

                pbar.set_description(

                    ("%11s" * 2 + "%11.4g" * 5)

                    % (f"{epoch}/{epochs - 1}", mem, *mloss, targets.shape[0], imgs.shape[-1])

                )

                callbacks.run("on_train_batch_end", model, ni, imgs, targets, paths, list(mloss))

                if callbacks.stop_training:

                    return

            # end batch ------------------------------------------------------------------------------------------------

        # Scheduler

        lr = [x["lr"] for x in optimizer.param_groups]  # for loggers

        scheduler.step()

        if RANK in {-1, 0}:

            # mAP

            callbacks.run("on_train_epoch_end", epoch=epoch)

            ema.update_attr(model, include=["yaml", "nc", "hyp", "names", "stride", "class_weights"])

            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop

            if not noval or final_epoch:  # Calculate mAP

                results, maps, _ = validate.run(

                    data_dict,

                    batch_size=batch_size // WORLD_SIZE * 2,

                    imgsz=imgsz,

                    half=amp,

                    model=ema.ema,

                    single_cls=single_cls,

                    dataloader=val_loader,

                    save_dir=save_dir,

                    plots=False,

                    callbacks=callbacks,

                    compute_loss=compute_loss,

                )

            # Update best mAP

            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]

            stop = stopper(epoch=epoch, fitness=fi)  # early stop check

            if fi > best_fitness:

                best_fitness = fi

            log_vals = list(mloss) + list(results) + lr

            callbacks.run("on_fit_epoch_end", log_vals, epoch, best_fitness, fi)

            # Save model

            if (not nosave) or (final_epoch and not evolve):  # if save

                ckpt = {

                    "epoch": epoch,

                    "best_fitness": best_fitness,

                    "model": deepcopy(de_parallel(model)).half(),

                    "ema": deepcopy(ema.ema).half(),

                    "updates": ema.updates,

                    "optimizer": optimizer.state_dict(),

                    "opt": vars(opt),

                    "git": GIT_INFO,  # {remote, branch, commit} if a git repo

                    "date": datetime.now().isoformat(),

                }

                # Save last, best and delete

                torch.save(ckpt, last)

                if best_fitness == fi:

                    torch.save(ckpt, best)

                if opt.save_period > 0 and epoch % opt.save_period == 0:

                    torch.save(ckpt, w / f"epoch{epoch}.pt")

                del ckpt

                callbacks.run("on_model_save", last, epoch, final_epoch, best_fitness, fi)

        # EarlyStopping

        if RANK != -1:  # if DDP training

            broadcast_list = [stop if RANK == 0 else None]

            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks

            if RANK != 0:

                stop = broadcast_list[0]

        if stop:

            break  # must break all DDP ranks

        # end epoch ----------------------------------------------------------------------------------------------------

    # end training -----------------------------------------------------------------------------------------------------

    if RANK in {-1, 0}:

        LOGGER.info(f"\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.")

        for f in last, best:

            if f.exists():

                strip_optimizer(f)  # strip optimizers

                if f is best:

                    LOGGER.info(f"\nValidating {f}...")

                    results, _, _ = validate.run(

                        data_dict,

                        batch_size=batch_size // WORLD_SIZE * 2,

                        imgsz=imgsz,

                        model=attempt_load(f, device).half(),

                        iou_thres=0.65 if is_coco else 0.60,  # best pycocotools at iou 0.65

                        single_cls=single_cls,

                        dataloader=val_loader,

                        save_dir=save_dir,

                        save_json=is_coco,

                        verbose=True,

                        plots=plots,

                        callbacks=callbacks,

                        compute_loss=compute_loss,

                    )  # val best model with plots

                    if is_coco:

                        callbacks.run("on_fit_epoch_end", list(mloss) + list(results) + lr, epoch, best_fitness, fi)

        callbacks.run("on_train_end", last, best, epoch, results)

    torch.cuda.empty_cache()

    return results

def parse_opt(known=False):

    parser = argparse.ArgumentParser()

    parser.add_argument("--weights", type=str, default=ROOT / "weights/yolov3.pt", help="initial weights path")

    parser.add_argument("--cfg", type=str, default=ROOT / "models/yolov3-tiny.yaml", help="model.yaml path")

    parser.add_argument("--data", type=str, default=ROOT / "voc.yaml", help="dataset.yaml path")

    parser.add_argument("--hyp", type=str, default=ROOT / "data/hyps/hyp.scratch-low.yaml", help="hyperparameters path")

    parser.add_argument("--epochs", type=int, default=30, help="total training epochs")

    parser.add_argument("--batch-size", type=int, default=16, help="total batch size for all GPUs, -1 for autobatch")

    parser.add_argument("--imgsz", "--img", "--img-size", type=int, default=640, help="train, val image size (pixels)")

    parser.add_argument("--rect", action="store_true", help="rectangular training")

    parser.add_argument("--resume", nargs="?", const=True, default=False, help="resume most recent training")

    parser.add_argument("--nosave", action="store_true", help="only save final checkpoint")

    parser.add_argument("--noval", action="store_true", help="only validate final epoch")

    parser.add_argument("--noautoanchor", action="store_true", help="disable AutoAnchor")

    parser.add_argument("--noplots", action="store_true", help="save no plot files")

    parser.add_argument("--evolve", type=int, nargs="?", const=300, help="evolve hyperparameters for x generations")

    parser.add_argument("--bucket", type=str, default="", help="gsutil bucket")

    parser.add_argument("--cache", type=str, nargs="?", const="ram", help="image --cache ram/disk")

    parser.add_argument("--image-weights", action="store_true", help="use weighted image selection for training")

    parser.add_argument("--device", default="0", help="cuda device, i.e. 0 or 0,1,2,3 or cpu")

    parser.add_argument("--multi-scale", action="store_true", help="vary img-size +/- 50%%")

    parser.add_argument("--single-cls", action="store_true", help="train multi-class data as single-class")

    parser.add_argument("--optimizer", type=str, choices=["SGD", "Adam", "AdamW"], default="SGD", help="optimizer")

    parser.add_argument("--sync-bn", action="store_true", help="use SyncBatchNorm, only available in DDP mode")

    parser.add_argument("--workers", type=int, default=8, help="max dataloader workers (per RANK in DDP mode)")

    parser.add_argument("--project", default=ROOT / "runs/train", help="save to project/name")

    parser.add_argument("--name", default="exp", help="save to project/name")

    parser.add_argument("--exist-ok", action="store_true", help="existing project/name ok, do not increment")

    parser.add_argument("--quad", action="store_true", help="quad dataloader")

    parser.add_argument("--cos-lr", action="store_true", help="cosine LR scheduler")

    parser.add_argument("--label-smoothing", type=float, default=0.0, help="Label smoothing epsilon")

    parser.add_argument("--patience", type=int, default=100, help="EarlyStopping patience (epochs without improvement)")

    parser.add_argument("--freeze", nargs="+", type=int, default=[0], help="Freeze layers: backbone=10, first3=0 1 2")

    parser.add_argument("--save-period", type=int, default=-1, help="Save checkpoint every x epochs (disabled if < 1)")

    parser.add_argument("--seed", type=int, default=0, help="Global training seed")

    parser.add_argument("--local_rank", type=int, default=-1, help="Automatic DDP Multi-GPU argument, do not modify")

    # Logger arguments

    parser.add_argument("--entity", default=None, help="Entity")

    parser.add_argument("--upload_dataset", nargs="?", const=True, default=False, help='Upload data, "val" option')

    parser.add_argument("--bbox_interval", type=int, default=-1, help="Set bounding-box image logging interval")

    parser.add_argument("--artifact_alias", type=str, default="latest", help="Version of dataset artifact to use")

    return parser.parse_known_args()[0] if known else parser.parse_args()

def main(opt, callbacks=Callbacks()):

    # Checks

    if RANK in {-1, 0}:

        print_args(vars(opt))

        check_git_status()

        check_requirements(ROOT / "requirements.txt")

    # Resume (from specified or most recent last.pt)

    if opt.resume and not check_comet_resume(opt) and not opt.evolve:

        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())

        opt_yaml = last.parent.parent / "opt.yaml"  # train options yaml

        opt_data = opt.data  # original dataset

        if opt_yaml.is_file():

            with open(opt_yaml, errors="ignore") as f:

                d = yaml.safe_load(f)

        else:

            d = torch.load(last, map_location="cpu")["opt"]

        opt = argparse.Namespace(**d)  # replace

        opt.cfg, opt.weights, opt.resume = "", str(last), True  # reinstate

        if is_url(opt_data):

            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout

    else:

        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = (

            check_file(opt.data),

            check_yaml(opt.cfg),

            check_yaml(opt.hyp),

            str(opt.weights),

            str(opt.project),

        )  # checks

        assert len(opt.cfg) or len(opt.weights), "either --cfg or --weights must be specified"

        if opt.evolve:

            if opt.project == str(ROOT / "runs/train"):  # if default project name, rename to runs/evolve

                opt.project = str(ROOT / "runs/evolve")

            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume

        if opt.name == "cfg":

            opt.name = Path(opt.cfg).stem  # use model.yaml as name

        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))

    # DDP mode

    device = select_device(opt.device, batch_size=opt.batch_size)

    if LOCAL_RANK != -1:

        msg = "is not compatible with YOLOv3 Multi-GPU DDP training"

        assert not opt.image_weights, f"--image-weights {msg}"

        assert not opt.evolve, f"--evolve {msg}"

        assert opt.batch_size != -1, f"AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size"

        assert opt.batch_size % WORLD_SIZE == 0, f"--batch-size {opt.batch_size} must be multiple of WORLD_SIZE"

        assert torch.cuda.device_count() > LOCAL_RANK, "insufficient CUDA devices for DDP command"

        torch.cuda.set_device(LOCAL_RANK)

        device = torch.device("cuda", LOCAL_RANK)

        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")

    # Train

    if not opt.evolve:

        train(opt.hyp, opt, device, callbacks)

    # Evolve hyperparameters (optional)

    else:

        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)

        meta = {

            "lr0": (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)

            "lrf": (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)

            "momentum": (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1

            "weight_decay": (1, 0.0, 0.001),  # optimizer weight decay

            "warmup_epochs": (1, 0.0, 5.0),  # warmup epochs (fractions ok)

            "warmup_momentum": (1, 0.0, 0.95),  # warmup initial momentum

            "warmup_bias_lr": (1, 0.0, 0.2),  # warmup initial bias lr

            "box": (1, 0.02, 0.2),  # box loss gain

            "cls": (1, 0.2, 4.0),  # cls loss gain

            "cls_pw": (1, 0.5, 2.0),  # cls BCELoss positive_weight

            "obj": (1, 0.2, 4.0),  # obj loss gain (scale with pixels)

            "obj_pw": (1, 0.5, 2.0),  # obj BCELoss positive_weight

            "iou_t": (0, 0.1, 0.7),  # IoU training threshold

            "anchor_t": (1, 2.0, 8.0),  # anchor-multiple threshold

            "anchors": (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)

            "fl_gamma": (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)

            "hsv_h": (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)

            "hsv_s": (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)

            "hsv_v": (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)

            "degrees": (1, 0.0, 45.0),  # image rotation (+/- deg)

            "translate": (1, 0.0, 0.9),  # image translation (+/- fraction)

            "scale": (1, 0.0, 0.9),  # image scale (+/- gain)

            "shear": (1, 0.0, 10.0),  # image shear (+/- deg)

            "perspective": (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001

            "flipud": (1, 0.0, 1.0),  # image flip up-down (probability)

            "fliplr": (0, 0.0, 1.0),  # image flip left-right (probability)

            "mosaic": (1, 0.0, 1.0),  # image mixup (probability)

            "mixup": (1, 0.0, 1.0),  # image mixup (probability)

            "copy_paste": (1, 0.0, 1.0),

        }  # segment copy-paste (probability)

        with open(opt.hyp, errors="ignore") as f:

            hyp = yaml.safe_load(f)  # load hyps dict

            if "anchors" not in hyp:  # anchors commented in hyp.yaml

                hyp["anchors"] = 3

        if opt.noautoanchor:

            del hyp["anchors"], meta["anchors"]

        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch

        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices

        evolve_yaml, evolve_csv = save_dir / "hyp_evolve.yaml", save_dir / "evolve.csv"

        if opt.bucket:

            # download evolve.csv if exists

            subprocess.run(

                [

                    "gsutil",

                    "cp",

                    f"gs://{opt.bucket}/evolve.csv",

                    str(evolve_csv),

                ]

            )

        for _ in range(opt.evolve):  # generations to evolve

            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate

                # Select parent(s)

                parent = "single"  # parent selection method: 'single' or 'weighted'

                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=",", skiprows=1)

                n = min(5, len(x))  # number of previous results to consider

                x = x[np.argsort(-fitness(x))][:n]  # top n mutations

                w = fitness(x) - fitness(x).min() + 1e-6  # weights (sum > 0)

                if parent == "single" or len(x) == 1:

                    # x = x[random.randint(0, n - 1)]  # random selection

                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection

                elif parent == "weighted":

                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination

                # Mutate

                mp, s = 0.8, 0.2  # mutation probability, sigma

                npr = np.random

                npr.seed(int(time.time()))

                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1

                ng = len(meta)

                v = np.ones(ng)

                while all(v == 1):  # mutate until a change occurs (prevent duplicates)

                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)

                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)

                    hyp[k] = float(x[i + 7] * v[i])  # mutate

            # Constrain to limits

            for k, v in meta.items():

                hyp[k] = max(hyp[k], v[1])  # lower limit

                hyp[k] = min(hyp[k], v[2])  # upper limit

                hyp[k] = round(hyp[k], 5)  # significant digits

            # Train mutation

            results = train(hyp.copy(), opt, device, callbacks)

            callbacks = Callbacks()

            # Write mutation results

            keys = (

                "metrics/precision",

                "metrics/recall",

                "metrics/mAP_0.5",

                "metrics/mAP_0.5:0.95",

                "val/box_loss",

                "val/obj_loss",

                "val/cls_loss",

            )

            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)

        # Plot results

        plot_evolve(evolve_csv)

        LOGGER.info(

            f'Hyperparameter evolution finished {opt.evolve} generations\n'

            f"Results saved to {colorstr('bold', save_dir)}\n"

            f'Usage example: $ python train.py --hyp {evolve_yaml}'

        )

def run(**kwargs):

    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolov5m.pt')

    opt = parse_opt(True)

    for k, v in kwargs.items():

        setattr(opt, k, v)

    main(opt)

    return opt

if __name__ == "__main__":

    opt = parse_opt()

    main(opt)

Detect.py（测试模型）

# YOLOv3 🚀 by Ultralytics, AGPL-3.0 license

"""

Run YOLOv3 detection inference on images, videos, directories, globs, YouTube, webcam, streams, etc.

Usage - sources:

    $ python detect.py --weights yolov5s.pt --source 0                               # webcam

                                                     img.jpg                         # image

                                                     vid.mp4                         # video

                                                     screen                          # screenshot

                                                     path/                           # directory

                                                     list.txt                        # list of images

                                                     list.streams                    # list of streams

                                                     'path/*.jpg'                    # glob

                                                     'https://youtu.be/LNwODJXcvt4'  # YouTube

                                                     'rtsp://example.com/media.mp4'  # RTSP, RTMP, HTTP stream

Usage - formats:

    $ python detect.py --weights yolov5s.pt                 # PyTorch

                                 yolov5s.torchscript        # TorchScript

                                 yolov5s.onnx               # ONNX Runtime or OpenCV DNN with --dnn

                                 yolov5s_openvino_model     # OpenVINO

                                 yolov5s.engine             # TensorRT

                                 yolov5s.mlmodel            # CoreML (macOS-only)

                                 yolov5s_saved_model        # TensorFlow SavedModel

                                 yolov5s.pb                 # TensorFlow GraphDef

                                 yolov5s.tflite             # TensorFlow Lite

                                 yolov5s_edgetpu.tflite     # TensorFlow Edge TPU

                                 yolov5s_paddle_model       # PaddlePaddle

"""

import argparse

import os

import platform

import sys

from pathlib import Path

import torch

FILE = Path(__file__).resolve()

ROOT = FILE.parents[0]  # YOLOv3 root directory

if str(ROOT) not in sys.path:

    sys.path.append(str(ROOT))  # add ROOT to PATH

ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative

from ultralytics.utils.plotting import Annotator, colors, save_one_box

from models.common import DetectMultiBackend

from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams

from utils.general import (

    LOGGER,

    Profile,

    check_file,

    check_img_size,

    check_imshow,

    check_requirements,

    colorstr,

    cv2,

    increment_path,

    non_max_suppression,

    print_args,

    scale_boxes,

    strip_optimizer,

    xyxy2xywh,

)

from utils.torch_utils import select_device, smart_inference_mode

@smart_inference_mode()

def run(

    weights=ROOT / "yolov5s.pt",  # model path or triton URL

    source=ROOT / "data/images",  # file/dir/URL/glob/screen/0(webcam)

    data=ROOT / "data/coco128.yaml",  # dataset.yaml path

    imgsz=(1262,1262),  # inference size (height, width)

    conf_thres=0.25,  # confidence threshold

    iou_thres=0.45,  # NMS IOU threshold

    max_det=1000,  # maximum detections per image

    device="",  # cuda device, i.e. 0 or 0,1,2,3 or cpu

    view_img=False,  # show results

    save_txt=False,  # save results to *.txt

    save_conf=False,  # save confidences in --save-txt labels

    save_crop=False,  # save cropped prediction boxes

    nosave=False,  # do not save images/videos

    classes=None,  # filter by class: --class 0, or --class 0 2 3

    agnostic_nms=False,  # class-agnostic NMS

    augment=False,  # augmented inference

    visualize=False,  # visualize features

    update=False,  # update all models

    project=ROOT / "runs/detect",  # save results to project/name

    name="exp",  # save results to project/name

    exist_ok=False,  # existing project/name ok, do not increment

    line_thickness=3,  # bounding box thickness (pixels)

    hide_labels=False,  # hide labels

    hide_conf=False,  # hide confidences

    half=False,  # use FP16 half-precision inference

    dnn=False,  # use OpenCV DNN for ONNX inference

    vid_stride=1,  # video frame-rate stride

):

    source = str(source)

    save_img = not nosave and not source.endswith(".txt")  # save inference images

    is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)

    is_url = source.lower().startswith(("rtsp://", "rtmp://", "http://", "https://"))

    webcam = source.isnumeric() or source.endswith(".streams") or (is_url and not is_file)

    screenshot = source.lower().startswith("screen")

    if is_url and is_file:

        source = check_file(source)  # download

    # Directories

    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

    # Load model

    device = select_device(device)

    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)

    stride, names, pt = model.stride, model.names, model.pt

    imgsz = check_img_size(imgsz, s=stride)  # check image size

    # Dataloader

    bs = 1  # batch_size

    if webcam:

        view_img = check_imshow(warn=True)

        dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)