FairMOT yolov5s转onnx

代码：https://github.com/ifzhang/FairMOT
论文讲解参考：https://blog.csdn.net/weixin_42398658/article/details/110873083
https://www.163.com/dy/article/H10AGKQM0552OFB6.html
https://zhuanlan.zhihu.com/p/393322545

损失函数

self.loss_stats, self.loss = self._get_losses(opt)
###IDLoss=crossentropy
###crit=focalloss
###crit_reg=l1loss
###crit_wh=l1loss

模型解码

with torch.no_grad():
    output = self.model(im_blob)[-1]
    hm = output['hm'].sigmoid_()
    wh = output['wh']
    id_feature = output['id']
    id_feature = F.normalize(id_feature, dim=1)
    reg = output['reg'] if self.opt.reg_offset else None
    dets, inds = mot_decode(hm, wh, reg=reg, ltrb=self.opt.ltrb, K=self.opt.K)
    id_feature = _tranpose_and_gather_feat(id_feature, inds)   #(1,500,64)
    id_feature = id_feature.squeeze(0)     #(500,64)
    id_feature = id_feature.cpu().numpy()     #(500,64)
dets = self.post_process(dets, meta)      #(500,5)
dets = self.merge_outputs([dets])[1]      #(500,5)

remain_inds = dets[:, 4] > self.opt.conf_thres
dets = dets[remain_inds]    #过滤完之后还剩的目标框(26,5)
id_feature = id_feature[remain_inds]    #过滤完之后还剩的id特征(26,64)
if len(dets) > 0:
    '''Detections'''
    detections = [STrack(STrack.tlbr_to_tlwh(tlbrs[:4]), tlbrs[4], f, 30) for
                  (tlbrs, f) in zip(dets[:, :5], id_feature)]
else:
    detections = []
''' Add newly detected tracklets to tracked_stracks'''
unconfirmed = []
tracked_stracks = []  # type: list[STrack]
for track in self.tracked_stracks:
     if not track.is_activated:
          unconfirmed.append(track)
     else:
          tracked_stracks.append(track)
''' Step 2: First association, with embedding'''
strack_pool = joint_stracks(tracked_stracks, self.lost_stracks)
# Predict the current location with KF
#for strack in strack_pool:
     #strack.predict()
STrack.multi_predict(strack_pool)
dists = matching.embedding_distance(strack_pool, detections)
#dists = matching.iou_distance(strack_pool, detections)
dists = matching.fuse_motion(self.kalman_filter, dists, strack_pool, detections)
matches, u_track, u_detection = matching.linear_assignment(dists, thresh=0.4)
...

‘hm’:(1,1,152,272)
‘wh’:(1,4,152,272)
‘id’:(1,64,152,272),长度为64的特征
‘reg’:(1,2,152,272)
dets:(1,500,6)
id_feature:(1,500,64)

推理时间对比

##### FairMOT
t1 = time_sync()
output = self.model(im_blob)[-1]
t2 = time_sync()
#####yolov5s
t1 = time_sync()
pred = model(img,augment=augment,visualize=increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False)[0]
pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det)
t2 = time_sync()

2个1080ti
u版yolov5s模型推理+nms， 640640，一张图片耗时6ms
yolov5s FairMOT模型推理，1088608，一张图片耗时10ms

FairMOT yolov5s转onnx

"""Export a YOLOv5 *.pt model to TorchScript, ONNX, CoreML formats Usage: $ python path/to/export.py --weights yolov5s.pt --img 640 --batch 1 """

import argparse
import sys
import time
from pathlib import Path
import torch
import torch.nn as nn
from torch.utils.mobile_optimizer import optimize_for_mobile
from lib.models.model import create_model, load_model
from lib.models.yolo import Detect
from lib.models.common import Conv

FILE = Path(__file__).absolute()
sys.path.append(FILE.parents[0].as_posix())  # add yolov5/ to path

def colorstr(*input):
    # Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e. colorstr('blue', 'hello world')
    *args, string = input if len(input) > 1 else ('blue', 'bold', input[0])  # color arguments, string
    colors = {
    'black': '\033[30m',  # basic colors
              'red': '\033[31m',
              'green': '\033[32m',
              'yellow': '\033[33m',
              'blue': '\033[34m',
              'magenta': '\033[35m',
              'cyan': '\033[36m',
              'white': '\033[37m',
              'bright_black': '\033[90m',  # bright colors
              'bright_red': '\033[91m',
              'bright_green': '\033[92m',
              'bright_yellow': '\033[93m',
              'bright_blue': '\033[94m',
              'bright_magenta': '\033[95m',
              'bright_cyan': '\033[96m',
              'bright_white': '\033[97m',
              'end': '\033[0m',  # misc
              'bold': '\033[1m',
              'underline': '\033[4m'}
    return ''.join(colors[x] for x in args) + f'{
      string}' + colors['end']


def export_onnx(model, img, file, opset_version, train, dynamic, simplify):
    # ONNX model export
    prefix = colorstr('ONNX:')
    import onnx

    print(f'\n{
      prefix} starting export with onnx {
      onnx.__version__}...')
    f = file.with_suffix('.onnx')
    torch.onnx.export(model, img, f, verbose=False, opset_version=opset_version,
                      training=torch.onnx.TrainingMode.TRAINING if train else torch.onnx.TrainingMode.EVAL,
                      do_constant_folding=not train,
                      input_names=['images'],
                      output_names=['output'],
                      dynamic_axes={
    'images': {
    0: 'batch', 2: 'height', 3: 'width'},  # shape(1,3,640,640)
                                    'output': {
    0: 'batch', 1: 'anchors'}  # shape(1,25200,85)
                                    } if dynamic else None)

    # Checks
    model_onnx = onnx.load(f)  # load onnx model
    onnx.checker.check_model(model_onnx)  # check onnx model
    # print(onnx.helper.printable_graph(model_onnx.graph)) # print

    # Simplify
    if simplify:
        try:
            import onnxsim

            print(f'{
      prefix} simplifying with onnx-simplifier {
      onnxsim.__version__}...')
            model_onnx, check = onnxsim.simplify(
                model_onnx,
                dynamic_input_shape=dynamic,
                input_shapes={
    'images': list(img.shape)} if dynamic else None)
            assert check, 'assert check failed'
            onnx.save(model_onnx, f)
        except Exception as e:
            print(f'{
      prefix} simplifier failure: {
      e}')


def select_device(device='', batch_size=None):
    device = str(device).strip().lower().replace('cuda:', '')  # to string, 'cuda:0' to '0'
    cpu = device == 'cpu'
    cuda = not cpu and torch.cuda.is_available()
    return torch.device('cuda:0' if cuda else 'cpu')
class SiLU(nn.Module):  # export-friendly version of nn.SiLU()
    @staticmethod
    def forward(x):
        return x * torch.sigmoid(x)


class Hardswish(nn.Module):  # export-friendly version of nn.Hardswish()
    @staticmethod
    def forward(x):
        # return x * F.hardsigmoid(x) # for torchscript and CoreML
        return x * F.hardtanh(x + 3, 0., 6.) / 6.  # for torchscript, CoreML and ONNX

def run(weights='',  # weights path
        img_size=(640, 640),  # image (height, width)
        batch_size=1,  # batch size
        device='cpu',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
        include=('torchscript', 'onnx', 'coreml'),  # include formats
        half=False,  # FP16 half-precision export
        inplace=False,  # set YOLOv5 Detect() inplace=True
        train=False,  # model.train() mode
        optimize=False,  # TorchScript: optimize for mobile
        dynamic=False,  # ONNX: dynamic axes
        simplify=False,  # ONNX: simplify model
        opset_version=12,  # ONNX: opset version
        arch='',
        head_conv=-1,
        model_path='',
        ltrb=True,
        reid_dim=64,
        ):

    if opt.head_conv == -1: # init default head_conv
      opt.head_conv = 256 if 'dla' in opt.arch else 256

    num_classes = 1
    heads = {
    'hm': num_classes,
                 'wh': 2 if not ltrb else 4,
                 'id': reid_dim}
    heads.update({
    'reg': 2})
    t = time.time()
    include = [x.lower() for x in include]
    img_size *= 2 if len(img_size) == 1 else 1  # expand
    file = Path(weights)

    # Load PyTorch model
    device = select_device(device)
    assert not (device.type == 'cpu' and half), '--half only compatible with GPU export, i.e. use --device 0'

    model = create_model(arch, heads, head_conv)
    model = load_model(model, model_path)

    # Input
    # gs = int(max(model.stride)) # grid size (max stride)
    # img_size = [check_img_size(x, gs) for x in img_size] # verify img_size are gs-multiples
    img = torch.zeros(batch_size, 3, *img_size).to(device)  # image size(1,3,320,192) iDetection

    # Update model
    if half:
        img, model = img.half(), model.half()  # to FP16
    model.train() if train else model.eval()  # training mode = no Detect() layer grid construction
    for k, m in model.backbone.named_modules():
        if isinstance(m, Conv):  # assign export-friendly activations
            if isinstance(m.act, nn.Hardswish):
                m.act = Hardswish()
            elif isinstance(m.act, nn.SiLU):
                m.act = SiLU()
        elif isinstance(m, Detect):
            m.inplace = inplace
            m.onnx_dynamic = dynamic
            # m.forward = m.forward_export # assign forward (optional)

    # Exports
    # if 'torchscript' in include:
    # export_torchscript(model, img, file, optimize)
    if 'onnx' in include:
        export_onnx(model.backbone, img, file, opset_version, train, dynamic, simplify)

    # Finish
    print(f'\nExport complete ({
      time.time() - t:.2f}s). Visualize with https://github.com/lutzroeder/netron.')


def parse_opt():
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='../exp/mot/all_yolov5s/model_60.pth', help='weights path')
    parser.add_argument('--img-size', nargs='+', type=int, default=[608, 1088], help='image (height, width)')
    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
    parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--include', nargs='+', default=['torchscript', 'onnx', 'coreml'], help='include formats')
    parser.add_argument('--half', action='store_true', help='FP16 half-precision export')
    parser.add_argument('--inplace', action='store_true', help='set YOLOv5 Detect() inplace=True')
    parser.add_argument('--train', action='store_true', help='model.train() mode')
    parser.add_argument('--optimize', action='store_true', help='TorchScript: optimize for mobile')
    parser.add_argument('--dynamic', action='store_true', help='ONNX: dynamic axes')
    parser.add_argument('--simplify', action='store_true', help='ONNX: simplify model')
    parser.add_argument('--opset-version', type=int, default=12, help='ONNX: opset version')
    parser.add_argument('--arch', default='yolo',
                             help='model architecture. Currently tested'
                                  'resdcn_34 | resdcn_50 | resfpndcn_34 |'
                                  'dla_34 | hrnet_18')
    parser.add_argument('--head_conv', type=int, default=-1,
                             help='conv layer channels for output head'
                                  '0 for no conv layer'
                                  '-1 for default setting: '
                                  '256 for resnets and 256 for dla.')
    parser.add_argument('--model_path', default='FairMOT-master/exp/mot/all_yolov5s/model_60.pth',
                             help='path to pretrained model')
    parser.add_argument('--ltrb', default=True,
                        help='regress left, top, right, bottom of bbox')
    parser.add_argument('--reid_dim', type=int, default=64,
                        help='feature dim for reid')
    opt = parser.parse_args()
    return opt


def main(opt):
    run(**vars(opt))


if __name__ == "__main__":
    opt = parse_opt()
    main(opt)

但是目前转出来的模型还有点问题，转出来的模型不包括4个检测头，只到特征部分，还差4个头的conv没有包含进去。