speed problem

[Hiwyl]

使用原始yolov5s6 640输入，速度为12ms，使用下面repvgg_block 640输入为13ms GPU v100

YOLOv5 🚀 by Ultralytics, GPL-3.0 license

Parameters

nc: 80 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple
anchors:

• [19,27, 44,40, 38,94] # P3/8
• [96,68, 86,152, 180,137] # P4/16
• [140,301, 303,264, 238,542] # P5/32
• [436,615, 739,380, 925,792] # P6/64

YOLOv5 v6.0 backbone

backbone:

[from, number, module, args]

[[-1, 1, Conv, [32, 6, 2, 2]], # 0-P1/2
[-1, 1, Conv, [64, 3, 2]], # 1-P2/4
[-1, 1, C3, [64]],
[-1, 1, RepVGGBlock, [128, 3, 2]], # 3-P3/8
[-1, 3, C3, [128]],
[-1, 1, RepVGGBlock, [256, 3, 2]], # 5-P4/16
[-1, 3, C3, [256]],
[-1, 1, RepVGGBlock, [512, 3, 2]], # 7-P5/32
[-1, 3, C3, [512]],
[-1, 1, RepVGGBlock, [768, 3, 2]], # 9-P6/64
[-1, 3, C3, [768]],
[-1, 1, SPPF, [768, 5]], # 11
]

YOLOv5 v6.0 head

head:
[[-1, 1, Conv, [512, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 8], 1, Concat, [1]], # cat backbone P5
[-1, 3, C3, [512, False]], # 15

[-1, 1, Conv, [256, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 6], 1, Concat, [1]], # cat backbone P4
[-1, 3, C3, [256, False]], # 19

[-1, 1, Conv, [128, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 4], 1, Concat, [1]], # cat backbone P3
[-1, 3, C3, [128, False]], # 23 (P3/8-small)

[-1, 1, Conv, [128, 3, 2]],
[[-1, 20], 1, Concat, [1]], # cat head P4
[-1, 3, C3, [256, False]], # 26 (P4/16-medium)

[-1, 1, Conv, [256, 3, 2]],
[[-1, 16], 1, Concat, [1]], # cat head P5
[-1, 3, C3, [512, False]], # 29 (P5/32-large)

[-1, 1, Conv, [512, 3, 2]],
[[-1, 12], 1, Concat, [1]], # cat head P6
[-1, 3, C3, [768, False]], # 32 (P6/64-xlarge)

[[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6)
]

[ppogg]

Hi, 朋友，非常欢迎使用这个仓库！
关于你的问题，你可以看以下博客:
https://zhuanlan.zhihu.com/p/410874403
里面介绍了我对于g模型的所有详细实验。
另外，对于你的问题，我有以下几点疑问:
①是否重参化？重参化的脚本rep_convert.py在script文件夹中
②是否使用了repvgg block fuse的功能，代码在yolo.py文件中

    def fuse(self): 
    # fuse repvgg block
        print('Fusing layers... ')
        for m in self.model.modules():
            # print(m)
            if type(m) is RepVGGBlock:
                if hasattr(m, 'rbr_1x1'):
                    # print(m)
                    kernel, bias = m.get_equivalent_kernel_bias()
                    rbr_reparam = nn.Conv2d(in_channels=m.rbr_dense.conv.in_channels,
                                            out_channels=m.rbr_dense.conv.out_channels,
                                            kernel_size=m.rbr_dense.conv.kernel_size,
                                            stride=m.rbr_dense.conv.stride,
                                            padding=m.rbr_dense.conv.padding, dilation=m.rbr_dense.conv.dilation,
                                            groups=m.rbr_dense.conv.groups, bias=True)
                    rbr_reparam.weight.data = kernel
                    rbr_reparam.bias.data = bias
                    for para in self.parameters():
                        para.detach_()
                    m.rbr_dense = rbr_reparam
                    # m.__delattr__('rbr_dense')
                    m.__delattr__('rbr_1x1')
                    if hasattr(self, 'rbr_identity'):
                        m.__delattr__('rbr_identity')
                    if hasattr(self, 'id_tensor'):
                        m.__delattr__('id_tensor')
                    m.deploy = True
                    m.forward = m.fusevggforward  # update forward
                # continue
                # print(m)
            if type(m) is Conv and hasattr(m, 'bn'):
                # print(m)
                m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
                delattr(m, 'bn')  # remove batchnorm
                m.forward = m.fuseforward  # update forward

③你测试的图片基数大概多少？推荐5000张以上的图片，分别测试bs＝1, bs＝16, bs＝32
④补充我之前做实验的一些见解，repvgg block reparam和fuse后，等价于3×3卷积，速度是一模一样的，在此已做过大量测试，详细可见上面的链接

[Hiwyl]

感谢！

1. 使用了rep_convert.py这个脚本

# YOLOv5 experimental modules
import torch
from utils.google_utils import attempt_download

if __name__ == "__main__":
    load = "runs/yolov5s6_repvgg/weights/best.pt"
    save = "runs/yolov5s6_repvgg/weights/best_deploy.pt"
    input_size = 640
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    for w in load if isinstance(load, list) else [load]:
        attempt_download(w)
        ckpt = torch.load(w, map_location=None)  # load

    torch.save(ckpt, save)

    print(f"Done. Befrom weights:({load})")
    print(f"Done. Befrom weights:({save})")

貌似不是这个脚本吧。

[ppogg]

是的，是这个脚本，他会调用yolo.py里面的函数接口，捕捉需要重参化的repvgg block

[Hiwyl]

我其他设置和您的一样，只是模型定义yaml使用的是四分支，训练结束后，先通过rep_convert再detect，相比于原始yolov5s6，使新模型的FLOPs是其1/4， bs1 detect了1000次 平均为12.5ms，原始yolov5s6为12ms

[ppogg]

检测的数据大概多少张呢

[Hiwyl]

1000

[ppogg]

好的，请问下使新模型的FLOPs是其1/4是什么意思呢，另外能否推理5000次试试，我对这个结果比较好奇

[Hiwyl]

好的，QQ群里回您

[ppogg]

好的

[ppogg]

修复了之前重参化脚本的漏洞：
#64 (comment)