yolov5 6.1 增加anchors

根据引用[1]中的信息，Yolov5在配置文件中默认保存了一些针对coco数据集的预设锚定框。以yolov5s.yaml为例，预设的锚定框尺寸如下：
- [10,13, 16,30, 33,23] # P3/8
- [30,61, 62,45, 59,119] # P4/16
- [116,90, 156,198, 373,326] # P5/32

如果你想在Yolov5 6.1版本中增加anchors，你可以手动修改配置文件中的anchors部分，添加你想要的锚定框尺寸。这样可以根据你的需求来调整锚定框的大小和数量。请确保你的锚定框尺寸符合你的数据集和目标的特点。

相关问题

ResRep 剪枝方法 YOLOv5 6.1中的应用，完整详细代码复现

ResRep 剪枝方法是一种基于残差网络的剪枝方法，可以有效地减少神经网络模型的参数和计算量，同时保持模型的精度。

在 YOLOv5 6.1 版本中，已经应用了 ResRep 剪枝方法，可以通过以下步骤进行完整详细代码复现：

1. 下载 YOLOv5 6.1 版本的代码，并安装相关依赖。

git clone https://github.com/ultralytics/yolov5.git
cd yolov5/
pip install -r requirements.txt

2. 下载 COCO 数据集，并将其放在 `data/` 文件夹中。

mkdir data/
cd data/
wget https://github.com/ultralytics/yolov5/releases/download/v6.0/coco128.zip
unzip coco128.zip
cd ..

3. 在 `models/yolov5s.yaml` 文件中，修改模型的 `anchors`、`nc` 和 `depth_multiple` 参数。

anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32
nc: 80
depth_multiple: 0.33

4. 在 `train.py` 文件中，修改训练参数，包括 `epochs`、`batch-size` 和 `img-size` 等。

python train.py --img 640 --batch 16 --epochs 300 --data coco.yaml --cfg models/yolov5s.yaml --weights '' --name yolov5s_resrep

5. 在 `models/yolo.py` 文件中，添加 ResRep 剪枝方法的相关代码，包括 `resrep_prune()` 和 `forward` 函数的修改。

import torch.nn as nn
import torch.nn.functional as F
from models.common import Conv, Bottleneck, SPP, DWConv, Focus, Concat
from utils.torch_utils import time_synchronized

class ResRep(nn.Module):
    def __init__(self, model, prune_idx):
        super(ResRep, self).__init__()
        self.model = model
        self.prune_idx = prune_idx

    def forward(self, x):
        # Forward pass through the pruned model
        for i, m in enumerate(self.model):
            x = m(x)
            if i == self.prune_idx:
                break
        return x

    def resrep_prune(self, threshold):
        # Prune the model based on the threshold
        pruned_idx = []
        for i, m in enumerate(self.model):
            if isinstance(m, Bottleneck):
                if m.bn3.weight is not None:
                    mask = m.bn3.weight.data.abs().ge(threshold).float().cuda()
                    m.bn3.weight.data.mul_(mask)
                    m.bn3.bias.data.mul_(mask)
                    m.conv3.weight.data.mul_(mask.view(-1, 1, 1, 1))
                    pruned_idx.append(i)
            elif isinstance(m, Conv):
                if m.bn.weight is not None:
                    mask = m.bn.weight.data.abs().ge(threshold).float().cuda()
                    m.bn.weight.data.mul_(mask)
                    m.bn.bias.data.mul_(mask)
                    m.conv.weight.data.mul_(mask.view(-1, 1, 1, 1))
                    pruned_idx.append(i)
        self.prune_idx = max(pruned_idx)

class YOLOLayer(nn.Module):
    def __init__(self, anchors, nc, img_size, yolo_idx):
        super(YOLOLayer, self).__init__()
        self.anchors = torch.Tensor(anchors)
        self.na = self.anchors.shape[0]  # number of anchors
        self.nc = nc  # number of classes
        self.no = self.nc + 5  # number of outputs per anchor
        self.img_size = img_size
        self.grid_size = 0  # grid size
        self.stride = 0  # stride
        self.grid = self.create_grid(img_size)
        self.scaled_anchors = torch.Tensor([(a_w / self.stride, a_h / self.stride) for a_w, a_h in self.anchors])
        self.anchor_w = self.scaled_anchors[:, 0:1].view((1, self.na, 1, 1))
        self.anchor_h = self.scaled_anchors[:, 1:2].view((1, self.na, 1, 1))
        self.yolo_idx = yolo_idx

    def forward(self, x):
        # Residual block
        x = self.residual(x, 1)
        # Feature extraction
        x = self.extract(x, 2)
        # Pruning
        x = self.prune(x, 3)
        # Detection
        x = self.detect(x, 4)
        return x

    def residual(self, x, n):
        for i in range(n):
            x = self.m[i](x) + x
        return x

    def extract(self, x, n):
        for i in range(n):
            x = self.m[i](x)
        return x

    def prune(self, x, n):
        self.resrep.resrep_prune(threshold=0.1)
        x = self.resrep(x)
        return x

    def detect(self, x, n):
        # Predict on center
        io = x.clone()[:, :, self.grid_size[1] // 2:self.grid_size[1] // 2 + 1,
             self.grid_size[0] // 2:self.grid_size[0] // 2 + 1]
        io[..., 4] += self.grid_x
        io[..., 5] += self.grid_y
        io[..., :2] = self.sigmoid(io[..., :2]) * 2. - 0.5 + self.grid.to(x.device)
        io[..., 2:4] = (self.sigmoid(io[..., 2:4]) * 2) ** 2 * self.anchor_wh[self.anchor_vec == self.yolo_idx]
        io[..., :4] *= self.stride
        return io.view(io.shape[0], -1, self.no), x

class Model(nn.Module):
    def __init__(self, cfg='models/yolov5s.yaml', ch=3, nc=None):
        super(Model, self).__init__()
        self.model, self.save = parse_model(deepcopy(yaml.load(open(cfg, 'r'))))
        self.stride = torch.tensor([32, 16, 8])
        self.ch = ch
        self.nc = nc
        self.hyper_params = self.model.pop('hyper_params')
        self.init_weights()

    def forward(self, x):
        y, dt = [], []
        for i, m in enumerate(self.model):
            x = m(x)
            if i in [2, 4, 6]:
                y.append(x)
                dt.append(None)
        return y, dt

    def init_weights(self):
        # Initialize weights
        for m in self.modules():
            t = type(m)
            if t is Conv:
                pass  # nn.init.kaiming_normal_(m.conv.weight, mode='fan_out', nonlinearity='relu')
            elif t is DWConv:
                pass  # nn.init.kaiming_normal_(m.conv.weight, mode='fan_out', nonlinearity='relu')
                m.bn.weight.data.fill_(1.0)
                m.bn.bias.data.fill_(0)
            elif t is nn.BatchNorm2d:
                m.eps = 1e-3
                m.momentum = 0.03

    def prune(self, threshold):
        # Apply ResRep pruning to the model
        pruned_idx = []
        for i, m in enumerate(self.model):
            if isinstance(m, Bottleneck):
                if m.bn3.weight is not None:
                    mask = m.bn3.weight.data.abs().ge(threshold).float().cuda()
                    m.bn3.weight.data.mul_(mask)
                    m.bn3.bias.data.mul_(mask)
                    m.conv3.weight.data.mul_(mask.view(-1, 1, 1, 1))
                    pruned_idx.append(i)
            elif isinstance(m, Conv):
                if m.bn.weight is not None:
                    mask = m.bn.weight.data.abs().ge(threshold).float().cuda()
                    m.bn.weight.data.mul_(mask)
                    m.bn.bias.data.mul_(mask)
                    m.conv.weight.data.mul_(mask.view(-1, 1, 1, 1))
                    pruned_idx.append(i)
            elif isinstance(m, YOLOLayer):
                m.resrep = ResRep(m.m, self.prune_idx)
                m.resrep.resrep_prune(threshold=0.1)
                pruned_idx.append(i)
        self.prune_idx = max(pruned_idx)

    def fuse(self):
        # Fuse Conv+BN and Conv+ReLU into Conv
        print('Fusing layers...')
        for m in self.modules():
            if type(m) is Conv and type(m.bn) is nn.BatchNorm2d:
                m.conv = fuse_conv_bn(m.conv, m.bn)
                delattr(m, 'bn')
            elif type(m) is nn.Sequential:
                for i, v in enumerate(m):
                    if type(v) is Conv and type(v.bn) is nn.BatchNorm2d:
                        v.conv = fuse_conv_bn(v.conv, v.bn)
                        delattr(v, 'bn')
                    elif type(v) is Conv and hasattr(m[i + 1], 'act'):
                        v.conv = fuse_conv_relu(v.conv, m[i + 1].act)
                        delattr(m[i + 1], 'act')
                    elif type(v) is nn.BatchNorm2d and hasattr(m[i + 1], 'act'):
                        delattr(m[i + 1], 'act')
            elif type(m) is nn.BatchNorm2d:
                if not hasattr(m, 'act'):
                    m.act = nn.ReLU(inplace=True)

    def info(self, verbose=False):
        # Print model information
        model_info(self, verbose)

def parse_model(d, ch=3, nc=None):  # model_dict, input_channels, num_classes
    anchors, nc = d['anchors'], d['nc']
    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors
    layers, save, c2 = [], [], ch
    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):
        m = eval(m) if isinstance(m, str) else m  # eval strings
        for j, a in enumerate(args):
            try:
                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
            except:
                pass
        n = '' if n == 1 else n
        if m in [Conv, DWConv, Focus, Bottleneck, SPP, Concat, Detect]:
            c1, c2 = c2, args[0]
            if isinstance(c2, list):
                c2 = [ch] + c2
            elif c2 == 'same':
                c2 = c1
            elif c2 == -1:
                c2 = [256, 512, 1024, 2048][max(2 + i - len(d['backbone']), 0)]
            elif c2 == -2:
                c2 = c1 // 2
            elif c2 == -3:
                c2 = c1 // 3
            elif c2 == -4:
                c2 = c1 // 4
            else:
                c2 = int(c2)
            args = [c1, c2, *args[1:]]
            if m in [Bottleneck, SPP]:
                args.insert(2, n)
                n = ''
        elif m is nn.BatchNorm2d:
            args = [c2]
        elif m is nn.Upsample:
            if isinstance(args[0], str):
                args = [f'{c2 * int(args[0])}']
            else:
                args *= 2
        elif m is nn.Linear:
            args = [nc, args[0]]
            if n == 'head':
                args[0] = args[0] * na * (nc + 5)
                n = ''
        elif m is Detect:
            args.append([anchors[i] for i in d['anchor_idx'][f]])
            args.append(nc)
            args.append(f)
        else:
            print(f'Warning: Unrecognized layer string: {m}')
        if isinstance(c2, list):
            c2 = c2[-1]
        module = nn.Sequential(*[m(*args) if m is not Detect else Detect(*args[:3]).to(f'cuda:{args[3]}') for m in [m]])
        module.nc = nc  # attach number of classes to Detect()
        module.stride = torch.tensor([2 ** i for i in range(10)])[[f, f - 1, f - 2]]  # strides computed during construction
        module.anchor_vec = d['anchor_idx'][f]
        module.training = False
        layers.append(module)
        if n:
            save.append(n)
    return nn.Sequential(*layers), sorted(save)

def fuse_conv_bn(conv, bn):
    # https://tehnokv.com/posts/fusing-batchnorm-and-conv/
    with torch.no_grad():
        # init
        fusedconv = Conv(
            conv.in_channels,
            conv.out_channels,
            kernel_size=conv.kernel_size,
            stride=conv.stride,
            padding=conv.padding,
            groups=conv.groups,
            bias=True,
            dilation=conv.dilation)
        fusedconv.weight.data = conv.weight.data.clone().reshape(
            fusedconv.weight.data.shape)  # copy conv weights

        # prepare filters
        bnmean = bn.running_mean
        bnstd = torch.sqrt(bn.running_var + bn.eps)
        if conv.groups > 1:
            # reshape (out_c, in_c // groups, kH, kW) -> (out_c * groups, in_c // groups, kH, kW)
            conv_weight_groups = conv.weight.data.reshape(
                conv.out_channels * conv.groups, -1, conv.kernel_size[0], conv.kernel_size[1])
            # reshape bn params (out_c) -> (out_c * groups)
            bnmean = bnmean.repeat(conv.groups)
            bnstd = bnstd.repeat(conv.groups)
        else:
            conv_weight_groups = conv.weight.data

        # fuse
        fusedconv.bias.data = bn.bias.data + (bn.weight.data / bnstd) * (conv.bias.data - bnmean)
        scale = (bn.weight.data / bnstd)
        fusedconv.weight.data *= scale.reshape(-1, 1, 1, 1)

        return fusedconv

def fuse_conv_relu(conv, relu):
    # Fuse Conv+ReLU into Conv
    with torch.no_grad():
        # init
        fusedconv = Conv(
            conv.in_channels,
            conv.out_channels,
            kernel_size=conv.kernel_size,
            stride=conv.stride,
            padding=conv.padding,
            groups=conv.groups,
            bias=True,
            dilation=conv.dilation)
        fusedconv.weight.data = conv.weight.data.clone().reshape(
            fusedconv.weight.data.shape)  # copy conv weights

        # fuse
        fusedconv.bias.data = conv.bias.data
        fusedconv.weight.data *= relu.inplace_slope.reshape(-1, 1, 1, 1)

        return fusedconv

6. 在 `train.py` 文件中，添加 ResRep 剪枝方法的调用。

# ResRep pruning
if epoch == 100:
    model.prune(threshold=0.1)
    print(f'Pruned model to {count_parameters(model)[0] / 1e6:.3g}M parameters')

7. 运行训练命令，开始训练。

python train.py --img 640 --batch 16 --epochs 300 --data coco.yaml --cfg models/yolov5s.yaml --weights '' --name yolov5s_resrep

完成以上步骤后，即可得到应用了 ResRep 剪枝方法的 YOLOv5 6.1 版本的模型，并进行训练。