YOLOv5代码解读[02] models/yolov5l.yaml文件解析

yolov5l.yaml文件

# YOLOv5 🚀 by Ultralytics, GPL-3.0 license

# Parameters
nc: 27  # number of classes
depth_multiple: 1.0  # model depth multiple
width_multiple: 1.0  # layer channel multiple
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

# YOLOv5 v6.0 backbone
backbone:
  # [from, number, module, args]
  [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
   [-1, 3, C3, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
   [-1, 6, C3, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
   [-1, 9, C3, [512]],
   [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
   [-1, 3, C3, [1024]],
   [-1, 1, SPPF, [1024, 5]],  # 9
  ]

# YOLOv5 v6.0 head
head:
  [[-1, 1, Conv, [512, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 3, C3, [512, False]],  # 13

   [-1, 1, Conv, [256, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 3, C3, [256, False]],  # 17 (P3/8-small)

   [-1, 1, Conv, [256, 3, 2]],
   [[-1, 14], 1, Concat, [1]],  # cat head P4
   [-1, 3, C3, [512, False]],  # 20 (P4/16-medium)

   [-1, 1, Conv, [512, 3, 2]],
   [[-1, 10], 1, Concat, [1]],  # cat head P5
   [-1, 3, C3, [1024, False]],  # 23 (P5/32-large)

   [[17, 20, 23], 1, Detect, [nc, anchors, False]],  # Detect(P3, P4, P5)
  ]

检测头1—>耦合头

class Detect(nn.Module):
    stride = None  
    onnx_dynamic = False
    export = False
    def __init__(self, nc=80, anchors=(), Decoupled=False, ch=(), inplace=True):  
        super().__init__()
        # 是否解耦头
        self.decoupled = Decoupled
        # 类别数目
        self.nc = nc  
        # 每个anchor输出维度 
        self.no = nc + 5  
        # 检测层的输出数量（不同尺度个数） 
        self.nl = len(anchors)  
        # 每个尺度特征图的anchor数量
        self.na = len(anchors[0]) // 2  
        # 初始化步长init grid
        self.grid = [torch.zeros(1)] * self.nl    
        # 初始化anchor grid
        self.anchor_grid = [torch.zeros(1)] * self.nl  
        # self.register_buffer("a", torch.ones(2,3))  
        # register_buffer的作用是将torch.ones(2,3)这个tensor注册到模型的buffers()属性中，并命名为a，
        # 这代表a对应的是一个持久态，不会有梯度传播给它，但是能被模型的state_dict记录下来，可以理解为模型的常数。
        self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2))  # (3,3,2) == (nl,na,2)
        # 检测头head输出卷积
        # 如果是解耦头
        if self.decoupled:
            self.m = nn.ModuleList(DecoupledHead(x, self.nc, anchors) for x in ch) 
        # 如果是耦合头
        else:
            self.m = nn.ModuleList(nn.Conv2d(x, self.no*self.na, 1) for x in ch) 
        # use in-place ops (e.g. slice assignment)
        self.inplace = inplace  
        
    def forward(self, x):
        # inference output
        z = []
        # 对于每个尺度的特征图来说
        for i in range(self.nl):
            # conv
            # P3: [1, 128, 80, 80]->[1, 3*(nc+5), 80, 80]
            # P4: [1, 256, 40, 40]->[1, 3*(nc+5), 40, 40]
            # P5: [1, 512, 20, 20]->[1, 3*(nc+5), 20, 20]
            x[i] = self.m[i](x[i])
            # 以coco数据集为例，x(bs,255,20,20) -> x(bs,3,20,20,85)   (x,y,w,h,c,c1,c2,.........)
            bs, _, ny, nx = x[i].shape
            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()

            # 推断过程inference
            if not self.training:
                # self.grid: [tensor([0.]), tensor([0.]), tensor([0.])]
                if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:
                    self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)

                y = x[i].sigmoid()

                if self.inplace:
                    # 中心点xy 网格grid
                    y[..., 0:2] = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]
                    # 长宽wh  锚anchor_grid
                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]
                else:
                    xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]
                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]
                    y = torch.cat((xy, wh, y[..., 4:]), -1)
                z.append(y.view(bs, -1, self.no))

        return x if self.training else (torch.cat(z, 1),) if self.export else (torch.cat(z, 1), x)
    
    # # 转成caffe时候的代码
    # def forward(self, x):
    #     # inference output
    #     z = []
    #     # 对于每个尺度的特征图来说
    #     for i in range(self.nl):
    #         # conv
    #         # P3: [1, 128, 80, 80]->[1, 3*(nc+5), 80, 80]
    #         # P4: [1, 256, 40, 40]->[1, 3*(nc+5), 40, 40]
    #         # P5: [1, 512, 20, 20]->[1, 3*(nc+5), 20, 20]
    #         x[i] = self.m[i](x[i])
    #         # y = x[i]
    #         y = x[i].sigmoid()
    #         z.append(y)
    #     return z

    def _make_grid(self, nx=20, ny=20, i=0, torch_1_10=check_version(torch.__version__, '1.10.0')):
        d = self.anchors[i].device
        t = self.anchors[i].dtype
        y, x = torch.arange(ny, device=d, dtype=t), torch.arange(nx, device=d, dtype=t)
        # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility
        if torch_1_10:
            yv, xv = torch.meshgrid(y, x, indexing='ij')
        else:
            yv, xv = torch.meshgrid(y, x)
        # 网格grid (x, y)
        # x[i] --> (bs,3,ny,nx,85)
        # grid --> (1,3,ny,nx,2)
        grid = torch.stack((xv, yv), 2).expand((1, self.na, ny, nx, 2))
        # 锚anchor (w, h)
        # x[i] --> (bs,3,ny,nx,85)
        # anchor_grid --> (1,3,ny,nx,2)
        # self.stride: tensor([ 8., 16., 32.])
        anchor_grid = (self.anchors[i].clone() * self.stride[i]).view((1, self.na, 1, 1, 2)).expand((1, self.na, ny, nx, 2))
        return grid, anchor_grid

检测头2—>解耦头

class DecoupledHead(nn.Module):
    def __init__(self, ch=256, nc=80, anchors=()):
        super().__init__()
        # 类别个数
        self.nc = nc
        # 检测层的数量
        self.nl = len(anchors)
        # 每一层anchor个数
        self.na = len(anchors[0]) // 2
        self.merge = Conv(ch, 128 , 1, 1)  # 默认256
        self.cls_convs1 = Conv(128, 64, 3, 1, 1)
        self.cls_convs2 = Conv(64, 64, 3, 1, 1)
        self.reg_convs1 = Conv(128, 64, 3, 1, 1)
        self.reg_convs2 = Conv(64, 64, 3, 1, 1)
        self.cls_preds = nn.Conv2d(64 , self.nc*self.na, 1)
        self.reg_preds = nn.Conv2d(64 , 4*self.na, 1)
        self.obj_preds = nn.Conv2d(64 , 1*self.na, 1)

    def forward(self, x):
        x = self.merge(x)
        x1 = self.cls_convs1(x)
        x1 = self.cls_convs2(x1)
        x1 = self.cls_preds(x1)
        x2 = self.reg_convs1(x)
        x2 = self.reg_convs2(x2)
        x21 = self.reg_preds(x2)
        x22 = self.obj_preds(x2)
        out = torch.cat([x21, x22, x1], 1)
        return out

检测头3—>ASFF检测头

class ASFF_Detect(nn.Module):  
    stride = None  
    onnx_dynamic = False   
    def __init__(self, nc=80, anchors=(), ch=(), multiplier=0.5, rfb=False, inplace=True):  
        super().__init__()
        # 类别数目
        self.nc = nc  
        # 每个anchor输出维度
        self.no = nc + 5  
        # 检测层的输出数量（不同尺度个数） 
        self.nl = len(anchors) 
        # 每个尺度特征图的anchor数量
        self.na = len(anchors[0]) // 2  
        # 初始化步长init grid
        self.grid = [torch.zeros(1)] * self.nl  
        # init anchor grid
        self.anchor_grid = [torch.zeros(1)] * self.nl
        # self.register_buffer("a", torch.ones(2,3))  
        # register_buffer的作用是将torch.ones(2,3)这个tensor注册到模型的buffers()属性中，并命名为a，
        # 这代表a对应的是一个持久态，不会有梯度传播给它，但是能被模型的state_dict记录下来，可以理解为模型的常数。
        self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2))  # (3,3,2) == (nl,na,2)
        # ASFF模块
        self.l0_fusion = ASFFV5(level=0, multiplier=multiplier, rfb=rfb)
        self.l1_fusion = ASFFV5(level=1, multiplier=multiplier, rfb=rfb)
        self.l2_fusion = ASFFV5(level=2, multiplier=multiplier, rfb=rfb)
        # 检测头head输出卷积
        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  
        # use in-place ops (e.g. slice assignment)
        self.inplace = inplace  

    def forward(self, x):
        # inference output
        z = []  
        result = []
        result.append(self.l2_fusion(x))
        result.append(self.l1_fusion(x))
        result.append(self.l0_fusion(x))
        x = result    
        
        # 对于每个尺度的特征图来说
        for i in range(self.nl):
            # conv 
            # P3: [1, 128, 80, 80]->[1, 3*(nc+5), 80, 80]
            # P4: [1, 256, 40, 40]->[1, 3*(nc+5), 40, 40]
            # P5: [1, 512, 20, 20]->[1, 3*(nc+5), 20, 20]
            x[i] = self.m[i](x[i])  
            # 以coco数据集为例，x(bs,255,20,20) -> x(bs,3,20,20,85)   (x,y,w,h,c,c1,c2,.........)
            bs, _, ny, nx = x[i].shape  
            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
            
            # 推断过程inference 
            if not self.training:  
                # self.grid: [tensor([0.]), tensor([0.]), tensor([0.])]
                if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:
                    self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)

                y = x[i].sigmoid()
              
                # 这块xy的计算存在大量疑惑？？？？？？？？？？？？？？？？？？？？？？？？？
                if self.inplace:
                    # 中心点xy 网格grid
                    y[..., 0:2] = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  
                    # 长宽wh  锚anchor_grid
                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] 
                else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
                    xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  
                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  
                    y = torch.cat((xy, wh, y[..., 4:]), -1)
                z.append(y.view(bs, -1, self.no))
        
        return x if self.training else (torch.cat(z, 1), x)
    
    def _make_grid(self, nx=20, ny=20, i=0):
        d = self.anchors[i].device
        if check_version(torch.__version__, '1.10.0'):  # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility
            yv, xv = torch.meshgrid([torch.arange(ny, device=d), torch.arange(nx, device=d)], indexing='ij')
        else:
            yv, xv = torch.meshgrid([torch.arange(ny, device=d), torch.arange(nx, device=d)])
        # 网格grid (x, y)
        # x[i] --> (bs,3,ny,nx,85)
        # grid --> (1,3,ny,nx,2)
        grid = torch.stack((xv, yv), 2).expand((1, self.na, ny, nx, 2)).float()
        # 锚anchor (w, h)
        # x[i] --> (bs,3,ny,nx,85)
        # anchor_grid --> (1,3,ny,nx,2)
        # self.stride: tensor([ 8., 16., 32.])
        anchor_grid = (self.anchors[i].clone() * self.stride[i]).view((1, self.na, 1, 1, 2)).expand((1, self.na, ny, nx, 2)).float()
        return grid, anchor_grid

Model类解析

class Model(nn.Module):
    def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None):  
        super().__init__()
        # 字典dict类型
        if isinstance(cfg, dict):
            self.yaml = cfg  
        # yaml文件
        else: 
            self.yaml_file = Path(cfg).name
            # 用ascii编码，忽略错误的形式打开文件cfg
            with open(cfg, encoding='ascii', errors='ignore') as f:
                self.yaml = yaml.safe_load(f)  
        
        # 输入通道
        ch = self.yaml['ch'] = self.yaml.get('ch', ch)  
        # 重写yaml文件中的nc
        if nc and nc != self.yaml['nc']:
            LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
            self.yaml['nc'] = nc  
        # 重写yaml文件中的anchors 
        if anchors:
            LOGGER.info(f'Overriding model.yaml anchors with anchors={anchors}')
            self.yaml['anchors'] = round(anchors)  
        
        # 根据yaml文件的model_dict解析模型
        self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) 
        # 默认类别名字 从0到nc-1
        self.names = [str(i) for i in range(self.yaml['nc'])] 
        self.inplace = self.yaml.get('inplace', True)
       
        # 设置Detect()中的inplace, stride, anchors
        m = self.model[-1]  
        if isinstance(m, Detect) or isinstance(m, ASFF_Detect):
            s = 256
            m.inplace = self.inplace
            # 根据前向传播forward 计算步长stride
            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])
            # 把anchors放缩到了3个不同的尺度
            # 这块的形状为什么这样变化？？？？？？
            m.anchors /= m.stride.view(-1, 1, 1)
            # 根据YOLOv5 Detect()模块m的步幅顺序检查给定锚框顺序，必要时进行纠正。
            check_anchor_order(m)
            self.stride = m.stride
            if m.decoupled:
                LOGGER.info('decoupled done')
                pass 
            else:
                self._initialize_biases()  # only run once  

        # 初始化权重weights和偏置biases
        initialize_weights(self)
        self.info()
        LOGGER.info('')

    def forward(self, x, augment=False, profile=False, visualize=False):
        # 推断时增强augmented inference
        if augment:
            return self._forward_augment(x)  
        # 单尺度推断single-scale inference 或者训练train
        return self._forward_once(x, profile, visualize)  

    def _forward_augment(self, x):
        # height, width
        img_size = x.shape[-2:]  
        s = [1, 0.83, 0.67]  # scales
        f = [None, 3, None]  # flips (2-ud, 3-lr)
        y = []  # outputs
        for si, fi in zip(s, f):
            xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
            yi = self._forward_once(xi)[0]  # forward
            # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1])  # save
            yi = self._descale_pred(yi, fi, si, img_size)
            y.append(yi)
        y = self._clip_augmented(y)  # clip augmented tails
        return torch.cat(y, 1), None  # augmented inference, train

    def _forward_once(self, x, profile=False, visualize=False):
        y, dt = [], []  
        for m in self.model:
            # 输入不是来自于上一个层的输出
            if m.f != -1:  
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]
            if profile:
                self._profile_one_layer(m, x, dt)
            # 计算输出
            x = m(x)
            y.append(x if m.i in self.save else None) 
            # 特征可视化
            if visualize:
                feature_visualization(x, m.type, m.i, save_dir=visualize)
        return x

    def _descale_pred(self, p, flips, scale, img_size):
        # de-scale predictions following augmented inference (inverse operation)
        if self.inplace:
            p[..., :4] /= scale  # de-scale
            if flips == 2:
                p[..., 1] = img_size[0] - p[..., 1]  # de-flip ud
            elif flips == 3:
                p[..., 0] = img_size[1] - p[..., 0]  # de-flip lr
        else:
            x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale  # de-scale
            if flips == 2:
                y = img_size[0] - y  # de-flip ud
            elif flips == 3:
                x = img_size[1] - x  # de-flip lr
            p = torch.cat((x, y, wh, p[..., 4:]), -1)
        return p

    def _clip_augmented(self, y):
        # Clip YOLOv5 augmented inference tails
        nl = self.model[-1].nl  # number of detection layers (P3-P5)
        g = sum(4 ** x for x in range(nl))  # grid points
        e = 1  # exclude layer count
        i = (y[0].shape[1] // g) * sum(4 ** x for x in range(e))  # indices
        y[0] = y[0][:, :-i]  # large
        i = (y[-1].shape[1] // g) * sum(4 ** (nl - 1 - x) for x in range(e))  # indices
        y[-1] = y[-1][:, i:]  # small
        return y

    def _profile_one_layer(self, m, x, dt):
        c = isinstance(m, Detect) or isinstance(m, ASFF_Detect) # is final layer, copy input as inplace fix
        o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0  # FLOPs
        t = time_sync()
        for _ in range(10):
            m(x.copy() if c else x)
        dt.append((time_sync() - t) * 100)
        if m == self.model[0]:
            LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s}  {'module'}")
        LOGGER.info(f'{dt[-1]:10.2f} {o:10.2f} {m.np:10.0f}  {m.type}')
        if c:
            LOGGER.info(f"{sum(dt):10.2f} {'-':>10s} {'-':>10s}  Total")

    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
        # https://arxiv.org/abs/1708.02002 section 3.3
        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
        m = self.model[-1]  
        # mi--> Conv2d(128, 255, kernel_size=(1, 1), stride=(1, 1)) 
        # s --> tensor(8.)
        for mi, s in zip(m.m, m.stride):  
            # conv.bias(255) to (3,85)
            b = mi.bias.view(m.na, -1)  
            b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
            b.data[:, 5:] += math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # cls
            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)

    def _print_biases(self):
        m = self.model[-1]  
        for mi in m.m:  
            b = mi.bias.detach().view(m.na, -1).T  
            LOGGER.info(('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean()))

    def _print_weights(self):
        for m in self.model.modules():
            if type(m) is Bottleneck:
                LOGGER.info('%10.3g' % (m.w.detach().sigmoid() * 2))  # shortcut weights

    def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers
        LOGGER.info('Fusing layers... ')
        for m in self.model.modules():
            if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'):
                m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
                delattr(m, 'bn')  # remove batchnorm
                m.forward = m.forward_fuse  # update forward
        self.info()
        return self

    def info(self, verbose=False, img_size=640):  
        # 打印模型信息
        model_info(self, verbose, img_size)

    def _apply(self, fn):
        # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers
        self = super()._apply(fn)
        m = self.model[-1]  # Detect()
        if isinstance(m, Detect) or isinstance(m, ASFF_Detect) or isinstance(m, Decoupled_Detect):
            m.stride = fn(m.stride)
            m.grid = list(map(fn, m.grid))
            if isinstance(m.anchor_grid, list):
                m.anchor_grid = list(map(fn, m.anchor_grid))
        return self

parse_model函数

def parse_model(d, ch):  
    # model_dict, input_channels(3)
    LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10}  {'module':<40}{'arguments':<30}")
    # nc:类别数; gd:'depth_multiple'; gw:'width_multiple'
    anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
    # anchor数目, 每层为3
    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors 
    # 每层的输出，na*(classes+5)
    no = na * (nc + 5)  
                                        
    # layers, savelist, ch_out
    layers, save, c2 = [], [], ch[-1] 
    # from, number, module, args
    # 以[-1, 1, Conv, [64, 6, 2, 2]为例, ch=[3], f=-1, n=1, m=Conv, args=[64, 6, 2, 2]
    #   [-1, 1, Conv, [128, 3, 2]
    #   [-1, 3, C3, [128]]
    #   [-1, 1, SPPF, [1024, 5]]
    #   [-1, 1, nn.Upsample, [None, 2, 'nearest']]
    #   [[-1, 6], 1, Concat, [1]]
    #   [-1, 3, C3, [512, False]]
    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):
        # 把strings转为本身的类型
        m = eval(m) if isinstance(m, str) else m  
        for j, a in enumerate(args):
            try:
                # 列表形式
                args[j] = eval(a) if isinstance(a, str) else a  
            except NameError:
                pass
        
        # depth_gain 深度缩放因子
        n = n_ = max(round(n*gd), 1) if n > 1 else n 
    
        # 对于不同类型的卷积模块   
        if m in [Conv, DWConv,  CrossConv, GhostConv, Bottleneck, GhostBottleneck,
                 BottleneckCSP, MobileBottleneck, SPP, SPPF, MixConv2d, Focus,
                 InvertedResidual, ConvBNReLU, C3, C3TR, C3SPP, C3Ghost, CoordAtt,
                 CoordAttv2, OSA_Stage]:
            # i=0, c1=3,  c2=64;  
            # i=1, c1=32, c2=128;  
            # i=2, c1=64, c2=128;
            # c1输入通道；c2输出通道；
            c1, c2 = ch[f], args[0]
            
            # width_gain 宽度缩放因子
            # 说明不是输出
            if c2 != no:  
                # 输出通道数必须为8的倍数
                c2 = make_divisible(c2*gw, 8)
            
            # i=0, [3,  32, 6, 2, 2]
            # i=1, [32, 64, 3, 2]
            # i=2, [64, 64]
            args = [c1, c2, *args[1:]]

            # 堆叠次数number of repeats
            # 注意网络设计理念：stage ---> block ---> layer
            if m in [BottleneckCSP, C3, C3TR, C3Ghost]:
                args.insert(2, n)  
                n = 1
        elif m is nn.BatchNorm2d:
            args = [ch[f]]
        elif m is Concat:
            c2 = sum(ch[x] for x in f)
        elif m is Detect:
            args.append([ch[x] for x in f])
            if isinstance(args[1], int):  # number of anchors
                args[1] = [list(range(args[1] * 2))] * len(f)
        elif m is ASFF_Detect :
            args.append([ch[x] for x in f])
            if isinstance(args[1], int):  # number of anchors
                args[1] = [list(range(args[1] * 2))] * len(f) 
        elif m is Contract:
            c2 = ch[f] * args[0] ** 2
        elif m is Expand:
            c2 = ch[f] // args[0] ** 2
        elif m is ConvNeXt_Block:
            c2 = args[0]
            args = args[1:]
        else:
            c2 = ch[f]
        
        # module
        # Conv(3, 32, 6, 2, 2]
        m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) 
        
        # m ===> <class 'models.common.Conv'>
        # str(m)[8:-2] ===> models.common.Conv
        t = str(m)[8:-2].replace('__main__.', '')  
        # 参数(parameters)/模型参数, 由模型通过学习得到的变量，比如权重和偏置.
        # m_.parameters(): <generator object Module.parameters at 0x7fcf4c2059d0>
        np = sum(x.numel() for x in m_.parameters()) 

        # attach index, 'from' index, type, number params
        m_.i, m_.f, m_.type, m_.np = i, f, t, np  
       
        LOGGER.info(f'{i:>3}{str(f):>18}{n_:>3}{np:10.0f}  {t:<40}{str(args):<30}')  
        
        # savelist  [6, 4, 14, 10, 17, 20, 23]
        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  
        # layers列表
        layers.append(m_)
        if i == 0:
            ch = []
        # ch列表
        ch.append(c2)

    return nn.Sequential(*layers), sorted(save)