class Focus(nn.Module): # Focus wh information into c-space def init(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups super().init() self.conv = Conv(c1 * 4, c2, k, s, p, g, act) # self.contract = Contract(gain=2) def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) # return self.conv(self.contract(x)) 详细分析这个程序的结构和实现的功能

这是一个名为 Focus 的 PyTorch 模块，实现了将输入的特征图进行空间上的压缩，即将高和宽的信息压缩到一半。具体实现过程如下：

- 构造函数 init 中，定义了该模块的输入通道数 c1，输出通道数 c2，卷积核大小 k，步长 s，padding p，分组数 g，以及是否使用激活函数 act。
- forward 函数中，首先将输入特征图 x 进行四分，即将每个 2x2 的区域合并成一个像素，得到大小为 (b, 4c, w/2, h/2) 的新特征图。然后，通过一个卷积层进行特征的降维，将通道数从 4c 降到 c2，得到输出特征图 y。

整个模块的作用就是将输入特征图中的高和宽信息进行压缩，减少计算量和参数数量。具体来说，对于下游任务，如果可以在一定程度上保留输入特征图的空间信息，那么模型的性能可能会更好。因此，这种空间信息的压缩操作被广泛应用于目标检测、图像分割等计算机视觉领域的任务中。

相关问题

class Focus(nn.Module): # Focus wh information into c-space def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups super().__init__() self.conv = Conv(c1 * 4, c2, k, s, p, g, act) # self.contract = Contract(gain=2) def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) # return self.conv(self.contract(x))

这是一个定义了一个叫做Focus的PyTorch模块。它的作用是将输入的宽高信息通过卷积运算变换到一个新的通道空间，通道数量为c2。具体来说，它将输入x按照空间位置进行4等分，并将这四个位置的信息在通道维度上进行拼接，得到一个4c1通道的张量，然后通过一个卷积层将其变换为一个c2通道的输出。这个卷积层的卷积核大小为k，步长为s，填充为p，分组数为g。如果act=True，则在卷积层后面加上一个激活函数。

yolov7 common.py 源码

以下是 YOLOv7 的 common.py 源码：

import math
import torch.nn.functional as F
import torch.nn as nn
import torch

def make_divisible(x, divisor):
    # Returns x evenly divisible by divisor
    return math.ceil(x / divisor) * divisor


def autopad(k, p=None):  # kernel, padding
    # Pad to 'same'
    if p is None:
        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]
    return p


class Conv(nn.Module):
    # Standard convolution
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):
        super(Conv, self).__init__()
        self.conv = nn.Conv2d(c1, c2, k, stride=s, padding=autopad(k, p), groups=g, bias=False)
        self.bn = nn.BatchNorm2d(c2)
        self.act = nn.Hardswish() if act else nn.Identity()

    def forward(self, x):
        return self.act(self.bn(self.conv(x)))


class Bottleneck(nn.Module):
    # Standard bottleneck
    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):
        super(Bottleneck, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_, c2, 3, 1, g=g)
        self.add = shortcut and c1 == c2
        self.identity = nn.Identity() if self.add else None

    def forward(self, x):
        return self.identity(x) + self.cv2(self.cv1(x))


class SPP(nn.Module):
    # Spatial pyramid pooling layer used in YOLOv3-SPP
    def __init__(self, c1, c2, k=(5, 9, 13)):
        super(SPP, self).__init__()
        c_ = c1 // 2  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])

    def forward(self, x):
        x = self.cv1(x)
        return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))


class DWConv(nn.Module):
    # Depthwise convolution
    def __init__(self, c1, c2, k=1, s=1, p=None):
        super(DWConv, self).__init__()
        self.conv = nn.Conv2d(c1, c1, k, stride=s, padding=autopad(k, p), groups=c1, bias=False)
        self.bn = nn.BatchNorm2d(c1)
        self.act = nn.Hardswish()
        self.project = nn.Conv2d(c1, c2, 1, bias=False)
        self.bn2 = nn.BatchNorm2d(c2)
        self.act2 = nn.Hardswish()

    def forward(self, x):
        return self.act2(self.bn2(self.project(self.act(self.bn(self.conv(x))))))


class Focus(nn.Module):
    # Focus wh information into c-space
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):
        super(Focus, self).__init__()
        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)

    def forward(self, x):
        # x(b,c,w,h) -> y(b,4c,w/2,h/2)
        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))


class Concat(nn.Module):
    # Concatenate a list of tensors along dimension
    def __init__(self, dimension=1):
        super(Concat, self).__init__()
        self.d = dimension

    def forward(self, x):
        return torch.cat(x, self.d)


class Detect(nn.Module):
    # Detect layer
    def __init__(self, nc, anchors):
        super(Detect, self).__init__()
        self.nc = nc  # number of classes
        self.no = nc + 5  # number of outputs per anchor
        self.na = len(anchors)  # number of anchors
        self.anchors = torch.tensor(anchors).float().view(self.na, -1)
        self.anchors /= self.anchors.sum(1).view(self.na, 1)  # normalized anchors
        self.register_buffer("anchor_grid", self.anchors.clone().view(1, -1, 1, 1))
        self.m = nn.Conv2d(self.no * self.na, self.no * self.na, 1)  # prediction conv

    def forward(self, x):
        # x(bs,255,h,w) -> p(bs,3,85,h,w)
        bs, _, ny, nx = x.shape
        device, dtype = x.device, x.dtype
        stride = self.anchor_grid.device / torch.tensor([nx, ny])[None, :, None, None].to(device)
        grid = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
        y = torch.stack(grid, 2).to(device).float()
        x = (x.sigmoid() * 2. - 0.5) * stride  # x(?,255,?,?) --sig--> x(?,255,?,?) --*2-0.5--> x(?,255,?,?) --*stride--> x(?,255,?,?)
        y = (y + 0.5) * stride  # y(?,2,?,?) --+0.5--> y(?,2,?,?) --*stride--> y(?,2,?,?)
        xy = torch.stack([x, y], 2).view(bs, 2, self.na * ny * nx).permute(0, 2, 1).contiguous().view(bs, self.na * ny * nx, 2)
        x = self.m(x.flatten(2).permute(0, 2, 1)).view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
        # x(bs,na,ny,nx,na) --view--> x(bs,na,ny,nx,no) --permute--> x(bs,na,ny,nx,no)
        if not self.training:
            x[..., 4:] = x[..., 4:].sigmoid()
            return x
        else:  # train
            return x, xy, self.anchor_grid.repeat(bs, 1, ny, nx)


class Model(nn.Module):
    # YOLOv7 model https://github.com/WongKinYiu/yolov7
    def __init__(self, nc=80, anchors=((10, 13), (16, 30), (33, 23), (30, 61), (62, 45), (59, 119), (116, 90), (156, 198), (373, 326)),
                 ch=[256, 512, 1024, 2048], depth=0.33):
        super(Model, self).__init__()
        assert depth in [0.33, 0.67, 1.0]
        self.depth = depth  # model depth multiplier
        self.grid = [torch.zeros(1)] * 5  # init grid
        self.stride = torch.tensor([8., 16., 32., 64., 128.])
        self.create_backbone(ch)
        self.create_neck()
        self.create_head(nc, anchors)

    def forward(self, x):
        z = []
        for i in range(5):
            x = self.backbone[i](x)
            z.append(x)
        x = self.neck(z)
        return self.head(x)

    def create_backbone(self, ch):
        # darknet backbone
        self.backbone = nn.ModuleList([Focus(3, ch[0], 3),
                                       Conv(ch[0], ch[1], 3, 2),
                                       Bottleneck(ch[1], ch[2]),
                                       Conv(ch[2], ch[3], 3, 2),
                                       Bottleneck(ch[3], ch[4]),
                                       Conv(ch[4], ch[5], 3, 2),
                                       SPP(ch[5], ch[5]),
                                       Bottleneck(ch[5], ch[6]),
                                       Conv(ch[6], ch[7], 1)])
        c2 = make_divisible(ch[7] * self.depth)  # ch_last
        self.backbone.append(Bottleneck(ch[7], c2, False))
        self.out_channels = [c2, ch[4], ch[2], ch[0]]

    def create_neck(self):
        # FPN-like attentional output
        self.neck = nn.Sequential(
            Concat(),
            Conv(self.out_channels[0], self.out_channels[0], 1),
            DWConv(self.out_channels[0], self.out_channels[1], 3, s=2),
            DWConv(self.out_channels[1], self.out_channels[2], 3, s=2),
            DWConv(self.out_channels[2], self.out_channels[3], 3, s=2),
            SPP(self.out_channels[3], self.out_channels[3]),
            DWConv(self.out_channels[3], self.out_channels[3], 3, dilation=3),
            DWConv(self.out_channels[3], self.out_channels[3], 3, dilation=3),
            DWConv(self.out_channels[3], self.out_channels[3], 3, dilation=3),
        )

    def create_head(self, nc, anchors):
        # detection head
        self.head = nn.Sequential(
            DWConv(self.out_channels[3], self.out_channels[3], 3, dilation=3),
            DWConv(self.out_channels[3], self.out_channels[3], 3, dilation=3),
            DWConv(self.out_channels[3], self.out_channels[3], 3, dilation=3),
            Concat(),
            Conv(self.out_channels[3] * 4, self.out_channels[3], 1),
            nn.Conv2d(self.out_channels[3], len(anchors) * (nc + 5), 1, bias=True),
            Detect(nc, anchors))


def attempt_load(weights, map_location=None, inplace=True):
    # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
    if isinstance(weights, (list, tuple)):
        # Load a list of models (ensemble)
        ensemble = nn.ModuleList()
        for w in weights:
            model = Model()
            model.to(next(w.parameters()).device)
            try:
                ckpt = torch.load(w, map_location=map_location)  # load
                state_dict = ckpt['model'].float().state_dict()  # to FP32
                state_dict = {k: v for k, v in state_dict.items() if model.state_dict()[k].shape == v.shape}  # filter
                model.load_state_dict(state_dict, strict=False)  # load
                print(f"Transferred {len(state_dict)} from {w}")
            except:
                print(f"Error loading {w}")
            ensemble.append(model.eval())
        return ensemble

    else:
        # Load a single model
        model = Model()
        model.to(next(weights.parameters()).device)
        try:
            ckpt = torch.load(weights, map_location=map_location)  # load
            state_dict = ckpt['model'].float().state_dict()  # to FP32
            state_dict = {k: v for k, v in state_dict.items() if model.state_dict()[k].shape == v.shape}  # filter
            model.load_state_dict(state_dict, strict=False)  # load
            print(f"Transferred {len(state_dict)} from {weights}")
        except:
            print(f"Error loading {weights}")
        return model.eval()