![](https://images.cnblogs.com/OutliningIndicators/ContractedBlock.gif)
import math import torch import torch.nn as nn import yaml import nn_utils # 制造一个整除的数字 def make_divisible(x, divisor): return math.ceil(x / divisor) * divisor def autopad(kernel_size, padding=None): # Pad to 'same' if padding is None: padding = kernel_size // 2 return padding class Conv(nn.Module): """ 标准卷积 """ def __init__(self, input_channel, output_channel, kernel_size=1, stride=1, padding=None, groups=1, activation=True): super().__init__() self.conv = nn.Conv2d(input_channel, output_channel, kernel_size, stride, padding=autopad(kernel_size, padding), groups=groups, bias=False) self.bn = nn.BatchNorm2d(output_channel) self.act = nn.LeakyReLU(0.1, inplace=True) if activation else nn.Identity() def forward(self, x): return self.act(self.bn(self.conv(x))) class Bottleneck(nn.Module): """ 标准瓶颈层 """ def __init__(self, input_channel, output_channel, shortcut=True, groups=1, expansion=0.5): super(Bottleneck, self).__init__() hidden_channel = int(output_channel * expansion) self.conv1 = Conv(input_channel, hidden_channel) self.conv2 = Conv(hidden_channel, output_channel, 3, 1, groups=groups) self.add = shortcut and input_channel == output_channel def forward(self, x): if self.add: return x + self.conv2(self.conv1(x)) else: return self.conv2(self.conv1(x)) class BottleneckCSP(nn.Module): def __init__(self, input_channel, output_channel, repeats=1, shortcut=True, groups=1, expansion=0.5): super().__init__() hidden_channel = int(output_channel * expansion) self.conv1 = Conv(input_channel, hidden_channel, 1, 1) self.conv2 = nn.Conv2d(input_channel, hidden_channel, 1, 1, bias=False) self.conv3 = nn.Conv2d(hidden_channel, hidden_channel, 1, 1, bias=False) self.conv4 = Conv(2 * hidden_channel, output_channel, 1, 1) self.bn = nn.BatchNorm2d(2 * hidden_channel) self.act = nn.LeakyReLU(0.1, inplace=True) self.repeat_blocks = nn.Sequential(*[ Bottleneck(hidden_channel, hidden_channel, shortcut, groups, expansion=1.0) for _ in range(repeats) ]) def forward(self, x): y1 = self.conv3(self.repeat_blocks(self.conv1(x))) y2 = self.conv2(x) ycat = torch.cat((y1, y2), dim=1) return self.conv4(self.act(self.bn(ycat))) class SPP(nn.Module): def __init__(self, input_channel, output_channel, kernel_size_list=(5, 9, 13)): super().__init__() hidden_channel = input_channel // 2 self.conv1 = Conv(input_channel, hidden_channel, 1, 1) self.conv2 = Conv(hidden_channel * (len(kernel_size_list) + 1), output_channel, 1, 1) self.spatial_pyramid_poolings = nn.Sequential(*[ nn.MaxPool2d(kernel_size=kernel_size, stride=1, padding=kernel_size // 2) for kernel_size in kernel_size_list ]) def forward(self, x): x = self.conv1(x) spp = torch.cat([x] + [m(x) for m in self.spatial_pyramid_poolings], dim=1) return self.conv2(spp) class Focus(nn.Module): """ 一种无损下采样,通过切片进行2倍下采样 input[1,3,100,100] out[1,12,50,50] """ def __init__(self, in_channel, out_channel, kernel_size=1, stride=1, padding=None, groups=1, activation=True): super().__init__() self.conv = Conv(in_channel * 4, out_channel, kernel_size, stride, padding, groups, activation) def forward(self, x): # block(y, x) # a(0, 0) b(1, 0) # c(0, 1) d(1, 1) a = x[..., ::2, ::2] b = x[..., 1::2, ::2] c = x[..., ::2, 1::2] d = x[..., 1::2, 1::2] return self.conv(torch.cat([a, b, c, d], dim=1)) class Concat(nn.Module): def __init__(self, dimension=1): super().__init__() self.dimension = dimension def forward(self, x): return torch.cat(x, dim=self.dimension) class Detect(nn.Module): def __init__(self, num_classes, num_anchor, reference_channels): super().__init__() self.num_anchor = num_anchor self.num_classes = num_classes self.num_output = self.num_classes + 5 self.heads = nn.ModuleList([ nn.Conv2d(input_channel, self.num_output * self.num_anchor, 1) for input_channel in reference_channels ]) def forward(self, x): for ilevel, head in enumerate(self.heads): x[ilevel] = head(x[ilevel]) return x class Yolo(nn.Module): def __init__(self, num_classes, config_file): super(Yolo, self).__init__() self.num_classes = num_classes # self.model, self.save_layer_index, self.anchors = self.build_model(config_file, num_classes) self.model, self.saved_index, self.anchors = self.build_model(num_classes, config_file) def forward(self, x): y = [] for module_instance in self.model: # 如果是Concat获得Detect层,输入数据需要拼接 if module_instance.from_index != -1: if isinstance(module_instance.from_index, int): x = y[module_instance.from_index] else: xout = [] for from_index in module_instance.from_index: if from_index == -1: xvalue = x else: xvalue = y[from_index] xout.append(xvalue) x = xout x = module_instance(x) # 保存需要拼接的x if module_instance.layer_index in self.saved_index: y.append(x) else: y.append(None) return x def parse_string(self, value): if value == "None": return None elif value == "True": return True elif value == "False": return False else: return value def build_model(self, num_classes, config_file, input_channel=3): with open(config_file) as f: self.yaml = yaml.load(f, Loader=yaml.FullLoader) layers_cfg_list = self.yaml["backbone"] + self.yaml["head"] anchors, depth_multiple, width_multiple = [self.yaml[item] for item in ["anchors", "depth_multiple", "width_multiple"]] num_anchor_per_level = len(anchors[0]) // 2 num_output_per_level = (5 + num_classes) * num_anchor_per_level # 用来存储所有的输入通道数 layers_channel = [input_channel] # 用来存储所有的通道 layers = [] # 需要concant的参数 saved_layer_index = [] for layer_index, (from_index, repeat, module_name, args) in enumerate(layers_cfg_list): args = [self.parse_string(item) for item in args] module_class_reference = eval(module_name) if repeat > 1: repeat = max(round(repeat * depth_multiple), 1) if module_class_reference in [Conv, Bottleneck, BottleneckCSP, SPP, Focus]: input_channel = layers_channel[from_index] output_channel = args[0] if output_channel != num_output_per_level: output_channel = make_divisible(output_channel * width_multiple, 8) # 网络的输出参数 args = [input_channel, output_channel, *args[1:]] if module_class_reference in [BottleneckCSP]: args.insert(2, repeat) repeat = 1 elif module_class_reference is Concat: output_channel = 0 for index in from_index: if index != -1: index += 1 output_channel += layers_channel[index] elif module_class_reference is Detect: reference_channel = [layers_channel[index + 1] for index in from_index] args = [num_classes, num_anchor_per_level, reference_channel] else: output_channel = layers_channel[from_index] if repeat > 1: module_instance = nn.ModuleList([ module_class_reference(*args) for _ in range(repeat) ]) else: module_instance = module_class_reference(*args) print(f"{layer_index}, input = {input_channel}, " f"output = {output_channel}, from_index = {from_index}, repeat = {repeat}") module_instance.from_index = from_index module_instance.layer_index = layer_index layers.append(module_instance) layers_channel.append(output_channel) if not isinstance(from_index, list): from_index = [from_index] # 前向传播的时候需要保存Concat或者Detect 需要拼接的参数 saved_layer_index.extend(filter(lambda x: x != -1, from_index)) # print(f"saved_layer_index {saved_layer_index}") return nn.Sequential(*layers), sorted(saved_layer_index), anchors if __name__ == '__main__': nn_utils.setup_seed(3) model = Yolo(20, "E:/杜老师课程/100_du/02.22/yolov5-2.0/models/yolov5m.yaml") input = torch.zeros((1, 3, 640, 640)) y = model(input) print(len(y)) print(y[0].shape) print(y[1].shape) print(y[2].shape) # print(y[0][0, 0, 0, 0].item()) # model = SPP(3, 2) # input = torch.zeros((1, 3, 640, 640)) # y = model(input) # print(y.shape)