往期回顾:YOLOv5源码解读1.0-目录_汉卿HanQ的博客-CSDN博客
学习了yolo.py网络模型后,今天学习common.py,common.py存放这YOLOv5网络搭建的通用模块,如果修改某一块,就要修改这个文件中对应的模块
目录
1.导入python包
# ----------------------------------1.导入python包----------------------------------
import json # 用于json和Python数据之间的相互转换
import math # 数学函数模块
import platform # 获取操作系统的信息
import warnings # 警告程序员关于语言或库功能的变化的方法
from copy import copy # 数据拷贝模块 分浅拷贝和深拷贝
from pathlib import Path # Path将str转换为Path对象 使字符串路径易于操作的模块
import cv2 # 调用OpenCV的cv库
import numpy as np # numpy数组操作模块
import pandas as pd # panda数组操作模块
import requests # Python的HTTP客户端库
import torch # pytorch深度学习框架
import torch.nn as nn # 专门为神经网络设计的模块化接口
from PIL import Image # 图像基础操作模块
from torch.cuda import amp # 混合精度训练模块
2.加载自定义模块
# ----------------------------------2.加载自定义模块----------------------------------
from utils.datasets import exif_transpose, letterbox # 加载数据集的函数
from utils.general import (LOGGER, check_requirements, check_suffix, colorstr, increment_path, make_divisible,
non_max_suppression, scale_coords, xywh2xyxy, xyxy2xywh) # 定义了一些常用的工具函数
from utils.plots import Annotator, colors, save_one_box # 定义了Annotator类,可以在图像上绘制矩形框和标注信息
from utils.torch_utils import time_sync # 定义了一些与PyTorch有关的工具函数
3.填充padautopad
# ----------------------------------3.填充padautopad----------------------------------
"""
很具输入的卷积核计算需要padding多少才能把tensor补成原理的形状
为same卷积或same池化自动扩充
k 卷积核的kernel_size
p 计算的需要pad值
"""
def autopad(k, p=None): # kernel, padding
# Pad to 'same'
if p is None: # 如果k是int 进行 k//2 否则 x//2
p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad
return p
4.Conv
# ----------------------------------4.Conv----------------------------------
"""
Conv是标准卷积层函数 是整个网络中最核心的模块,由卷积层+BN+激活函数组成
实现了将输入特征经过卷积层 激活函数 归一化层(指定是否使用) 得到输出层
c1 输入的channel
c2 输出的channel
k 卷积核的kernel_seze
s 卷积的stride
p 卷积的padding
act 激活函数类型
"""
class Conv(nn.Module):
# Standard convolution
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 = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) # 卷积层
self.bn = nn.BatchNorm2d(c2) # 归一化层
self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) # 激活函数
def forward(self, x): # 前向计算 网络执行顺序按照forward决定
return self.act(self.bn(self.conv(x))) # conv->bn->act激活
def forward_fuse(self, x): # 前向融合计算
return self.act(self.conv(x)) #卷积->激活
5.深度可分离卷积DWConv
# ----------------------------------5.深度可分离卷积DWConv----------------------------------
"""
将通道数按输入输出的最大公约数进行切分,在不同的通道图层上进行特征学习深度分离卷积层
分组数量=输入通道数量 每个通道作为一个小组分布进行卷积,结果连结作为输出
"""
class DWConv(Conv):
# Depth-wise convolution class
def __init__(self, c1, c2, k=1, s=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
6.注意力层TransformerLayer
# ----------------------------------6.注意力层TransformerLayer----------------------------------
"""
单个Encoder部分 但移除了两个Norm部分
"""
class TransformerLayer(nn.Module):
# Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance)
def __init__(self, c, num_heads):
super().__init__()
# 初始化 query key value
self.q = nn.Linear(c, c, bias=False)
self.k = nn.Linear(c, c, bias=False)
self.v = nn.Linear(c, c, bias=False)
# 输出0 attn_output 即通过self-attention之后,从每一个词语位置输出的attention和输入的query形状意义
# 1 attn_output_weights 即同attention weights 每个单词和另一个单词之间产生一个weight
self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads)
self.fc1 = nn.Linear(c, c, bias=False)
self.fc2 = nn.Linear(c, c, bias=False)
def forward(self, x):
x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x # 多注意力机制+残差
x = self.fc2(self.fc1(x)) + x # 前馈神经网络+残差
return x
7.注意力模块TransformerBlock
# ----------------------------------7.注意力模块TransformerBlock----------------------------------
class TransformerBlock(nn.Module):
# Vision Transformer https://arxiv.org/abs/2010.11929
def __init__(self, c1, c2, num_heads, num_layers):
super().__init__()
self.conv = None
if c1 != c2:
self.conv = Conv(c1, c2)
self.linear = nn.Linear(c2, c2) # learnable position embedding
self.tr = nn.Sequential(*(TransformerLayer(c2, num_heads) for _ in range(num_layers)))
self.c2 = c2
def forward(self, x):
if self.conv is not None:
x = self.conv(x)
b, _, w, h = x.shape
p = x.flatten(2).permute(2, 0, 1)
return self.tr(p + self.linear(p)).permute(1, 2, 0).reshape(b, self.c2, w, h)
8.瓶颈层Bottleneck
# ----------------------------------8.瓶颈层Bottleneck----------------------------------
"""
标准瓶颈层 由1*1 3*3卷积核和残差快组成
主要作用可以更有效提取特征,即减少了参数量 又优化了计算 保持了原有的精度
先经过1*1降维 然后3*3卷积 最后通过残差链接在一起
c1 第一个卷积的输入channel
c2 第二个卷积的输出channel
shortcut bool是否又shortcut连接
g 从输入通道到输出通道的阻塞链接为1
e e*c2就是 第一个输出channel=第二个输入channel
"""
class Bottleneck(nn.Module):
# Standard bottleneck
def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
self.cv1 = Conv(c1, c_, 1, 1) # 1*1卷积层
self.cv2 = Conv(c_, c2, 3, 1, g=g) # 3*3卷积层
self.add = shortcut and c1 == c2 # 如果shortcut=True 将输入输出相加后再数码处
def forward(self, x):
return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
9.CSP瓶颈层BottleneckCSP
# ----------------------------------9.CSP瓶颈层BottleneckCSP----------------------------------
"""
由几个Bottleneck堆叠+CSP结构组成
CSP结构主要思想是再输入block之前,将输入分为两个部分,一部分通过block计算,另一部分通过一个怠倦极shortcut进行concat
可以加强CNN的学习能力,减少内存消耗,减少计算瓶颈
c1 整个BottleneckCSP的输入channel
c2 整个BottleneckCSP的输出channel
n 有几个Bottleneck
g 从输入通道到输出通道的阻塞链接
e 中间层卷积个数/channel数
torch.cat 再11维度(channel)进行合并
c_ BottlenctCSP结构的中间层的通道数 由e决定
"""
class BottleneckCSP(nn.Module):
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
# 4个1*1卷积层
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
self.cv4 = Conv(2 * c_, c2, 1, 1)
self.bn = nn.BatchNorm2d(2 * c_) # BN层
self.act = nn.SiLU() # 激活函数
# m 堆叠n次Bottleneck操作 *可以把list拆开成一个个独立元素
self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
def forward(self, x):
# y1做BottleneckCSP上分支操作 先做一次cv1 再做cv3 即 x->conv->n*Bottleneck->conv->y1
y1 = self.cv3(self.m(self.cv1(x)))
# y2做BottleneckCSP下分支操作
y2 = self.cv2(x)
# y1与y2拼接 接着进入BN层归一化 然后act激活 最后返回cv4
return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
10.简化的CSP瓶颈层C3
# ----------------------------------10.简化的CSP瓶颈层C3----------------------------------
class C3(nn.Module):
# CSP Bottleneck with 3 convolutions
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
# 三个 1*1卷积核
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = Conv(c1, c_, 1, 1)
self.cv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2)
self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
# self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
def forward(self, x):
# 将第一个卷积层与第二个卷积层的结果拼接到一起
return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))
11.自注意力模块的C3TR
# ----------------------------------11.自注意力模块的C3TR----------------------------------
"""
继承C3模块,将n个Bottleneck更换为1个TransformerBlock
"""
class C3TR(C3):
# C3 module with TransformerBlock()
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
super().__init__(c1, c2, n, shortcut, g, e)
c_ = int(c2 * e)
self.m = TransformerBlock(c_, c_, 4, n)
12.SPP的C3SPP
# ----------------------------------12.SPP的C3SPP----------------------------------
"""
继承C3模块,将n个Bottleneck更换为1个SPP
"""
class C3SPP(C3):
# C3 module with SPP()
def __init__(self, c1, c2, k=(5, 9, 13), n=1, shortcut=True, g=1, e=0.5):
super().__init__(c1, c2, n, shortcut, g, e)
c_ = int(c2 * e)
self.m = SPP(c_, c_, k)
13.GhostBottleneck
# ----------------------------------13.GhostBottleneck----------------------------------
"""
继承C3模块,将n个Bottleneck更换为ChostBottleneck
"""
class C3Ghost(C3):
# C3 module with GhostBottleneck()
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
super().__init__(c1, c2, n, shortcut, g, e)
c_ = int(c2 * e) # hidden channels
self.m = nn.Sequential(*(GhostBottleneck(c_, c_) for _ in range(n)))
14.空间金字塔池化模块SPP
# ----------------------------------14.空间金字塔池化模块SPP----------------------------------
"""
空间金字塔池化,用于融合多尺度特征
c1 SPP输入channel
c2 SPP输出channel
k 保存着三个maxpoll的卷积核大小 5 9 13
"""
class SPP(nn.Module):
# Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729
def __init__(self, c1, c2, k=(5, 9, 13)):
super().__init__()
c_ = c1 // 2 # hidden channels
self.cv1 = Conv(c1, c_, 1, 1) # 1*1卷积核
self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) # +1因为由len+1个输入
# m先进行最大池化操作,然后通过nn.ModuleLost进行构造一个模块,再构造时对每一个k都要进行最大池化
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) # cv1操作
with warnings.catch_warnings():
warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning
# 最每一个m进行最大池化,和没有做池化的每一个输入进行叠加,然后拼接 最后做cv2操作
return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
15.快速版SPPF
# ----------------------------------15.快速版SPPF----------------------------------
"""
SPPF是快速版的SPP
"""
class SPPF(nn.Module):
# Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
def __init__(self, c1, c2, k=5): # equivalent to SPP(k=(5, 9, 13))
super().__init__()
c_ = c1 // 2 # hidden channels
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = Conv(c_ * 4, c2, 1, 1)
self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
def forward(self, x):
x = self.cv1(x)
with warnings.catch_warnings():
warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning
y1 = self.m(x)
y2 = self.m(y1)
return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
16.Focus
# ----------------------------------16.Focus----------------------------------
"""
Focus模块在模型的一开始,把wh整合到c空间
在图片进入到Backbone前,对图片进行切片操作(在一张图片中每隔一个像素就拿一个值 类似临近下采样)生成四张图,四张图互补且不丢失信息
通过上述操作,wh就集中到了channel通道空间,输入通道却扩充4倍 即RGB*4=12
最后将得到的新图片进行卷积操作,最终得到了没有丢失信息的二倍下采样特征图
c1 slice后的channel
c2 Focus最终输出的channel
k 最后卷积核的kernel_size
s 最后卷积核的stride
p 最后卷积的padding
g 输入通道到输出通道的阻塞连接
act 激活函数
"""
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))
17.幻象卷积GhostConv
# ----------------------------------17.幻象卷积GhostConv----------------------------------
"""
轻量化网络卷积模块 其不能增加mAP 但可以大大减少模型计算量
c1 输入的channel值
c2 输出的channel值
k 卷积的kernel_size
s 卷积的stride
....
"""
class GhostConv(nn.Module):
# Ghost Convolution https://github.com/huawei-noah/ghostnet
def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups
super().__init__()
c_ = c2 // 2 # hidden channels
self.cv1 = Conv(c1, c_, k, s, None, g, act) # 少量卷积 一般是一半的计算量
self.cv2 = Conv(c_, c_, 5, 1, None, c_, act) # cheap operaions 使用3*3 5*5卷积核 并且逐个特征图卷积
def forward(self, x):
y = self.cv1(x)
return torch.cat([y, self.cv2(y)], 1)
18.幻象模块的瓶颈层
# ----------------------------------18.幻象模块的瓶颈层----------------------------------
"""
一个可复用模块,类似ResNEt的基本残差,由两个堆叠的Ghost模块组成
第一个Ghost模块由于扩展层,增加了通道数。
第二个Ghost模块减少通道数与shortcut路径品牌,然后使用shortcat连接俩个模块的输入和输出
第二个Ghost模块不适用Relu其他层在其他每一层都使用了批归一化BN和Relu
"""
class GhostBottleneck(nn.Module):
# Ghost Bottleneck https://github.com/huawei-noah/ghostnet
def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride
super().__init__()
c_ = c2 // 2
self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw
DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw
GhostConv(c_, c2, 1, 1, act=False)) # pw-linear
# 先将给一个DWconv 然后进行shotcut操作
self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
def forward(self, x):
return self.conv(x) + self.shortcut(x)
19.收缩模块Contract
# ----------------------------------19.收缩模块Contract----------------------------------
"""
收缩模块,调整张量大小,将特征图中的w h 收缩到通道c中
"""
class Contract(nn.Module):
# Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40)
def __init__(self, gain=2):
super().__init__()
self.gain = gain
def forward(self, x):
b, c, h, w = x.size() # assert (h / s == 0) and (W / s == 0), 'Indivisible gain'
s = self.gain
x = x.view(b, c, h // s, s, w // s, s) # x(1,64,40,2,40,2)
x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40)
return x.view(b, c * s * s, h // s, w // s) # x(1,256,40,40)
20.扩张模块Expand
# ----------------------------------20.扩张模块Expand----------------------------------
"""
Contract逆操作,扩张模块,将特征图像素变大
改变输入特征的shape 将channel的数据扩展到w h
"""
class Expand(nn.Module):
# Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160)
def __init__(self, gain=2):
super().__init__()
self.gain = gain
def forward(self, x):
b, c, h, w = x.size() # assert C / s ** 2 == 0, 'Indivisible gain'
s = self.gain
x = x.view(b, s, s, c // s ** 2, h, w) # x(1,2,2,16,80,80)
x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2)
return x.view(b, c // s ** 2, h * s, w * s) # x(1,16,160,160)
21.拼接模块Concat
# ----------------------------------21.拼接模块Concat----------------------------------
"""
将两个tensor进行拼接 demesion是维度意思
"""
class Concat(nn.Module):
# Concatenate a list of tensors along dimension
def __init__(self, dimension=1):
super().__init__()
self.d = dimension
def forward(self, x):
return torch.cat(x, self.d)
22.后端推理DetectMultiBackend
# ----------------------------------22.后端推理DetectMultiBackend----------------------------------
"""
根据不同后端选择相应的模型类型 然后进行推理
"""
class DetectMultiBackend(nn.Module):
# YOLOv5 MultiBackend class for python inference on various backends
def __init__(self, weights='yolov5s.pt', device=None, dnn=True):
# Usage:
# PyTorch: weights = *.pt
# TorchScript: *.torchscript.pt
# CoreML: *.mlmodel
# TensorFlow: *_saved_model
# TensorFlow: *.pb
# TensorFlow Lite: *.tflite
# ONNX Runtime: *.onnx
# OpenCV DNN: *.onnx with dnn=True
super().__init__()
w = str(weights[0] if isinstance(weights, list) else weights)
suffix, suffixes = Path(w).suffix.lower(), ['.pt', '.onnx', '.tflite', '.pb', '', '.mlmodel']
check_suffix(w, suffixes) # check weights have acceptable suffix
pt, onnx, tflite, pb, saved_model, coreml = (suffix == x for x in suffixes) # backend booleans
jit = pt and 'torchscript' in w.lower()
stride, names = 64, [f'class{i}' for i in range(1000)] # assign defaults
if jit: # TorchScript
LOGGER.info(f'Loading {w} for TorchScript inference...')
extra_files = {'config.txt': ''} # model metadata
model = torch.jit.load(w, _extra_files=extra_files)
if extra_files['config.txt']:
d = json.loads(extra_files['config.txt']) # extra_files dict
stride, names = int(d['stride']), d['names']
elif pt: # PyTorch
from models.experimental import attempt_load # scoped to avoid circular import
model = torch.jit.load(w) if 'torchscript' in w else attempt_load(weights, map_location=device)
stride = int(model.stride.max()) # model stride
names = model.module.names if hasattr(model, 'module') else model.names # get class names
elif coreml: # CoreML *.mlmodel
import coremltools as ct
model = ct.models.MLModel(w)
elif dnn: # ONNX OpenCV DNN
LOGGER.info(f'Loading {w} for ONNX OpenCV DNN inference...')
check_requirements(('opencv-python>=4.5.4',))
net = cv2.dnn.readNetFromONNX(w)
elif onnx: # ONNX Runtime
LOGGER.info(f'Loading {w} for ONNX Runtime inference...')
check_requirements(('onnx', 'onnxruntime-gpu' if torch.has_cuda else 'onnxruntime'))
import onnxruntime
session = onnxruntime.InferenceSession(w, None)
else: # TensorFlow model (TFLite, pb, saved_model)
import tensorflow as tf
if pb: # https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt
def wrap_frozen_graph(gd, inputs, outputs):
x = tf.compat.v1.wrap_function(lambda: tf.compat.v1.import_graph_def(gd, name=""), []) # wrapped
return x.prune(tf.nest.map_structure(x.graph.as_graph_element, inputs),
tf.nest.map_structure(x.graph.as_graph_element, outputs))
LOGGER.info(f'Loading {w} for TensorFlow *.pb inference...')
graph_def = tf.Graph().as_graph_def()
graph_def.ParseFromString(open(w, 'rb').read())
frozen_func = wrap_frozen_graph(gd=graph_def, inputs="x:0", outputs="Identity:0")
elif saved_model:
LOGGER.info(f'Loading {w} for TensorFlow saved_model inference...')
model = tf.keras.models.load_model(w)
elif tflite: # https://www.tensorflow.org/lite/guide/python#install_tensorflow_lite_for_python
if 'edgetpu' in w.lower():
LOGGER.info(f'Loading {w} for TensorFlow Edge TPU inference...')
import tflite_runtime.interpreter as tfli
delegate = {'Linux': 'libedgetpu.so.1', # install https://coral.ai/software/#edgetpu-runtime
'Darwin': 'libedgetpu.1.dylib',
'Windows': 'edgetpu.dll'}[platform.system()]
interpreter = tfli.Interpreter(model_path=w, experimental_delegates=[tfli.load_delegate(delegate)])
else:
LOGGER.info(f'Loading {w} for TensorFlow Lite inference...')
interpreter = tf.lite.Interpreter(model_path=w) # load TFLite model
interpreter.allocate_tensors() # allocate
input_details = interpreter.get_input_details() # inputs
output_details = interpreter.get_output_details() # outputs
self.__dict__.update(locals()) # assign all variables to self
def forward(self, im, augment=False, visualize=False, val=False):
# YOLOv5 MultiBackend inference
b, ch, h, w = im.shape # batch, channel, height, width
if self.pt: # PyTorch
y = self.model(im) if self.jit else self.model(im, augment=augment, visualize=visualize)
return y if val else y[0]
elif self.coreml: # CoreML *.mlmodel
im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3)
im = Image.fromarray((im[0] * 255).astype('uint8'))
# im = im.resize((192, 320), Image.ANTIALIAS)
y = self.model.predict({'image': im}) # coordinates are xywh normalized
box = xywh2xyxy(y['coordinates'] * [[w, h, w, h]]) # xyxy pixels
conf, cls = y['confidence'].max(1), y['confidence'].argmax(1).astype(np.float)
y = np.concatenate((box, conf.reshape(-1, 1), cls.reshape(-1, 1)), 1)
elif self.onnx: # ONNX
im = im.cpu().numpy() # torch to numpy
if self.dnn: # ONNX OpenCV DNN
self.net.setInput(im)
y = self.net.forward()
else: # ONNX Runtime
y = self.session.run([self.session.get_outputs()[0].name], {self.session.get_inputs()[0].name: im})[0]
else: # TensorFlow model (TFLite, pb, saved_model)
im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3)
if self.pb:
y = self.frozen_func(x=self.tf.constant(im)).numpy()
elif self.saved_model:
y = self.model(im, training=False).numpy()
elif self.tflite:
input, output = self.input_details[0], self.output_details[0]
int8 = input['dtype'] == np.uint8 # is TFLite quantized uint8 model
if int8:
scale, zero_point = input['quantization']
im = (im / scale + zero_point).astype(np.uint8) # de-scale
self.interpreter.set_tensor(input['index'], im)
self.interpreter.invoke()
y = self.interpreter.get_tensor(output['index'])
if int8:
scale, zero_point = output['quantization']
y = (y.astype(np.float32) - zero_point) * scale # re-scale
y[..., 0] *= w # x
y[..., 1] *= h # y
y[..., 2] *= w # w
y[..., 3] *= h # h
y = torch.tensor(y)
return (y, []) if val else y
23.模型扩展模块AutoShape
# ----------------------------------23.模型扩展模块AutoShape----------------------------------
"""
给模型封装成包含预处理,推理和NMS
在train中不会被调用,当模型训练结束后,会通过这个模块对图片进行重塑,来方便模型预测
"""
class AutoShape(nn.Module):
# YOLOv5 input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS
conf = 0.25 # NMS confidence threshold
iou = 0.45 # NMS IoU threshold
classes = None # (optional list) filter by class, i.e. = [0, 15, 16] for COCO persons, cats and dogs
multi_label = False # NMS multiple labels per box
max_det = 1000 # maximum number of detections per image
def __init__(self, model):
super().__init__()
self.model = model.eval()
def autoshape(self):
LOGGER.info('AutoShape already enabled, skipping... ') # model already converted to model.autoshape()
return self
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.model[-1] # 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
@torch.no_grad()
def forward(self, imgs, size=640, augment=False, profile=False):
# Inference from various sources. For height=640, width=1280, RGB images example inputs are:
# file: imgs = 'data/images/zidane.jpg' # str or PosixPath
# URI: = 'https://ultralytics.com/images/zidane.jpg'
# OpenCV: = cv2.imread('image.jpg')[:,:,::-1] # HWC BGR to RGB x(640,1280,3)
# PIL: = Image.open('image.jpg') or ImageGrab.grab() # HWC x(640,1280,3)
# numpy: = np.zeros((640,1280,3)) # HWC
# torch: = torch.zeros(16,3,320,640) # BCHW (scaled to size=640, 0-1 values)
# multiple: = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...] # list of images
t = [time_sync()]
p = next(self.model.parameters()) # for device and type
if isinstance(imgs, torch.Tensor): # torch
with amp.autocast(enabled=p.device.type != 'cpu'):
return self.model(imgs.to(p.device).type_as(p), augment, profile) # inference
# Pre-process
n, imgs = (len(imgs), imgs) if isinstance(imgs, list) else (1, [imgs]) # number of images, list of images
shape0, shape1, files = [], [], [] # image and inference shapes, filenames
for i, im in enumerate(imgs):
f = f'image{i}' # filename
if isinstance(im, (str, Path)): # filename or uri
im, f = Image.open(requests.get(im, stream=True).raw if str(im).startswith('http') else im), im
im = np.asarray(exif_transpose(im))
elif isinstance(im, Image.Image): # PIL Image
im, f = np.asarray(exif_transpose(im)), getattr(im, 'filename', f) or f
files.append(Path(f).with_suffix('.jpg').name)
if im.shape[0] < 5: # image in CHW
im = im.transpose((1, 2, 0)) # reverse dataloader .transpose(2, 0, 1)
im = im[..., :3] if im.ndim == 3 else np.tile(im[..., None], 3) # enforce 3ch input
s = im.shape[:2] # HWC
shape0.append(s) # image shape
g = (size / max(s)) # gain
shape1.append([y * g for y in s])
imgs[i] = im if im.data.contiguous else np.ascontiguousarray(im) # update
shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)] # inference shape
x = [letterbox(im, new_shape=shape1, auto=False)[0] for im in imgs] # pad
x = np.stack(x, 0) if n > 1 else x[0][None] # stack
x = np.ascontiguousarray(x.transpose((0, 3, 1, 2))) # BHWC to BCHW
x = torch.from_numpy(x).to(p.device).type_as(p) / 255 # uint8 to fp16/32
t.append(time_sync())
with amp.autocast(enabled=p.device.type != 'cpu'):
# Inference
y = self.model(x, augment, profile)[0] # forward
t.append(time_sync())
# Post-process
y = non_max_suppression(y, self.conf, iou_thres=self.iou, classes=self.classes,
multi_label=self.multi_label, max_det=self.max_det) # NMS
for i in range(n):
scale_coords(shape1, y[i][:, :4], shape0[i])
t.append(time_sync())
return Detections(imgs, y, files, t, self.names, x.shape)
24.推理模块Detections
# ----------------------------------24.推理模块Detections----------------------------------
"""
针对目标检测的封装类,对推理结果进行处理
"""
class Detections:
# YOLOv5 detections class for inference results
def __init__(self, imgs, pred, files, times=None, names=None, shape=None):
super().__init__()
d = pred[0].device # device
gn = [torch.tensor([*(im.shape[i] for i in [1, 0, 1, 0]), 1, 1], device=d) for im in imgs] # normalizations
self.imgs = imgs # list of images as numpy arrays 原图
self.pred = pred # list of tensors pred[0] = (xyxy, conf, cls) 预测值
self.names = names # class names 类名
self.files = files # image filenames 图像文件名
self.xyxy = pred # xyxy pixels 左上角+右下角格式
self.xywh = [xyxy2xywh(x) for x in pred] # xywh pixels 中心点+宽长格式
self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)] # xyxy normalized xyxy标准化
self.xywhn = [x / g for x, g in zip(self.xywh, gn)] # xywh normalized xywh标准化
self.n = len(self.pred) # number of images (batch size)
self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3)) # timestamps (ms)
self.s = shape # inference BCHW shape
def display(self, pprint=False, show=False, save=False, crop=False, render=False, save_dir=Path('')):
crops = []
for i, (im, pred) in enumerate(zip(self.imgs, self.pred)):
s = f'image {i + 1}/{len(self.pred)}: {im.shape[0]}x{im.shape[1]} ' # string
if pred.shape[0]:
for c in pred[:, -1].unique():
n = (pred[:, -1] == c).sum() # detections per class
s += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, " # add to string
if show or save or render or crop:
annotator = Annotator(im, example=str(self.names))
for *box, conf, cls in reversed(pred): # xyxy, confidence, class
label = f'{self.names[int(cls)]} {conf:.2f}'
if crop:
file = save_dir / 'crops' / self.names[int(cls)] / self.files[i] if save else None
crops.append({'box': box, 'conf': conf, 'cls': cls, 'label': label,
'im': save_one_box(box, im, file=file, save=save)})
else: # all others
annotator.box_label(box, label, color=colors(cls))
im = annotator.im
else:
s += '(no detections)'
im = Image.fromarray(im.astype(np.uint8)) if isinstance(im, np.ndarray) else im # from np
if pprint:
LOGGER.info(s.rstrip(', '))
if show:
im.show(self.files[i]) # show
if save:
f = self.files[i]
im.save(save_dir / f) # save
if i == self.n - 1:
LOGGER.info(f"Saved {self.n} image{'s' * (self.n > 1)} to {colorstr('bold', save_dir)}")
if render:
self.imgs[i] = np.asarray(im)
if crop:
if save:
LOGGER.info(f'Saved results to {save_dir}\n')
return crops
def print(self):
self.display(pprint=True) # print results
LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' %
self.t)
def show(self):
self.display(show=True) # show results
def save(self, save_dir='runs/detect/exp'):
save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) # increment save_dir
self.display(save=True, save_dir=save_dir) # save results
def crop(self, save=True, save_dir='runs/detect/exp'):
save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) if save else None
return self.display(crop=True, save=save, save_dir=save_dir) # crop results
def render(self):
self.display(render=True) # render results
return self.imgs
def pandas(self):
# return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0])
new = copy(self) # return copy
ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name' # xyxy columns
cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name' # xywh columns
for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]):
a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)] # update
setattr(new, k, [pd.DataFrame(x, columns=c) for x in a])
return new
def tolist(self):
# return a list of Detections objects, i.e. 'for result in results.tolist():'
x = [Detections([self.imgs[i]], [self.pred[i]], self.names, self.s) for i in range(self.n)]
for d in x:
for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']:
setattr(d, k, getattr(d, k)[0]) # pop out of list
return x
def __len__(self):
return self.n
25.二级分类模块Classify
# ----------------------------------25.二级分类模块Classify----------------------------------
"""
对模型进行二次识别或分类 如车牌识别
"""
class Classify(nn.Module):
# Classification head, i.e. x(b,c1,20,20) to x(b,c2)
def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups
super().__init__()
self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1)
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g) # to x(b,c2,1,1) 自适应平均池化操作
self.flat = nn.Flatten() # 展平
def forward(self, x):
z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list 先自适应平均池化操作 然后拼接
return self.flat(self.conv(z)) # flatten to x(b,c2) 对z进行展品操作
26.整体代码
# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
Common modules
"""
# ----------------------------------1.导入python包----------------------------------
import json # 用于json和Python数据之间的相互转换
import math # 数学函数模块
import platform # 获取操作系统的信息
import warnings # 警告程序员关于语言或库功能的变化的方法
from copy import copy # 数据拷贝模块 分浅拷贝和深拷贝
from pathlib import Path # Path将str转换为Path对象 使字符串路径易于操作的模块
import cv2 # 调用OpenCV的cv库
import numpy as np # numpy数组操作模块
import pandas as pd # panda数组操作模块
import requests # Python的HTTP客户端库
import torch # pytorch深度学习框架
import torch.nn as nn # 专门为神经网络设计的模块化接口
from PIL import Image # 图像基础操作模块
from torch.cuda import amp # 混合精度训练模块
# ----------------------------------2.加载自定义模块----------------------------------
from utils.datasets import exif_transpose, letterbox # 加载数据集的函数
from utils.general import (LOGGER, check_requirements, check_suffix, colorstr, increment_path, make_divisible,
non_max_suppression, scale_coords, xywh2xyxy, xyxy2xywh) # 定义了一些常用的工具函数
from utils.plots import Annotator, colors, save_one_box # 定义了Annotator类,可以在图像上绘制矩形框和标注信息
from utils.torch_utils import time_sync # 定义了一些与PyTorch有关的工具函数
# ----------------------------------3.填充padautopad----------------------------------
"""
很具输入的卷积核计算需要padding多少才能把tensor补成原理的形状
为same卷积或same池化自动扩充
k 卷积核的kernel_size
p 计算的需要pad值
"""
def autopad(k, p=None): # kernel, padding
# Pad to 'same'
if p is None: # 如果k是int 进行 k//2 否则 x//2
p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad
return p
# ----------------------------------4.Conv----------------------------------
"""
Conv是标准卷积层函数 是整个网络中最核心的模块,由卷积层+BN+激活函数组成
实现了将输入特征经过卷积层 激活函数 归一化层(指定是否使用) 得到输出层
c1 输入的channel
c2 输出的channel
k 卷积核的kernel_seze
s 卷积的stride
p 卷积的padding
act 激活函数类型
"""
class Conv(nn.Module):
# Standard convolution
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 = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) # 卷积层
self.bn = nn.BatchNorm2d(c2) # 归一化层
self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) # 激活函数
def forward(self, x): # 前向计算 网络执行顺序按照forward决定
return self.act(self.bn(self.conv(x))) # conv->bn->act激活
def forward_fuse(self, x): # 前向融合计算
return self.act(self.conv(x)) #卷积->激活
# ----------------------------------5.深度可分离卷积DWConv----------------------------------
"""
将通道数按输入输出的最大公约数进行切分,在不同的通道图层上进行特征学习深度分离卷积层
分组数量=输入通道数量 每个通道作为一个小组分布进行卷积,结果连结作为输出
"""
class DWConv(Conv):
# Depth-wise convolution class
def __init__(self, c1, c2, k=1, s=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
# ----------------------------------6.注意力层TransformerLayer----------------------------------
"""
单个Encoder部分 但移除了两个Norm部分
"""
class TransformerLayer(nn.Module):
# Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance)
def __init__(self, c, num_heads):
super().__init__()
# 初始化 query key value
self.q = nn.Linear(c, c, bias=False)
self.k = nn.Linear(c, c, bias=False)
self.v = nn.Linear(c, c, bias=False)
# 输出0 attn_output 即通过self-attention之后,从每一个词语位置输出的attention和输入的query形状意义
# 1 attn_output_weights 即同attention weights 每个单词和另一个单词之间产生一个weight
self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads)
self.fc1 = nn.Linear(c, c, bias=False)
self.fc2 = nn.Linear(c, c, bias=False)
def forward(self, x):
x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x # 多注意力机制+残差
x = self.fc2(self.fc1(x)) + x # 前馈神经网络+残差
return x
# ----------------------------------7.注意力模块TransformerBlock----------------------------------
class TransformerBlock(nn.Module):
# Vision Transformer https://arxiv.org/abs/2010.11929
def __init__(self, c1, c2, num_heads, num_layers):
super().__init__()
self.conv = None
if c1 != c2:
self.conv = Conv(c1, c2)
self.linear = nn.Linear(c2, c2) # learnable position embedding
self.tr = nn.Sequential(*(TransformerLayer(c2, num_heads) for _ in range(num_layers)))
self.c2 = c2
def forward(self, x):
if self.conv is not None:
x = self.conv(x)
b, _, w, h = x.shape
p = x.flatten(2).permute(2, 0, 1)
return self.tr(p + self.linear(p)).permute(1, 2, 0).reshape(b, self.c2, w, h)
# ----------------------------------8.瓶颈层Bottleneck----------------------------------
"""
标准瓶颈层 由1*1 3*3卷积核和残差快组成
主要作用可以更有效提取特征,即减少了参数量 又优化了计算 保持了原有的精度
先经过1*1降维 然后3*3卷积 最后通过残差链接在一起
c1 第一个卷积的输入channel
c2 第二个卷积的输出channel
shortcut bool是否又shortcut连接
g 从输入通道到输出通道的阻塞链接为1
e e*c2就是 第一个输出channel=第二个输入channel
"""
class Bottleneck(nn.Module):
# Standard bottleneck
def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
self.cv1 = Conv(c1, c_, 1, 1) # 1*1卷积层
self.cv2 = Conv(c_, c2, 3, 1, g=g) # 3*3卷积层
self.add = shortcut and c1 == c2 # 如果shortcut=True 将输入输出相加后再数码处
def forward(self, x):
return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
# ----------------------------------9.CSP瓶颈层BottleneckCSP----------------------------------
"""
由几个Bottleneck堆叠+CSP结构组成
CSP结构主要思想是再输入block之前,将输入分为两个部分,一部分通过block计算,另一部分通过一个怠倦极shortcut进行concat
可以加强CNN的学习能力,减少内存消耗,减少计算瓶颈
c1 整个BottleneckCSP的输入channel
c2 整个BottleneckCSP的输出channel
n 有几个Bottleneck
g 从输入通道到输出通道的阻塞链接
e 中间层卷积个数/channel数
torch.cat 再11维度(channel)进行合并
c_ BottlenctCSP结构的中间层的通道数 由e决定
"""
class BottleneckCSP(nn.Module):
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
# 4个1*1卷积层
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
self.cv4 = Conv(2 * c_, c2, 1, 1)
self.bn = nn.BatchNorm2d(2 * c_) # BN层
self.act = nn.SiLU() # 激活函数
# m 堆叠n次Bottleneck操作 *可以把list拆开成一个个独立元素
self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
def forward(self, x):
# y1做BottleneckCSP上分支操作 先做一次cv1 再做cv3 即 x->conv->n*Bottleneck->conv->y1
y1 = self.cv3(self.m(self.cv1(x)))
# y2做BottleneckCSP下分支操作
y2 = self.cv2(x)
# y1与y2拼接 接着进入BN层归一化 然后act激活 最后返回cv4
return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
# ----------------------------------10.简化的CSP瓶颈层C3----------------------------------
class C3(nn.Module):
# CSP Bottleneck with 3 convolutions
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
# 三个 1*1卷积核
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = Conv(c1, c_, 1, 1)
self.cv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2)
self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
# self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
def forward(self, x):
# 将第一个卷积层与第二个卷积层的结果拼接到一起
return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))
# ----------------------------------11.自注意力模块的C3TR----------------------------------
"""
继承C3模块,将n个Bottleneck更换为1个TransformerBlock
"""
class C3TR(C3):
# C3 module with TransformerBlock()
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
super().__init__(c1, c2, n, shortcut, g, e)
c_ = int(c2 * e)
self.m = TransformerBlock(c_, c_, 4, n)
# ----------------------------------12.SPP的C3SPP----------------------------------
"""
继承C3模块,将n个Bottleneck更换为1个SPP
"""
class C3SPP(C3):
# C3 module with SPP()
def __init__(self, c1, c2, k=(5, 9, 13), n=1, shortcut=True, g=1, e=0.5):
super().__init__(c1, c2, n, shortcut, g, e)
c_ = int(c2 * e)
self.m = SPP(c_, c_, k)
# ----------------------------------13.GhostBottleneck----------------------------------
"""
继承C3模块,将n个Bottleneck更换为ChostBottleneck
"""
class C3Ghost(C3):
# C3 module with GhostBottleneck()
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
super().__init__(c1, c2, n, shortcut, g, e)
c_ = int(c2 * e) # hidden channels
self.m = nn.Sequential(*(GhostBottleneck(c_, c_) for _ in range(n)))
# ----------------------------------14.空间金字塔池化模块SPP----------------------------------
"""
空间金字塔池化,用于融合多尺度特征
c1 SPP输入channel
c2 SPP输出channel
k 保存着三个maxpoll的卷积核大小 5 9 13
"""
class SPP(nn.Module):
# Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729
def __init__(self, c1, c2, k=(5, 9, 13)):
super().__init__()
c_ = c1 // 2 # hidden channels
self.cv1 = Conv(c1, c_, 1, 1) # 1*1卷积核
self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) # +1因为由len+1个输入
# m先进行最大池化操作,然后通过nn.ModuleLost进行构造一个模块,再构造时对每一个k都要进行最大池化
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) # cv1操作
with warnings.catch_warnings():
warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning
# 最每一个m进行最大池化,和没有做池化的每一个输入进行叠加,然后拼接 最后做cv2操作
return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
# ----------------------------------15.快速版SPPF----------------------------------
"""
SPPF是快速版的SPP
"""
class SPPF(nn.Module):
# Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
def __init__(self, c1, c2, k=5): # equivalent to SPP(k=(5, 9, 13))
super().__init__()
c_ = c1 // 2 # hidden channels
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = Conv(c_ * 4, c2, 1, 1)
self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
def forward(self, x):
x = self.cv1(x)
with warnings.catch_warnings():
warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning
y1 = self.m(x)
y2 = self.m(y1)
return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
# ----------------------------------16.Focus----------------------------------
"""
Focus模块在模型的一开始,把wh整合到c空间
在图片进入到Backbone前,对图片进行切片操作(在一张图片中每隔一个像素就拿一个值 类似临近下采样)生成四张图,四张图互补且不丢失信息
通过上述操作,wh就集中到了channel通道空间,输入通道却扩充4倍 即RGB*4=12
最后将得到的新图片进行卷积操作,最终得到了没有丢失信息的二倍下采样特征图
c1 slice后的channel
c2 Focus最终输出的channel
k 最后卷积核的kernel_size
s 最后卷积核的stride
p 最后卷积的padding
g 输入通道到输出通道的阻塞连接
act 激活函数
"""
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))
# ----------------------------------17.幻象卷积GhostConv----------------------------------
"""
轻量化网络卷积模块 其不能增加mAP 但可以大大减少模型计算量
c1 输入的channel值
c2 输出的channel值
k 卷积的kernel_size
s 卷积的stride
....
"""
class GhostConv(nn.Module):
# Ghost Convolution https://github.com/huawei-noah/ghostnet
def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups
super().__init__()
c_ = c2 // 2 # hidden channels
self.cv1 = Conv(c1, c_, k, s, None, g, act) # 少量卷积 一般是一半的计算量
self.cv2 = Conv(c_, c_, 5, 1, None, c_, act) # cheap operaions 使用3*3 5*5卷积核 并且逐个特征图卷积
def forward(self, x):
y = self.cv1(x)
return torch.cat([y, self.cv2(y)], 1)
# ----------------------------------18.幻象模块的瓶颈层----------------------------------
"""
一个可复用模块,类似ResNEt的基本残差,由两个堆叠的Ghost模块组成
第一个Ghost模块由于扩展层,增加了通道数。
第二个Ghost模块减少通道数与shortcut路径品牌,然后使用shortcat连接俩个模块的输入和输出
第二个Ghost模块不适用Relu其他层在其他每一层都使用了批归一化BN和Relu
"""
class GhostBottleneck(nn.Module):
# Ghost Bottleneck https://github.com/huawei-noah/ghostnet
def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride
super().__init__()
c_ = c2 // 2
self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw
DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw
GhostConv(c_, c2, 1, 1, act=False)) # pw-linear
# 先将给一个DWconv 然后进行shotcut操作
self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
def forward(self, x):
return self.conv(x) + self.shortcut(x)
# ----------------------------------19.收缩模块Contract----------------------------------
"""
收缩模块,调整张量大小,将特征图中的w h 收缩到通道c中
"""
class Contract(nn.Module):
# Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40)
def __init__(self, gain=2):
super().__init__()
self.gain = gain
def forward(self, x):
b, c, h, w = x.size() # assert (h / s == 0) and (W / s == 0), 'Indivisible gain'
s = self.gain
x = x.view(b, c, h // s, s, w // s, s) # x(1,64,40,2,40,2)
x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40)
return x.view(b, c * s * s, h // s, w // s) # x(1,256,40,40)
# ----------------------------------20.扩张模块Expand----------------------------------
"""
Contract逆操作,扩张模块,将特征图像素变大
改变输入特征的shape 将channel的数据扩展到w h
"""
class Expand(nn.Module):
# Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160)
def __init__(self, gain=2):
super().__init__()
self.gain = gain
def forward(self, x):
b, c, h, w = x.size() # assert C / s ** 2 == 0, 'Indivisible gain'
s = self.gain
x = x.view(b, s, s, c // s ** 2, h, w) # x(1,2,2,16,80,80)
x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2)
return x.view(b, c // s ** 2, h * s, w * s) # x(1,16,160,160)
# ----------------------------------21.拼接模块Concat----------------------------------
"""
将两个tensor进行拼接 demesion是维度意思
"""
class Concat(nn.Module):
# Concatenate a list of tensors along dimension
def __init__(self, dimension=1):
super().__init__()
self.d = dimension
def forward(self, x):
return torch.cat(x, self.d)
# ----------------------------------22.后端推理DetectMultiBackend----------------------------------
"""
根据不同后端选择相应的模型类型 然后进行推理
"""
class DetectMultiBackend(nn.Module):
# YOLOv5 MultiBackend class for python inference on various backends
def __init__(self, weights='yolov5s.pt', device=None, dnn=True):
# Usage:
# PyTorch: weights = *.pt
# TorchScript: *.torchscript.pt
# CoreML: *.mlmodel
# TensorFlow: *_saved_model
# TensorFlow: *.pb
# TensorFlow Lite: *.tflite
# ONNX Runtime: *.onnx
# OpenCV DNN: *.onnx with dnn=True
super().__init__()
w = str(weights[0] if isinstance(weights, list) else weights)
suffix, suffixes = Path(w).suffix.lower(), ['.pt', '.onnx', '.tflite', '.pb', '', '.mlmodel']
check_suffix(w, suffixes) # check weights have acceptable suffix
pt, onnx, tflite, pb, saved_model, coreml = (suffix == x for x in suffixes) # backend booleans
jit = pt and 'torchscript' in w.lower()
stride, names = 64, [f'class{i}' for i in range(1000)] # assign defaults
if jit: # TorchScript
LOGGER.info(f'Loading {w} for TorchScript inference...')
extra_files = {'config.txt': ''} # model metadata
model = torch.jit.load(w, _extra_files=extra_files)
if extra_files['config.txt']:
d = json.loads(extra_files['config.txt']) # extra_files dict
stride, names = int(d['stride']), d['names']
elif pt: # PyTorch
from models.experimental import attempt_load # scoped to avoid circular import
model = torch.jit.load(w) if 'torchscript' in w else attempt_load(weights, map_location=device)
stride = int(model.stride.max()) # model stride
names = model.module.names if hasattr(model, 'module') else model.names # get class names
elif coreml: # CoreML *.mlmodel
import coremltools as ct
model = ct.models.MLModel(w)
elif dnn: # ONNX OpenCV DNN
LOGGER.info(f'Loading {w} for ONNX OpenCV DNN inference...')
check_requirements(('opencv-python>=4.5.4',))
net = cv2.dnn.readNetFromONNX(w)
elif onnx: # ONNX Runtime
LOGGER.info(f'Loading {w} for ONNX Runtime inference...')
check_requirements(('onnx', 'onnxruntime-gpu' if torch.has_cuda else 'onnxruntime'))
import onnxruntime
session = onnxruntime.InferenceSession(w, None)
else: # TensorFlow model (TFLite, pb, saved_model)
import tensorflow as tf
if pb: # https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt
def wrap_frozen_graph(gd, inputs, outputs):
x = tf.compat.v1.wrap_function(lambda: tf.compat.v1.import_graph_def(gd, name=""), []) # wrapped
return x.prune(tf.nest.map_structure(x.graph.as_graph_element, inputs),
tf.nest.map_structure(x.graph.as_graph_element, outputs))
LOGGER.info(f'Loading {w} for TensorFlow *.pb inference...')
graph_def = tf.Graph().as_graph_def()
graph_def.ParseFromString(open(w, 'rb').read())
frozen_func = wrap_frozen_graph(gd=graph_def, inputs="x:0", outputs="Identity:0")
elif saved_model:
LOGGER.info(f'Loading {w} for TensorFlow saved_model inference...')
model = tf.keras.models.load_model(w)
elif tflite: # https://www.tensorflow.org/lite/guide/python#install_tensorflow_lite_for_python
if 'edgetpu' in w.lower():
LOGGER.info(f'Loading {w} for TensorFlow Edge TPU inference...')
import tflite_runtime.interpreter as tfli
delegate = {'Linux': 'libedgetpu.so.1', # install https://coral.ai/software/#edgetpu-runtime
'Darwin': 'libedgetpu.1.dylib',
'Windows': 'edgetpu.dll'}[platform.system()]
interpreter = tfli.Interpreter(model_path=w, experimental_delegates=[tfli.load_delegate(delegate)])
else:
LOGGER.info(f'Loading {w} for TensorFlow Lite inference...')
interpreter = tf.lite.Interpreter(model_path=w) # load TFLite model
interpreter.allocate_tensors() # allocate
input_details = interpreter.get_input_details() # inputs
output_details = interpreter.get_output_details() # outputs
self.__dict__.update(locals()) # assign all variables to self
def forward(self, im, augment=False, visualize=False, val=False):
# YOLOv5 MultiBackend inference
b, ch, h, w = im.shape # batch, channel, height, width
if self.pt: # PyTorch
y = self.model(im) if self.jit else self.model(im, augment=augment, visualize=visualize)
return y if val else y[0]
elif self.coreml: # CoreML *.mlmodel
im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3)
im = Image.fromarray((im[0] * 255).astype('uint8'))
# im = im.resize((192, 320), Image.ANTIALIAS)
y = self.model.predict({'image': im}) # coordinates are xywh normalized
box = xywh2xyxy(y['coordinates'] * [[w, h, w, h]]) # xyxy pixels
conf, cls = y['confidence'].max(1), y['confidence'].argmax(1).astype(np.float)
y = np.concatenate((box, conf.reshape(-1, 1), cls.reshape(-1, 1)), 1)
elif self.onnx: # ONNX
im = im.cpu().numpy() # torch to numpy
if self.dnn: # ONNX OpenCV DNN
self.net.setInput(im)
y = self.net.forward()
else: # ONNX Runtime
y = self.session.run([self.session.get_outputs()[0].name], {self.session.get_inputs()[0].name: im})[0]
else: # TensorFlow model (TFLite, pb, saved_model)
im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3)
if self.pb:
y = self.frozen_func(x=self.tf.constant(im)).numpy()
elif self.saved_model:
y = self.model(im, training=False).numpy()
elif self.tflite:
input, output = self.input_details[0], self.output_details[0]
int8 = input['dtype'] == np.uint8 # is TFLite quantized uint8 model
if int8:
scale, zero_point = input['quantization']
im = (im / scale + zero_point).astype(np.uint8) # de-scale
self.interpreter.set_tensor(input['index'], im)
self.interpreter.invoke()
y = self.interpreter.get_tensor(output['index'])
if int8:
scale, zero_point = output['quantization']
y = (y.astype(np.float32) - zero_point) * scale # re-scale
y[..., 0] *= w # x
y[..., 1] *= h # y
y[..., 2] *= w # w
y[..., 3] *= h # h
y = torch.tensor(y)
return (y, []) if val else y
# ----------------------------------23.模型扩展模块AutoShape----------------------------------
"""
给模型封装成包含预处理,推理和NMS
在train中不会被调用,当模型训练结束后,会通过这个模块对图片进行重塑,来方便模型预测
"""
class AutoShape(nn.Module):
# YOLOv5 input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS
conf = 0.25 # NMS confidence threshold
iou = 0.45 # NMS IoU threshold
classes = None # (optional list) filter by class, i.e. = [0, 15, 16] for COCO persons, cats and dogs
multi_label = False # NMS multiple labels per box
max_det = 1000 # maximum number of detections per image
def __init__(self, model):
super().__init__()
self.model = model.eval()
def autoshape(self):
LOGGER.info('AutoShape already enabled, skipping... ') # model already converted to model.autoshape()
return self
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.model[-1] # 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
@torch.no_grad()
def forward(self, imgs, size=640, augment=False, profile=False):
# Inference from various sources. For height=640, width=1280, RGB images example inputs are:
# file: imgs = 'data/images/zidane.jpg' # str or PosixPath
# URI: = 'https://ultralytics.com/images/zidane.jpg'
# OpenCV: = cv2.imread('image.jpg')[:,:,::-1] # HWC BGR to RGB x(640,1280,3)
# PIL: = Image.open('image.jpg') or ImageGrab.grab() # HWC x(640,1280,3)
# numpy: = np.zeros((640,1280,3)) # HWC
# torch: = torch.zeros(16,3,320,640) # BCHW (scaled to size=640, 0-1 values)
# multiple: = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...] # list of images
t = [time_sync()]
p = next(self.model.parameters()) # for device and type
if isinstance(imgs, torch.Tensor): # torch
with amp.autocast(enabled=p.device.type != 'cpu'):
return self.model(imgs.to(p.device).type_as(p), augment, profile) # inference
# Pre-process
n, imgs = (len(imgs), imgs) if isinstance(imgs, list) else (1, [imgs]) # number of images, list of images
shape0, shape1, files = [], [], [] # image and inference shapes, filenames
for i, im in enumerate(imgs):
f = f'image{i}' # filename
if isinstance(im, (str, Path)): # filename or uri
im, f = Image.open(requests.get(im, stream=True).raw if str(im).startswith('http') else im), im
im = np.asarray(exif_transpose(im))
elif isinstance(im, Image.Image): # PIL Image
im, f = np.asarray(exif_transpose(im)), getattr(im, 'filename', f) or f
files.append(Path(f).with_suffix('.jpg').name)
if im.shape[0] < 5: # image in CHW
im = im.transpose((1, 2, 0)) # reverse dataloader .transpose(2, 0, 1)
im = im[..., :3] if im.ndim == 3 else np.tile(im[..., None], 3) # enforce 3ch input
s = im.shape[:2] # HWC
shape0.append(s) # image shape
g = (size / max(s)) # gain
shape1.append([y * g for y in s])
imgs[i] = im if im.data.contiguous else np.ascontiguousarray(im) # update
shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)] # inference shape
x = [letterbox(im, new_shape=shape1, auto=False)[0] for im in imgs] # pad
x = np.stack(x, 0) if n > 1 else x[0][None] # stack
x = np.ascontiguousarray(x.transpose((0, 3, 1, 2))) # BHWC to BCHW
x = torch.from_numpy(x).to(p.device).type_as(p) / 255 # uint8 to fp16/32
t.append(time_sync())
with amp.autocast(enabled=p.device.type != 'cpu'):
# Inference
y = self.model(x, augment, profile)[0] # forward
t.append(time_sync())
# Post-process
y = non_max_suppression(y, self.conf, iou_thres=self.iou, classes=self.classes,
multi_label=self.multi_label, max_det=self.max_det) # NMS
for i in range(n):
scale_coords(shape1, y[i][:, :4], shape0[i])
t.append(time_sync())
return Detections(imgs, y, files, t, self.names, x.shape)
# ----------------------------------24.推理模块Detections----------------------------------
"""
针对目标检测的封装类,对推理结果进行处理
"""
class Detections:
# YOLOv5 detections class for inference results
def __init__(self, imgs, pred, files, times=None, names=None, shape=None):
super().__init__()
d = pred[0].device # device
gn = [torch.tensor([*(im.shape[i] for i in [1, 0, 1, 0]), 1, 1], device=d) for im in imgs] # normalizations
self.imgs = imgs # list of images as numpy arrays 原图
self.pred = pred # list of tensors pred[0] = (xyxy, conf, cls) 预测值
self.names = names # class names 类名
self.files = files # image filenames 图像文件名
self.xyxy = pred # xyxy pixels 左上角+右下角格式
self.xywh = [xyxy2xywh(x) for x in pred] # xywh pixels 中心点+宽长格式
self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)] # xyxy normalized xyxy标准化
self.xywhn = [x / g for x, g in zip(self.xywh, gn)] # xywh normalized xywh标准化
self.n = len(self.pred) # number of images (batch size)
self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3)) # timestamps (ms)
self.s = shape # inference BCHW shape
def display(self, pprint=False, show=False, save=False, crop=False, render=False, save_dir=Path('')):
crops = []
for i, (im, pred) in enumerate(zip(self.imgs, self.pred)):
s = f'image {i + 1}/{len(self.pred)}: {im.shape[0]}x{im.shape[1]} ' # string
if pred.shape[0]:
for c in pred[:, -1].unique():
n = (pred[:, -1] == c).sum() # detections per class
s += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, " # add to string
if show or save or render or crop:
annotator = Annotator(im, example=str(self.names))
for *box, conf, cls in reversed(pred): # xyxy, confidence, class
label = f'{self.names[int(cls)]} {conf:.2f}'
if crop:
file = save_dir / 'crops' / self.names[int(cls)] / self.files[i] if save else None
crops.append({'box': box, 'conf': conf, 'cls': cls, 'label': label,
'im': save_one_box(box, im, file=file, save=save)})
else: # all others
annotator.box_label(box, label, color=colors(cls))
im = annotator.im
else:
s += '(no detections)'
im = Image.fromarray(im.astype(np.uint8)) if isinstance(im, np.ndarray) else im # from np
if pprint:
LOGGER.info(s.rstrip(', '))
if show:
im.show(self.files[i]) # show
if save:
f = self.files[i]
im.save(save_dir / f) # save
if i == self.n - 1:
LOGGER.info(f"Saved {self.n} image{'s' * (self.n > 1)} to {colorstr('bold', save_dir)}")
if render:
self.imgs[i] = np.asarray(im)
if crop:
if save:
LOGGER.info(f'Saved results to {save_dir}\n')
return crops
def print(self):
self.display(pprint=True) # print results
LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' %
self.t)
def show(self):
self.display(show=True) # show results
def save(self, save_dir='runs/detect/exp'):
save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) # increment save_dir
self.display(save=True, save_dir=save_dir) # save results
def crop(self, save=True, save_dir='runs/detect/exp'):
save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) if save else None
return self.display(crop=True, save=save, save_dir=save_dir) # crop results
def render(self):
self.display(render=True) # render results
return self.imgs
def pandas(self):
# return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0])
new = copy(self) # return copy
ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name' # xyxy columns
cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name' # xywh columns
for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]):
a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)] # update
setattr(new, k, [pd.DataFrame(x, columns=c) for x in a])
return new
def tolist(self):
# return a list of Detections objects, i.e. 'for result in results.tolist():'
x = [Detections([self.imgs[i]], [self.pred[i]], self.names, self.s) for i in range(self.n)]
for d in x:
for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']:
setattr(d, k, getattr(d, k)[0]) # pop out of list
return x
def __len__(self):
return self.n
# ----------------------------------25.二级分类模块Classify----------------------------------
"""
对模型进行二次识别或分类 如车牌识别
"""
class Classify(nn.Module):
# Classification head, i.e. x(b,c1,20,20) to x(b,c2)
def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups
super().__init__()
self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1)
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g) # to x(b,c2,1,1) 自适应平均池化操作
self.flat = nn.Flatten() # 展平
def forward(self, x):
z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list 先自适应平均池化操作 然后拼接
return self.flat(self.conv(z)) # flatten to x(b,c2) 对z进行展品操作