yolov4-keras每个细节详细展示
本篇文章以以下标题为索引，进行组文，如有错误，望不吝指正！
参考代码链接：https://github.com/bubbliiiing/yolov4-keras
文章目录
- 1. yolov4的数据预处理
- 2. yolov4的datasets生成的具体方法
- 3. yolov4的损失函数
- 4. yolov4的训练设置
1. yolov4的数据预处理
首先明确一下输入yolov4网络的图像内容及形状：
"""
time: 2022/04/14
author: cong
theme: image read show
"""
from PIL import Image
import cv2
import matplotlib.pyplot as plt
import numpy as np
# yolov4读入图像的两种方式
# opencv
image0 = cv2.imread('1_10.jpg') # BGR通道
image1 = cv2.cvtColor(image0, cv2.COLOR_BGR2RGB)
image2 = Image.fromarray(np.uint8(image1)) # 转变成Image
image3 = image2.resize((416, 416), Image.BICUBIC)
image_data4 = np.expand_dims(np.array(image3, dtype='float32')/255.0, 0)
# Image
imagea = Image.open('1_10.jpg')
imageb = imagea.resize((416, 416), Image.BICUBIC)
image_datac = np.expand_dims(np.array(imageb, dtype='float32')/255.0, 0)以上只是先整理了yolov4读入图像的两种方式：opencv和Image。
以下代码是get_random_data函数的具体实现方式，过程涉及到：
- 图像随机缩放，paste到416*416的灰度图上
- 图像镜像
- 色域变换
"""
time : 2022/03/29
author : cong
theme: yolov4_get_random_data
"""
from PIL import Image
import numpy as np
import cv2
# 将图像转换成RGB图像，防止灰度图在预测时报错。代码仅仅支持RGB图像的预测，所有其它类型的图像都会转化成RGB
def cvtColor(image):
if len(np.shape(image)) == 3 and np.shape(image)[2] == 3:
return image
else:
image = image.convert('RGB')
return image
# 生成随机数
def rand(a=0, b=1):
return np.random.rand() * (b - a) + a
# ================================================================================ #
# train_lines
train_annotation_path = '2007_train.txt'
with open(train_annotation_path) as f:
train_lines = f.readlines()
input_shape = [416, 416]
max_boxes = 100
random = True
# train_lines[0]: '/Users/cong/yolov4-keras-master/VOCdevkit/VOC2007/JPEGImages/007826.jpg 80,217,320,273,10 197,193,257,326,8 258,180,312,314,8 10,195,93,358,8 82,252,243,372,8\n'
annotation_line = train_lines[0]
line = annotation_line.split() # 默认切分 空格，换行，制表符,返回列表。# ['/Users/cong/yolov4-keras-master/VOCdevkit/VOC2007/JPEGImages/007826.jpg', '80,217,320,273,10','197,193,257,326,8', '258,180,312,314,8', '10,195,93,358,8', '82,252,243,372,8']
# 读取图像并转换成RGB图像
# line[0]:'/Users/cong/yolov4-keras-master/VOCdevkit/VOC2007/JPEGImages/2009_002295.jpg'
image = Image.open(line[0])
# cvtColor函数,,比如说截图。
if len(np.shape(image)) == 3 and np.shape(image)[2] == 3:
image = image
else:
image = image.convert('RGB')
# 获得图像的高宽与目标高宽
iw, ih = image.size
h, w = input_shape
# 获得预测框, int向下取整, box.split(','): ['80', '217', '320', '273', '10']
box = np.array([np.array(list(map(int, box.split(',')))) for box in line[1:]])
# array([[80, 217, 320, 273, 10],
# [197, 193, 257, 326, 8],
# [258, 180, 312, 314, 8],
# [10, 195, 93, 358, 8],
# [82, 252, 243, 372, 8]])
# not random ==================================
if not random:
scale = min(w / iw, h / ih) # w,h:416,,iw:375, ih:500
# nw,nh中 ，最多和w,h一样大
nw = int(iw * scale)
nh = int(ih * scale)
dx = (w - nw) // 2
dy = (h - nh) // 2
# ---------------------------------#
# 将图像多余的部分加上灰条
# ---------------------------------#
image = image.resize((nw, nh), Image.BICUBIC)
image.show()
new_image = Image.new('RGB', (w, h), (128, 128, 128))
new_image.show()
new_image.paste(image, (dx, dy))
new_image.show()
image_data = np.array(new_image, np.float32)
# 对真实框进行调整
box_data = np.zeros((max_boxes, 5))
if len(box) > 0:
np.random.shuffle(box)
box[:, [0, 2]] = box[:, [0, 2]] * nw / iw + dx
box[:, [1, 3]] = box[:, [1, 3]] * nh / ih + dy
box[:, 0:2][box[:, 0:2] < 0] = 0
box[:, 2][box[:, 2] > w] = w
box[:, 3][box[:, 3] > h] = h
box_w = box[:, 2] - box[:, 0]
box_h = box[:, 3] - box[:, 1]
box = box[np.logical_and(box_w > 1, box_h > 1)]# 从box中取出符合条件的。logical_and逻辑与。
if len(box) > max_boxes:
box = box[:max_boxes]
box_data[:len(box)] = box
# random ===========================================
# 多余部分加上灰度条//翻转//色域扭曲
if random:
jitter = .3
hue = .1
sat = 1.5
val = 1.5
new_ar = w / h * rand(1 - jitter, 1 + jitter) / rand(1 - jitter, 1 + jitter)
scale = rand(.25, 2)
if new_ar < 1:
nh = int(scale * h)
nw = int(nh * new_ar)
else:
nw = int(scale * w)
nh = int(nw / new_ar)
image = image.resize((nw, nh), Image.BICUBIC)
# ------------------------------------------#
# 将图像多余的部分加上灰条
# ------------------------------------------#
dx = int(rand(0, w - nw))
dy = int(rand(0, h - nh))
new_image = Image.new('RGB', (w, h), (128, 128, 128))
# new_image.show()
new_image.paste(image, (dx, dy))
image = new_image
# image.show()
# ------------------------------------------#
# 翻转图像
# ------------------------------------------#
flip = rand() < .5
if flip: image = image.transpose(Image.FLIP_LEFT_RIGHT) # 左右翻转
# ------------------------------------------#
# 色域扭曲 RGB==>HSV==>RGB
# ------------------------------------------#
hue = rand(-hue, hue)
sat = rand(1, sat) if rand() < .5 else 1 / rand(1, sat)
val = rand(1, val) if rand() < .5 else 1 / rand(1, val)
x = cv2.cvtColor(np.array(image, np.float32) / 255, cv2.COLOR_RGB2HSV)
x[..., 0] += hue * 360
x[..., 0][x[..., 0] > 1] -= 1
x[..., 0][x[..., 0] < 0] += 1
x[..., 1] *= sat
x[..., 2] *= val
x[x[:, :, 0] > 360, 0] = 360
x[:, :, 1:][x[:, :, 1:] > 1] = 1
x[x < 0] = 0
image_data = cv2.cvtColor(x, cv2.COLOR_HSV2RGB) * 255 # numpy array, 0 to 1
image_uint8 = image_data.astype(np.uint8)
image = Image.fromarray(image_uint8)
image.show()
cv2.imshow('image:', cv2.cvtColor(image_data, cv2.COLOR_RGB2BGR))
cv2.waitKey(0)
# ---------------------------------#
# 对真实框进行调整
# ---------------------------------#
box_data = np.zeros((max_boxes, 5))
if len(box) > 0:
np.random.shuffle(box)
box[:, [0, 2]] = box[:, [0, 2]] * nw / iw + dx
box[:, [1, 3]] = box[:, [1, 3]] * nh / ih + dy
if flip: box[:, [0, 2]] = w - box[:, [2, 0]]
box[:, 0:2][box[:, 0:2] < 0] = 0
box[:, 2][box[:, 2] > w] = w
box[:, 3][box[:, 3] > h] = h
box_w = box[:, 2] - box[:, 0]
box_h = box[:, 3] - box[:, 1]
box = box[np.logical_and(box_w > 1, box_h > 1)] # discard invalid box
if len(box) > max_boxes: box = box[:max_boxes]
box_data[:len(box)] = box2. yolov4的datasets生成的具体方法
以下是YoloDatasets类的详细过程，涉及到训练过程中，怎样生成batch数据：
"""
time: 2022/03/28
author: cong
"""
import numpy as np
import math
from PIL import Image
from random import shuffle, sample
import cv2
import keras
import matplotlib
matplotlib.use('Agg')
# Agg 渲染器是非交互式的后端，没有GUI界面，所以不显示图片，它是用来生成图像文件。
# Qt5Agg 是意思是Agg渲染器输出到Qt5绘图面板，它是交互式的后端，拥有在屏幕上展示的能力3。
def sigmoid(x):
s = 1 / (1 + np.exp(-x))
return s
# 获得目标类别
def get_classes(classes_path):
with open(classes_path, encoding='utf-8') as f:
class_names = f.readlines()
class_names = [c.strip() for c in class_names]
return class_names, len(class_names)
# 获得先验框
def get_anchors(anchors_path):
'''loads the anchors from a file'''
with open(anchors_path, encoding='utf-8') as f:
anchors = f.readline()
anchors = [float(x) for x in anchors.split(',')]
anchors = np.array(anchors).reshape(-1, 2)
return anchors, len(anchors)
# 将图像转换成RGB图像，防止灰度图在预测时报错。代码仅仅支持RGB图像的预测，所有其它类型的图像都会转化成RGB
def cvtColor(image):
if len(np.shape(image)) == 3 and np.shape(image)[2] == 3:
return image
else:
image = image.convert('RGB')
return image
# 图像归一化
def preprocess_input(image):
image /= 255.0
return image
# 1、__init__方法的使用和功能：
# 1、用来构造初始化函数，用来给类的实例进行初始化属性，所以不需要返回值。
# 2、在创建实例时系统自动调用
# 3、自定义类如果不定义的话，默认调用父类的，同理继承也是，子类若无，调用父类，若有，调用自己的.
#Sequence类必须重载三个私有方法__init__、__len__和__getitem__，主要是__getitem__。__init__是构造方法，用于初始化数据的，
# 只要能让样本数据集通过形参顺利传进来就行了。__len__基本上不用改写，用于计算样本数据长度。__getitem__用于生成批量数据，
# 喂给神经网络模型训练用，其输出格式是元组。元组里面有两个元素，每个元素各是一个列表，第一个元素是batch data构成的列表，
# 第二个元素是label构成的列表。在第一个列表中每个元素是一个batch，每个batch里面才是图像张量，所有的batch串成一个列表。第二个列表比较普通，
# 每个元素是个实数，表示标签。这点跟生成器不一样，生成器是在执行时通过yield关键字把每个batch数据喂给模型，一次喂一个batch。__getitem__相当于生成器的作用，
# 如同ImageDataGenerator，但注意编写方法时返回数据不要用yield，而要用return，像一个普通函数一样。至于后台怎么迭代调用这个生成器，
# 这是keras后台程序处理好的事，我们不用担心。我这里程序假定样本数据已经转成pickle的形式，是字典构成的列表，字典的键值分别是feature和label，其中feature是图像张量。
class YoloDatasets(keras.utils.Sequence):
def __init__(self, annotation_lines, input_shape, anchors, batch_size, num_classes, anchors_mask, mosaic, train):
self.annotation_lines = annotation_lines
self.length = len(self.annotation_lines)
self.input_shape = input_shape
self.anchors = anchors
self.batch_size = batch_size
self.num_classes = num_classes
self.anchors_mask = anchors_mask
self.mosaic = mosaic
self.train = train
# 覆盖keras.utils.Sequence中的__len__函数, 返回类实例的长度
def __len__(self):
return math.ceil(len(self.annotation_lines) / float(self.batch_size)) # ceil:大于结果的整数
# 覆盖keras.utils.Sequence中的__getitem__函数，这个是返回一个可迭代的对象，如果一个类实现了这个魔法函数，那么这个类就是可迭代对象
# 在__getitem__中没有结束条件
def __getitem__(self, index):
image_data = []
box_data = []
for i in range(index * self.batch_size, (index + 1) * self.batch_size):
i = i % self.length # voc:17416
# ---------------------------------------------------#
# 训练时进行数据的随机增强,验证时不进行数据的随机增强
# ---------------------------------------------------#
if self.mosaic:
if self.rand() < 0.5:
lines = sample(self.annotation_lines, 3) # sample(序列a，n)从序列a中随机抽取n个元素，并以list形式返回。
lines.append(self.annotation_lines[i])
shuffle(lines)
image, box = self.get_random_data_with_Mosaic(lines, self.input_shape)
else:
image, box = self.get_random_data(self.annotation_lines[i], self.input_shape, random=self.train)
else:
image, box = self.get_random_data(self.annotation_lines[i], self.input_shape, random=self.train)
image_data.append(preprocess_input(np.array(image)))
box_data.append(box)
image_data = np.array(image_data)
print(image_data.shape)
box_data = np.array(box_data)
print(box_data.shape)
y_true = self.preprocess_true_boxes(box_data, self.input_shape, self.anchors, self.num_classes)
return [image_data, *y_true], np.zeros(self.batch_size) # 在列表前加*号，会将列表拆分成一个一个的独立元素，
def on_epoch_begin(self):
shuffle(self.annotation_lines)
def rand(self, a=0, b=1):
return np.random.rand() * (b - a) + a
def get_random_data(self, annotation_line, input_shape, max_boxes=100, jitter=.3, hue=.1, sat=1.5, val=1.5,
random=True):
line = annotation_line.split()
# ------------------------------#
# 读取图像并转换成RGB图像
# ------------------------------#
image = Image.open(line[0])
image = cvtColor(image)
# ------------------------------#
# 获得图像的高宽与目标高宽
# ------------------------------#
iw, ih = image.size
h, w = input_shape
# ------------------------------#
# 获得预测框
# ------------------------------#
box = np.array([np.array(list(map(int, box.split(',')))) for box in line[1:]])
if not random:
scale = min(w / iw, h / ih)
nw = int(iw * scale)
nh = int(ih * scale)
dx = (w - nw) // 2
dy = (h - nh) // 2
# ---------------------------------#
# 将图像多余的部分加上灰条
# ---------------------------------#
image = image.resize((nw, nh), Image.BICUBIC)
new_image = Image.new('RGB', (w, h), (128, 128, 128))
new_image.paste(image, (dx, dy))
image_data = np.array(new_image, np.float32)
# ---------------------------------#
# 对真实框进行调整
# ---------------------------------#
box_data = np.zeros((max_boxes, 5))
if len(box) > 0:
np.random.shuffle(box)
box[:, [0, 2]] = box[:, [0, 2]] * nw / iw + dx
box[:, [1, 3]] = box[:, [1, 3]] * nh / ih + dy
box[:, 0:2][box[:, 0:2] < 0] = 0
box[:, 2][box[:, 2] > w] = w
box[:, 3][box[:, 3] > h] = h
box_w = box[:, 2] - box[:, 0]
box_h = box[:, 3] - box[:, 1]
box = box[np.logical_and(box_w > 1, box_h > 1)]
if len(box) > max_boxes: box = box[:max_boxes]
box_data[:len(box)] = box
return image_data, box_data
# ------------------------------------------#
# 对图像进行缩放并且进行长和宽的扭曲
# ------------------------------------------#
new_ar = w / h * self.rand(1 - jitter, 1 + jitter) / self.rand(1 - jitter, 1 + jitter)
scale = self.rand(.25, 2)
if new_ar < 1:
nh = int(scale * h)
nw = int(nh * new_ar)
else:
nw = int(scale * w)
nh = int(nw / new_ar)
image = image.resize((nw, nh), Image.BICUBIC)
# ------------------------------------------#
# 将图像多余的部分加上灰条
# ------------------------------------------#
dx = int(self.rand(0, w - nw))
dy = int(self.rand(0, h - nh))
new_image = Image.new('RGB', (w, h), (128, 128, 128))
new_image.paste(image, (dx, dy))
image = new_image
# ------------------------------------------#
# 翻转图像
# ------------------------------------------#
flip = self.rand() < .5
if flip: image = image.transpose(Image.FLIP_LEFT_RIGHT)
# ------------------------------------------#
# 色域扭曲
# ------------------------------------------#
hue = self.rand(-hue, hue)
sat = self.rand(1, sat) if self.rand() < .5 else 1 / self.rand(1, sat)
val = self.rand(1, val) if self.rand() < .5 else 1 / self.rand(1, val)
x = cv2.cvtColor(np.array(image, np.float32) / 255, cv2.COLOR_RGB2HSV)
x[..., 0] += hue * 360
x[..., 0][x[..., 0] > 1] -= 1
x[..., 0][x[..., 0] < 0] += 1
x[..., 1] *= sat
x[..., 2] *= val
x[x[:, :, 0] > 360, 0] = 360
x[:, :, 1:][x[:, :, 1:] > 1] = 1
x[x < 0] = 0
image_data = cv2.cvtColor(x, cv2.COLOR_HSV2RGB) * 255 # numpy array, 0 to 1
# ---------------------------------#
# 对真实框进行调整
# ---------------------------------#
box_data = np.zeros((max_boxes, 5))
if len(box) > 0:
np.random.shuffle(box)
box[:, [0, 2]] = box[:, [0, 2]] * nw / iw + dx
box[:, [1, 3]] = box[:, [1, 3]] * nh / ih + dy
if flip: box[:, [0, 2]] = w - box[:, [2, 0]]
box[:, 0:2][box[:, 0:2] < 0] = 0
box[:, 2][box[:, 2] > w] = w
box[:, 3][box[:, 3] > h] = h
box_w = box[:, 2] - box[:, 0]
box_h = box[:, 3] - box[:, 1]
box = box[np.logical_and(box_w > 1, box_h > 1)] # discard invalid box
if len(box) > max_boxes: box = box[:max_boxes]
box_data[:len(box)] = box
return image_data, box_data
def merge_bboxes(self, bboxes, cutx, cuty):
merge_bbox = []
for i in range(len(bboxes)):
for box in bboxes[i]:
tmp_box = []
x1, y1, x2, y2 = box[0], box[1], box[2], box[3]
if i == 0:
if y1 > cuty or x1 > cutx:
continue
if y2 >= cuty and y1 <= cuty:
y2 = cuty
if x2 >= cutx and x1 <= cutx:
x2 = cutx
if i == 1:
if y2 < cuty or x1 > cutx:
continue
if y2 >= cuty and y1 <= cuty:
y1 = cuty
if x2 >= cutx and x1 <= cutx:
x2 = cutx
if i == 2:
if y2 < cuty or x2 < cutx:
continue
if y2 >= cuty and y1 <= cuty:
y1 = cuty
if x2 >= cutx and x1 <= cutx:
x1 = cutx
if i == 3:
if y1 > cuty or x2 < cutx:
continue
if y2 >= cuty and y1 <= cuty:
y2 = cuty
if x2 >= cutx and x1 <= cutx:
x1 = cutx
tmp_box.append(x1)
tmp_box.append(y1)
tmp_box.append(x2)
tmp_box.append(y2)
tmp_box.append(box[-1])
merge_bbox.append(tmp_box)
return merge_bbox
def get_random_data_with_Mosaic(self, annotation_line, input_shape, max_boxes=100, hue=.1, sat=1.5, val=1.5):
h, w = input_shape
min_offset_x = self.rand(0.25, 0.75)
min_offset_y = self.rand(0.25, 0.75)
nws = [int(w * self.rand(0.4, 1)), int(w * self.rand(0.4, 1)), int(w * self.rand(0.4, 1)),
int(w * self.rand(0.4, 1))]
nhs = [int(h * self.rand(0.4, 1)), int(h * self.rand(0.4, 1)), int(h * self.rand(0.4, 1)),
int(h * self.rand(0.4, 1))]
place_x = [int(w * min_offset_x) - nws[0], int(w * min_offset_x) - nws[1], int(w * min_offset_x),
int(w * min_offset_x)]
place_y = [int(h * min_offset_y) - nhs[0], int(h * min_offset_y), int(h * min_offset_y),
int(h * min_offset_y) - nhs[3]]
image_datas = []
box_datas = []
index = 0
for line in annotation_line:
# 每一行进行分割
line_content = line.split()
# 打开图片
image = Image.open(line_content[0])
image = cvtColor(image)
# 图片的大小
iw, ih = image.size
# 保存框的位置
box = np.array([np.array(list(map(int, box.split(',')))) for box in line_content[1:]])
# 是否翻转图片
flip = self.rand() < .5
if flip and len(box) > 0:
image = image.transpose(Image.FLIP_LEFT_RIGHT)
box[:, [0, 2]] = iw - box[:, [2, 0]]
nw = nws[index]
nh = nhs[index]
image = image.resize((nw, nh), Image.BICUBIC)
# 将图片进行放置，分别对应四张分割图片的位置
dx = place_x[index]
dy = place_y[index]
new_image = Image.new('RGB', (w, h), (128, 128, 128))
new_image.paste(image, (dx, dy))
image_data = np.array(new_image)
index = index + 1
box_data = []
# 对box进行重新处理
if len(box) > 0:
np.random.shuffle(box)
box[:, [0, 2]] = box[:, [0, 2]] * nw / iw + dx
box[:, [1, 3]] = box[:, [1, 3]] * nh / ih + dy
box[:, 0:2][box[:, 0:2] < 0] = 0
box[:, 2][box[:, 2] > w] = w
box[:, 3][box[:, 3] > h] = h
box_w = box[:, 2] - box[:, 0]
box_h = box[:, 3] - box[:, 1]
box = box[np.logical_and(box_w > 1, box_h > 1)]
box_data = np.zeros((len(box), 5))
box_data[:len(box)] = box
image_datas.append(image_data)
box_datas.append(box_data)
# 将图片分割，放在一起
cutx = int(w * min_offset_x)
cuty = int(h * min_offset_y)
new_image = np.zeros([h, w, 3])
new_image[:cuty, :cutx, :] = image_datas[0][:cuty, :cutx, :]
new_image[cuty:, :cutx, :] = image_datas[1][cuty:, :cutx, :]
new_image[cuty:, cutx:, :] = image_datas[2][cuty:, cutx:, :]
new_image[:cuty, cutx:, :] = image_datas[3][:cuty, cutx:, :]
# 进行色域变换
hue = self.rand(-hue, hue)
sat = self.rand(1, sat) if self.rand() < .5 else 1 / self.rand(1, sat)
val = self.rand(1, val) if self.rand() < .5 else 1 / self.rand(1, val)
x = cv2.cvtColor(np.array(new_image / 255, np.float32), cv2.COLOR_RGB2HSV)
x[..., 0] += hue * 360
x[..., 0][x[..., 0] > 1] -= 1
x[..., 0][x[..., 0] < 0] += 1
x[..., 1] *= sat
x[..., 2] *= val
x[x[:, :, 0] > 360, 0] = 360
x[:, :, 1:][x[:, :, 1:] > 1] = 1
x[x < 0] = 0
new_image = cv2.cvtColor(x, cv2.COLOR_HSV2RGB) * 255
# 对框进行进一步的处理
new_boxes = self.merge_bboxes(box_datas, cutx, cuty)
# 将box进行调整
box_data = np.zeros((max_boxes, 5))
if len(new_boxes) > 0:
if len(new_boxes) > max_boxes: new_boxes = new_boxes[:max_boxes]
box_data[:len(new_boxes)] = new_boxes
return new_image, box_data
def preprocess_true_boxes(self, true_boxes, input_shape, anchors, num_classes):
assert (true_boxes[..., 4] < num_classes).all(), 'class id must be less than num_classes'
# -----------------------------------------------------------#
# 获得框的坐标和图片的大小
# -----------------------------------------------------------#
true_boxes = np.array(true_boxes, dtype='float32')
input_shape = np.array(input_shape, dtype='int32')
# -----------------------------------------------------------#
# 一共有三个特征层数
# -----------------------------------------------------------#
num_layers = len(self.anchors_mask)
# -----------------------------------------------------------#
# m为图片数量，grid_shapes为网格的shape
# -----------------------------------------------------------#
m = true_boxes.shape[0]
grid_shapes = [input_shape // {0: 32, 1: 16, 2: 8}[l] for l in range(num_layers)]
# -----------------------------------------------------------#
# y_true的格式为(m,13,13,3,85)(m,26,26,3,85)(m,52,52,3,85)
# -----------------------------------------------------------#
y_true = [np.zeros((m, grid_shapes[l][0], grid_shapes[l][1], len(self.anchors_mask[l]), 5 + num_classes),
dtype='float32') for l in range(num_layers)]
# -----------------------------------------------------------#
# 通过计算获得真实框的中心和宽高
# 中心点(m,n,2) 宽高(m,n,2)
# -----------------------------------------------------------#
boxes_xy = (true_boxes[..., 0:2] + true_boxes[..., 2:4]) // 2
boxes_wh = true_boxes[..., 2:4] - true_boxes[..., 0:2]
# -----------------------------------------------------------#
# 将真实框归一化到小数形式
# -----------------------------------------------------------#
true_boxes[..., 0:2] = boxes_xy / input_shape[::-1]
true_boxes[..., 2:4] = boxes_wh / input_shape[::-1]
# -----------------------------------------------------------#
# [9,2] -> [1,9,2]
# -----------------------------------------------------------#
anchors = np.expand_dims(anchors, 0)
anchor_maxes = anchors / 2.
anchor_mins = -anchor_maxes
# -----------------------------------------------------------#
# 长宽要大于0才有效
# -----------------------------------------------------------#
valid_mask = boxes_wh[..., 0] > 0
for b in range(m):
# -----------------------------------------------------------#
# 对每一张图进行处理
# -----------------------------------------------------------#
wh = boxes_wh[b, valid_mask[b]]
if len(wh) == 0: continue
# -----------------------------------------------------------#
# [n,2] -> [n,1,2]
# -----------------------------------------------------------#
wh = np.expand_dims(wh, -2)
box_maxes = wh / 2.
box_mins = - box_maxes
# -----------------------------------------------------------#
# 计算所有真实框和先验框的交并比
# intersect_area [n,9]
# box_area [n,1]
# anchor_area [1,9]
# iou [n,9]
# -----------------------------------------------------------#
intersect_mins = np.maximum(box_mins, anchor_mins)
intersect_maxes = np.minimum(box_maxes, anchor_maxes)
intersect_wh = np.maximum(intersect_maxes - intersect_mins, 0.)
intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
box_area = wh[..., 0] * wh[..., 1]
anchor_area = anchors[..., 0] * anchors[..., 1]
iou = intersect_area / (box_area + anchor_area - intersect_area)
# -----------------------------------------------------------#
# 维度是[n,] 感谢 消尽不死鸟 的提醒
# -----------------------------------------------------------#
best_anchor = np.argmax(iou, axis=-1)
for t, n in enumerate(best_anchor):
# -----------------------------------------------------------#
# 找到每个真实框所属的特征层
# -----------------------------------------------------------#
for l in range(num_layers):
if n in self.anchors_mask[l]:
# -----------------------------------------------------------#
# floor用于向下取整，找到真实框所属的特征层对应的x、y轴坐标
# -----------------------------------------------------------#
i = np.floor(true_boxes[b, t, 0] * grid_shapes[l][1]).astype('int32')
j = np.floor(true_boxes[b, t, 1] * grid_shapes[l][0]).astype('int32')
# -----------------------------------------------------------#
# k指的的当前这个特征点的第k个先验框
# -----------------------------------------------------------#
k = self.anchors_mask[l].index(n)
# -----------------------------------------------------------#
# c指的是当前这个真实框的种类
# -----------------------------------------------------------#
c = true_boxes[b, t, 4].astype('int32')
# -----------------------------------------------------------#
# y_true的shape为(m,13,13,3,85)(m,26,26,3,85)(m,52,52,3,85)
# 最后的85可以拆分成4+1+80，4代表的是框的中心与宽高、
# 1代表的是置信度、80代表的是种类
# -----------------------------------------------------------#
y_true[l][b, j, i, k, 0:4] = true_boxes[b, t, 0:4]
y_true[l][b, j, i, k, 4] = 1
y_true[l][b, j, i, k, 5 + c] = 1
return y_true
# train_lines
train_annotation_path = '2007_train.txt'
with open(train_annotation_path) as f:
train_lines = f.readlines()
# input_shape
input_shape = [416, 416]
# anchors
anchors_path = 'model_data/yolo_anchors.txt'
anchors, num_anchors = get_anchors(anchors_path)
# batch_size
Freeze_batch_size = 8
Unfreeze_batch_size = 4
batch_size = Freeze_batch_size
# num_classes,
classes_path = 'model_data/voc_classes.txt'
class_names, num_classes = get_classes(classes_path)
# anchors_mask
anchors_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
# mosaic
mosaic = False
train_dataloader = YoloDatasets(train_lines, input_shape, anchors, batch_size,
num_classes, anchors_mask, mosaic=mosaic, train=True)
dataset_len = len(train_dataloader)
print('dataset_len:', dataset_len)
train_lines_len = len(train_lines)
dataset_len_cal = math.ceil(train_lines_len / batch_size)
print('dataset_len_cal:', dataset_len_cal)
# a = train_dataloader[2180]3. yolov4的损失函数
以下代码使用一张图片计算loss：忽略样本那一块的代码注释感觉有问题，正在研究
"""
time : 2022/03/31
author: cong
"""
import math
import numpy as np
from model import yolo_body
from PIL import Image
import tensorflow as tf
from keras import backend as K
import cv2
def sigmoid(x):
return 1 / (1 + np.exp(-x))
# ---------------------------------------------------#
# 获得目标类别
#---------------------------------------------------#
def get_classes(classes_path):
with open(classes_path, encoding='utf-8') as f:
class_names = f.readlines()
class_names = [c.strip() for c in class_names]
return class_names, len(class_names)
#---------------------------------------------------#
# 获得先验框
#---------------------------------------------------#
def get_anchors(anchors_path):
'''loads the anchors from a file'''
with open(anchors_path, encoding='utf-8') as f:
anchors = f.readline()
anchors = [float(x) for x in anchors.split(',')]
anchors = np.array(anchors).reshape(-1, 2)
return anchors, len(anchors)
#---------------------------------------------------------#
# 将图像转换成RGB图像，防止灰度图在预测时报错。
# 代码仅仅支持RGB图像的预测，所有其它类型的图像都会转化成RGB
#---------------------------------------------------------#
def cvtColor(image):
if len(np.shape(image)) == 3 and np.shape(image)[2] == 3:
return image
else:
image = image.convert('RGB')
return image
#---------------------------------------------------#
# 对输入图像进行resize
#---------------------------------------------------#
def resize_image(image, size, letterbox_image):
# ---------------------------------------------------------------------#
# 该变量用于控制是否使用letterbox_image对输入图像进行不失真的resize，
# 在多次测试后，发现关闭letterbox_image直接resize的效果更好
# ---------------------------------------------------------------------#
iw, ih = image.size
w, h = size
if letterbox_image:
scale = min(w/iw, h/ih)
nw = int(iw*scale)
nh = int(ih*scale)
image = image.resize((nw,nh), Image.BICUBIC)
new_image = Image.new('RGB', size, (128,128,128))
# 将image 贴到new_image的图像上，起始位置在((w-nw)//2, (h-nh)//2)
new_image.paste(image, ((w-nw)//2, (h-nh)//2))
else:
new_image = image.resize((w, h), Image.BICUBIC)
return new_image
def preprocess_input(image):
image /= 255.0
return image
# 生成随机数
def rand(a=0, b=1):
return np.random.rand() * (b - a) + a
# 将预测值的每个特征层调成真实值
def get_anchors_and_decode(feats, anchors, num_classes, input_shape, calc_loss=False):
num_anchors = len(anchors)
# 生成grid
grid_shape = feats.shape[1:3]
grid_x = np.tile(np.reshape(range(grid_shape[1]), [1, -1, 1, 1]), [grid_shape[0], 1, num_anchors, 1])
grid_y = np.tile(np.reshape(range(grid_shape[0]), [-1, 1, 1, 1]), [1, grid_shape[1], num_anchors, 1])
grid = np.concatenate([grid_x, grid_y], axis=3)
# 生成anchors_reshape
anchors_reshape = np.reshape(anchors, [1, 1, num_anchors, 2])
anchors_reshape = np.tile(anchors_reshape, [grid_shape[0], grid_shape[1], 1, 1])
#
feats = np.reshape(feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])
out_xy = feats[..., :2] # 网络预测输出的x、y的偏移量
out_wh = feats[..., 2:4] # 网络预测输出的w、h的缩放尺度
xy = sigmoid(out_xy) + grid # 实际的网格中x、y的坐标
wh = np.exp(out_wh) * anchors_reshape
# 在Python中“/”表示浮点数除法，返回浮点结果，也就是结果为浮点数，
# 而“//”在Python中表示整数除法，返回不大于结果的一个最大的整数，意思就是除法结果向下取整。
box_xy = xy / grid_shape[::-1] #对应位置的元素除以13
box_wh = wh / input_shape[::-1] #对应位置的元素除以416
# ------------------------------------------#
# 获得预测框的置信度
# ------------------------------------------#
box_confidence = sigmoid(feats[..., 4:5])
box_class_probs = sigmoid(feats[..., 5:])
# ---------------------------------------------------------------------#
# 在计算loss的时候返回grid, feats, box_xy, box_wh
# 在预测的时候返回box_xy, box_wh, box_confidence, box_class_probs
# ---------------------------------------------------------------------#
if calc_loss == True:
return grid, feats, box_xy, box_wh
return box_xy, box_wh, box_confidence, box_class_probs
def box_ciou(b1, b2):
"""
输入为：
----------
b1: tensor, shape=(batch, feat_w, feat_h, anchor_num, 4), xywh
b2: tensor, shape=(batch, feat_w, feat_h, anchor_num, 4), xywh
返回为：
-------
ciou: tensor, shape=(batch, feat_w, feat_h, anchor_num, 1)
"""
# -----------------------------------------------------------#
# 求出预测框左上角右下角
# b1_mins (batch, feat_w, feat_h, anchor_num, 2)
# b1_maxes (batch, feat_w, feat_h, anchor_num, 2)
# -----------------------------------------------------------#
b1_xy = b1[..., :2]
b1_wh = b1[..., 2:4]
b1_wh_half = b1_wh / 2.
b1_mins = b1_xy - b1_wh_half
b1_maxes = b1_xy + b1_wh_half
# -----------------------------------------------------------#
# 求出真实框左上角右下角
# b2_mins (batch, feat_w, feat_h, anchor_num, 2)
# b2_maxes (batch, feat_w, feat_h, anchor_num, 2)
# -----------------------------------------------------------#
b2_xy = b2[..., :2]
b2_wh = b2[..., 2:4]
b2_wh_half = b2_wh / 2.
b2_mins = b2_xy - b2_wh_half
b2_maxes = b2_xy + b2_wh_half
# -----------------------------------------------------------#
# 求真实框和预测框所有的iou
# iou (batch, feat_w, feat_h, anchor_num)
# -----------------------------------------------------------#
intersect_mins = K.maximum(b1_mins, b2_mins)
intersect_maxes = K.minimum(b1_maxes, b2_maxes)
intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)
intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
b1_area = b1_wh[..., 0] * b1_wh[..., 1]
b2_area = b2_wh[..., 0] * b2_wh[..., 1]
union_area = b1_area + b2_area - intersect_area
iou = intersect_area / K.maximum(union_area, K.epsilon())
# -----------------------------------------------------------#
# 计算中心的差距
# center_distance (batch, feat_w, feat_h, anchor_num)
# -----------------------------------------------------------#
center_distance = K.sum(K.square(b1_xy - b2_xy), axis=-1)
enclose_mins = K.minimum(b1_mins, b2_mins)
enclose_maxes = K.maximum(b1_maxes, b2_maxes)
enclose_wh = K.maximum(enclose_maxes - enclose_mins, 0.0)
# -----------------------------------------------------------#
# 计算对角线距离
# enclose_diagonal (batch, feat_w, feat_h, anchor_num)
# -----------------------------------------------------------#
enclose_diagonal = K.sum(K.square(enclose_wh), axis=-1)
ciou = iou - 1.0 * (center_distance) / K.maximum(enclose_diagonal, K.epsilon())
v = 4 * K.square(tf.math.atan2(b1_wh[..., 0], K.maximum(b1_wh[..., 1],
K.epsilon())) - tf.math.atan2(b2_wh[..., 0],
K.maximum(b2_wh[..., 1],K.epsilon()))) /(math.pi * math.pi)
alpha = v / K.maximum((1.0 - iou + v), K.epsilon())
ciou = ciou - alpha * v
ciou = K.expand_dims(ciou, -1)
return ciou
# ---------------------------------------------------#
# 平滑标签
# ---------------------------------------------------#
def _smooth_labels(y_true, label_smoothing):
num_classes = tf.cast(K.shape(y_true)[-1], dtype=K.floatx())
label_smoothing = K.constant(label_smoothing, dtype=K.floatx())
return y_true * (1.0 - label_smoothing) + label_smoothing / num_classes
# ---------------------------------------------------#
# 用于计算每个预测框与真实框的iou
# ---------------------------------------------------#
def box_iou(b1, b2):
# ---------------------------------------------------#
# num_anchor,1,4
# 计算左上角的坐标和右下角的坐标
# ---------------------------------------------------#
b1 = K.expand_dims(b1, -2)
b1_xy = b1[..., :2]
b1_wh = b1[..., 2:4]
b1_wh_half = b1_wh / 2.
b1_mins = b1_xy - b1_wh_half
b1_maxes = b1_xy + b1_wh_half
# ---------------------------------------------------#
# 1,n,4
# 计算左上角和右下角的坐标
# ---------------------------------------------------#
b2 = K.expand_dims(b2, 0)
b2_xy = b2[..., :2]
b2_wh = b2[..., 2:4]
b2_wh_half = b2_wh / 2.
b2_mins = b2_xy - b2_wh_half
b2_maxes = b2_xy + b2_wh_half
# ---------------------------------------------------#
# 计算重合面积
# ---------------------------------------------------#
intersect_mins = K.maximum(b1_mins, b2_mins)
intersect_maxes = K.minimum(b1_maxes, b2_maxes)
intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)
intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
b1_area = b1_wh[..., 0] * b1_wh[..., 1]
b2_area = b2_wh[..., 0] * b2_wh[..., 1]
iou = intersect_area / (b1_area + b2_area - intersect_area)
return iou
def yolo_loss(args, input_shape, anchors, anchors_mask, num_classes, ignore_thresh=.5, label_smoothing=0.1,
print_loss=False):
num_layers = len(anchors_mask)
# ---------------------------------------------------------------------------------------------------#
# 将预测结果和实际ground truth分开，args是[*model_body.output, *y_true]
# y_true是一个列表，包含三个特征层，shape分别为:
# (m,13,13,3,85)
# (m,26,26,3,85)
# (m,52,52,3,85)
# yolo_outputs是一个列表，包含三个特征层，shape分别为:
# (m,13,13,3,85)
# (m,26,26,3,85)
# (m,52,52,3,85)
# ---------------------------------------------------------------------------------------------------#
y_true = args[num_layers:]
yolo_outputs = args[:num_layers]
# -----------------------------------------------------------#
# 得到input_shpae为416,416
# -----------------------------------------------------------#
#input_shape = K.cast(input_shape, K.dtype(y_true[0]))
# input_shape = K.constant(input_shape)
# -----------------------------------------------------------#
# 取出每一张图片
# m的值就是batch_size
# -----------------------------------------------------------#
m = K.shape(yolo_outputs[0])[0]
loss = 0
num_pos = 0
# ---------------------------------------------------------------------------------------------------#
# y_true是一个列表，包含三个特征层，shape分别为(m,13,13,3,85),(m,26,26,3,85),(m,52,52,3,85)。
# yolo_outputs是一个列表，包含三个特征层，shape分别为(m,13,13,3,85),(m,26,26,3,85),(m,52,52,3,85)。
# ---------------------------------------------------------------------------------------------------#
for l in range(num_layers):
# -----------------------------------------------------------#
# 以第一个特征层(m,13,13,3,85)为例子
# 取出该特征层中存在目标的点的位置。(m,13,13,3,1)
# -----------------------------------------------------------#
object_mask = y_true[l][..., 4:5]
# -----------------------------------------------------------#
# 取出其对应的种类(m,13,13,3,80)
# -----------------------------------------------------------#
true_class_probs = y_true[l][..., 5:]
if label_smoothing:
true_class_probs = _smooth_labels(true_class_probs, label_smoothing)
# -----------------------------------------------------------#
# 将yolo_outputs的特征层输出进行处理、获得四个返回值
# 其中：
# grid (13,13,1,2) 网格坐标
# raw_pred (m,13,13,3,85) 尚未处理的预测结果
# pred_xy (m,13,13,3,2) 解码后的中心坐标
# pred_wh (m,13,13,3,2) 解码后的宽高坐标
# -----------------------------------------------------------#
grid, raw_pred, pred_xy, pred_wh = get_anchors_and_decode(yolo_outputs[l],
anchors[anchors_mask[l]], num_classes, input_shape,
calc_loss=True)
# -----------------------------------------------------------#
# pred_box是解码后的预测的box的位置
# (m,13,13,3,4)
# -----------------------------------------------------------#
pred_box = K.concatenate([pred_xy, pred_wh])
# -----------------------------------------------------------#
# 找到负样本群组，第一步是创建一个数组，[]
# TensorArray可以看做是具有动态size功能的Tensor数组。通常都是跟while_loop或map_fn结合使用
# -----------------------------------------------------------#
ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)
object_mask_bool = K.cast(object_mask, 'bool')
# -----------------------------------------------------------#
# 对每一张图片计算ignore_mask
# -----------------------------------------------------------#
def loop_body(b, ignore_mask):
# -----------------------------------------------------------#
# 取出n个真实框：n,4
# -----------------------------------------------------------#
# tf.boolean_mask的作用是通过布尔值过滤元素
true_box = tf.boolean_mask(y_true[l][b, ..., 0:4], object_mask_bool[b, ..., 0])
# -----------------------------------------------------------#
# 计算预测框与真实框的iou
# pred_box 13,13,3,4 预测框的坐标
# true_box n,4 真实框的坐标
# iou 13,13,3,n 预测框和真实框的iou
# -----------------------------------------------------------#
iou = box_iou(pred_box[b], true_box)
# -----------------------------------------------------------#
# best_iou 13,13,3 每个特征点与真实框的最大重合程度
# -----------------------------------------------------------#
#选出最大的iou
best_iou = K.max(iou, axis=-1)
# -----------------------------------------------------------#
# 判断预测框和真实框的最大iou小于ignore_thresh
# 则认为该预测框没有与之对应的真实框
# 该操作的目的是：
# 忽略预测结果与真实框非常对应特征点，因为这些框已经比较准了
# 不适合当作负样本，所以忽略掉。
# -----------------------------------------------------------#
ignore_mask = ignore_mask.write(b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))
return b + 1, ignore_mask
# -----------------------------------------------------------#
# 在这个地方进行一个循环、循环是对每一张图片进行的
# -----------------------------------------------------------#
_, ignore_mask = tf.while_loop(lambda b, *args: b < m, loop_body, [0, ignore_mask])
# -----------------------------------------------------------#
# ignore_mask用于提取出作为负样本的特征点
# (m,13,13,3)
# -----------------------------------------------------------#
ignore_mask = ignore_mask.stack()
# (m,13,13,3,1)
ignore_mask = K.expand_dims(ignore_mask, -1)
# -----------------------------------------------------------#
# 真实框越大，比重越小，小框的比重更大。
# -----------------------------------------------------------#
box_loss_scale = 2 - y_true[l][..., 2:3] * y_true[l][..., 3:4]
# box_loss_scale:(1,52,52,3,1)
# -----------------------------------------------------------#
# 计算Ciou loss
# -----------------------------------------------------------#
raw_true_box = y_true[l][..., 0:4]
ciou = box_ciou(pred_box, raw_true_box) # (1,52,52,3,1)
ciou_loss = object_mask * box_loss_scale * (1 - ciou) # (1,52,52,3,1)
# ------------------------------------------------------------------------------#
# 如果该位置本来有框，那么计算1与置信度的交叉熵
# 如果该位置本来没有框，那么计算0与置信度的交叉熵
# 在这其中会忽略一部分样本，这些被忽略的样本满足条件best_iou<ignore_thresh
# 该操作的目的是：
# 忽略预测结果与真实框非常对应特征点，因为这些框已经比较准了
# 不适合当作负样本，所以忽略掉。
# ------------------------------------------------------------------------------#
confidence_loss = object_mask * K.binary_crossentropy(object_mask, raw_pred[..., 4:5], from_logits=True) + \
(1 - object_mask) * K.binary_crossentropy(object_mask, raw_pred[..., 4:5],
from_logits=True) * ignore_mask
class_loss = object_mask * K.binary_crossentropy(true_class_probs, raw_pred[..., 5:], from_logits=True)
location_loss = K.sum(ciou_loss)
confidence_loss = K.sum(confidence_loss)
class_loss = K.sum(class_loss)
# -----------------------------------------------------------#
# 计算正样本数量
# -----------------------------------------------------------#
num_pos += tf.maximum(K.sum(K.cast(object_mask, tf.float32)), 1)
loss += location_loss + confidence_loss + class_loss
# if print_loss:
# loss = tf.Print(loss, [loss, location_loss, confidence_loss, class_loss, K.sum(ignore_mask)], message='loss: ')
loss = loss / num_pos
return loss
class_path = 'model_data/coco_classes.txt'
model_path = 'model_data/yolo4_weight.h5'
anchors_path = 'model_data/yolo_anchors.txt'
max_boxes = 100 # 最大框的数量
confidence = 0.5 # 只有得分大于置信度的预测框会被保留下来
nms_iou = 0.5 # 非极大抑制所用到的nms_iou大小
# 该变量用于控制是否使用letterbox_image对输入图像进行不失真的resize，
# 在多次测试后，发现关闭letterbox_image直接resize的效果更好
# ---------------------------------------------------------------------#
letterbox_image = False
input_shape = [416, 416]
anchors_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
class_names, num_classes = get_classes(class_path)
anchors_all, num_anchors = get_anchors(anchors_path)
annotation_line = '/Users/cong/yolov4-keras-master/VOCdevkit/VOC2007/JPEGImages/007826.jpg 80,217,320,273,10 197,193,257,326,8 258,180,312,314,8 10,195,93,358,8 82,252,243,372,8\n'
line = annotation_line.split()
image = Image.open(line[0])
print('names:', class_names)
image = cvtColor(image) # print('image_shape:', np.shape(image))# image_shape: (1080, 1920, 3)
image_data = resize_image(image, input_shape, letterbox_image=False)
# 添加上batch_size维度，并进行归一化,
# np.expand_dims(a, axis=0)表示在0位置添加数据,
image_data = np.expand_dims(preprocess_input(np.array(image_data, dtype='float32')), 0)
#print(image_data.shape) (1, 416, 416, 3)
# 加载模型
model = yolo_body((None, None, 3), anchors_mask, num_classes)
# model.summary()
model.load_weights(model_path)
model_output = model.predict(image_data)
# print('model:', model)
# 以上为模型推理的结果
###########################################################
# 以下找出该图片的y_true
iw, ih = image.size
h, w = input_shape
# 获得预测框, int向下取整, box.split(','): ['80', '217', '320', '273', '10']
box = np.array([np.array(list(map(int, box.split(',')))) for box in line[1:]])
# anchors_mask
anchors_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
box_data = np.zeros((max_boxes, 5))
if len(box) > 0:
# np.random.shuffle(box)
if len(box) > max_boxes:
box = box[:max_boxes]
box_data[:len(box)] = box
box_datas_array = np.expand_dims(box_data, axis=0)
assert (box_datas_array[..., 4] < num_classes).all(), 'class id must be less than num_classes'
# -----------------------------------------------------------#
# 获得框的坐标和图片的大小
# -----------------------------------------------------------#
true_boxes = np.array(box_datas_array, dtype='float32')
input_shape = np.array(input_shape, dtype='int32')
# -----------------------------------------------------------#
# 一共有三个特征层数
# -----------------------------------------------------------#
num_layers = len(anchors_mask)
# -----------------------------------------------------------#
# m为图片数量，grid_shapes为网格的shape
# -----------------------------------------------------------#
m = true_boxes.shape[0]
grid_shapes = [input_shape // {0: 32, 1: 16, 2: 8}[l] for l in range(num_layers)]
# -----------------------------------------------------------#
# y_true的格式为(m,13,13,3,85)(m,26,26,3,85)(m,52,52,3,85)
# -----------------------------------------------------------#
y_true = [np.zeros((m, grid_shapes[l][0], grid_shapes[l][1], len(anchors_mask[l]), 5 + num_classes),
dtype='float32') for l in range(num_layers)]
# -----------------------------------------------------------#
# 通过计算获得真实框的中心和宽高
# 中心点(m,n,2) 宽高(m,n,2)
# -----------------------------------------------------------#
boxes_xy = (true_boxes[..., 0:2] + true_boxes[..., 2:4]) // 2 # (xmax - xmin)/2 + xmin = (xmin + xmax)/2
boxes_wh = true_boxes[..., 2:4] - true_boxes[..., 0:2]
# -----------------------------------------------------------#
# 将真实框归一化到小数形式
# -----------------------------------------------------------#
true_boxes[..., 0:2] = boxes_xy / input_shape[::-1]
true_boxes[..., 2:4] = boxes_wh / input_shape[::-1]
# -----------------------------------------------------------#
# [9,2] -> [1,9,2]
# -----------------------------------------------------------#
anchors = np.expand_dims(anchors_all, 0)
anchor_maxes = anchors / 2.
anchor_mins = -anchor_maxes
# -----------------------------------------------------------#
# 长宽要大于0才有效
# -----------------------------------------------------------#
valid_mask = boxes_wh[..., 0] > 0 # box_wh:(100,100,2)
# valid_mask:(100)
for b in range(m):
# -----------------------------------------------------------#
# 对每一张图进行处理
# -----------------------------------------------------------#
wh = boxes_wh[b, valid_mask[b]] # array:(5,2)
if len(wh) == 0: continue
# -----------------------------------------------------------#
# [n,2] -> [n,1,2]
# -----------------------------------------------------------#
wh = np.expand_dims(wh, -2) #array:(5,1,2)
box_maxes = wh / 2.
box_mins = - box_maxes
# -----------------------------------------------------------#
# 计算所有真实框和先验框的交并比
# intersect_area [n,9]
# box_area [n,1]
# anchor_area [1,9]
# iou [n,9]
# -----------------------------------------------------------#
# 一张图像上 所有目标框与anchor的交并比
intersect_mins = np.maximum(box_mins, anchor_mins)
intersect_maxes = np.minimum(box_maxes, anchor_maxes)
intersect_wh = np.maximum(intersect_maxes - intersect_mins, 0.)
intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
box_area = wh[..., 0] * wh[..., 1]
anchor_area = anchors[..., 0] * anchors[..., 1]
iou = intersect_area / (box_area + anchor_area - intersect_area) # (n,9)
# -----------------------------------------------------------#
# 维度是[n,] 感谢 消尽不死鸟 的提醒
# -----------------------------------------------------------#
best_anchor = np.argmax(iou, axis=-1) #(6,5,5,5,6)# 索引为5的anchor在26*26的特征图上同时预测3个物体
for t, n in enumerate(best_anchor):
# 找到每个真实框所属的特征层，
# n:best_anchor的最大索引，
# b：第几张图片
# t: 这张图片上的第几个框
# l：哪个特征层上的anchor
for l in range(num_layers):
if n in anchors_mask[l]:
# floor用于向下取整，找到真实框所属的特征层对应的x、y轴顶点坐标#
i = np.floor(true_boxes[b, t, 0] * grid_shapes[l][1]).astype('int32')
j = np.floor(true_boxes[b, t, 1] * grid_shapes[l][0]).astype('int32')
# k指的的当前这个特征层的第k个先验框
k = anchors_mask[l].index(n)
# c指的是当前这个真实框的种类
c = true_boxes[b, t, 4].astype('int32')#
# y_true的shape为(m,13,13,3,85)(m,26,26,3,85)(m,52,52,3,85)
# 最后的85可以拆分成4+1+80，4代表的是框的中心与宽高、
# 1代表的是置信度、80代表的是种类
# -----------------------------------------------------------#
y_true[l][b, j, i, k, 0:4] = true_boxes[b, t, 0:4]# J是第几行，所以算的是ymin的长度 i是第几列，要计算x方向的长度
y_true[l][b, j, i, k, 4] = 1
y_true[l][b, j, i, k, 5 + c] = 1
# 将预测结果和实际ground truth分开，args是[*model_body.output, *y_true]
args = [*model_output, *y_true]
ignore_thresh=.5,
label_smoothing=0.1,
print_loss=False
num_layers = len(anchors_mask)
y_true = [K.constant(i) for i in args[num_layers:]]
yolo_outputs = args[:num_layers]
# -----------------------------------------------------------#
# 得到input_shpae为416,416
# -----------------------------------------------------------#
# input_shape = K.constant(input_shape)
# -----------------------------------------------------------#
# 取出每一张图片
# m的值就是batch_size
# -----------------------------------------------------------#
m = K.shape(yolo_outputs[0])[0]
loss = 0
num_pos = 0
# ---------------------------------------------------------------------------------------------------#
# y_true是一个列表，包含三个特征层，shape分别为(m,13,13,3,85),(m,26,26,3,85),(m,52,52,3,85)。
# yolo_outputs是一个列表，包含三个特征层，shape分别为(m,13,13,3,85),(m,26,26,3,85),(m,52,52,3,85)。
# ---------------------------------------------------------------------------------------------------#
def loop_body(b, ignore_mask):
# 取出n个真实框：(n,4)
# tf.boolean_mask的作用是通过布尔值过滤元素
true_box = tf.boolean_mask(y_true[l][b, ..., 0:4], object_mask_bool[b, ..., 0])
print(true_box.shape)
# -----------------------------------------------------------#
# 计算预测框与真实框的iou
# pred_box 13,13,3,4 预测框的坐标
# true_box n,4 真实框的坐标
# iou 13,13,3,n 预测框和真实框的iou
# -----------------------------------------------------------#
iou = box_iou(pred_box[b], true_box) # (13, 13, 3, ?)
print(iou.shape)
# -----------------------------------------------------------#
# best_iou (13,13,3) 每个特征点的3个预测框与真实框的最大重合程度
# -----------------------------------------------------------#
best_iou = K.max(iou, axis=-1)
print(best_iou.shape)
ignore_mask = ignore_mask.write(b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))
print('ignore_mask_shape:', K.cast(best_iou < ignore_thresh, K.dtype(true_box)).shape)
return b + 1, ignore_mask
for l in range(num_layers-2):
print('l:', l)
# -----------------------------------------------------------#
# 以第一个特征层(m,13,13,3,85)为例子
# 取出该特征层中存在目标的点的位置。(m,13,13,3,1)
# -----------------------------------------------------------#
object_mask = y_true[l][..., 4:5]
# -----------------------------------------------------------#
# 取出其对应的种类(m,13,13,3,80)
# -----------------------------------------------------------#
true_class_probs = y_true[l][..., 5:]
if label_smoothing:
true_class_probs = _smooth_labels(true_class_probs, label_smoothing)
#标签平滑的意思比如真实值编码为[1, 0, 0], 现在将真实值变为[0.95, 0.025, 0.025]
#其实就是对训练过程进行惩罚。
# -----------------------------------------------------------#
# 将yolo_outputs的特征层输出进行处理、获得四个返回值
# 其中：
# grid (13,13,1,2) 网格坐标
# raw_pred (m,13,13,3,85) 尚未处理的预测结果
# pred_xy (m,13,13,3,2) 解码后的中心坐标
# pred_wh (m,13,13,3,2) 解码后的宽高坐标
# -----------------------------------------------------------#
grid, raw_pred, pred_xy, pred_wh = get_anchors_and_decode(yolo_outputs[l],
anchors_all[anchors_mask[l]], num_classes, input_shape,
calc_loss=True)
# -----------------------------------------------------------#
# pred_box是解码后的预测的box的位置
# (m,13,13,3,4)
# -----------------------------------------------------------#
pred_box = K.concatenate([K.constant(pred_xy), K.constant(pred_wh)])
# y_true = K.constant(y_true)
# -----------------------------------------------------------#
# 找到负样本群组，第一步是创建一个数组，[]
# dynamic_size指定数组长度可变
# -----------------------------------------------------------#
ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)
object_mask_bool = K.cast(object_mask, 'bool') # (1,52,52,3,1)
# -----------------------------------------------------------#
# 对每一张图片计算ignore_mask
# -----------------------------------------------------------#
# -----------------------------------------------------------#
# 在这个地方进行一个循环、循环是对每一张图片进行的
# -----------------------------------------------------------#
_, ignore_mask = tf.while_loop(lambda b, *args: b < m, loop_body, [0, ignore_mask])
#
# -----------------------------------------------------------#
# ignore_mask用于提取出作为负样本的特征点
# (m,13,13,3)
# -----------------------------------------------------------#
ignore_mask = ignore_mask.stack() # 将TensorArray中元素叠起来当做一个Tensor输出
# (m,13,13,3,1)
ignore_mask = K.expand_dims(ignore_mask, -1)
# (?,13,13,3,1)
# -----------------------------------------------------------#
# 真实框越大，比重越小，小框的比重更大。
# -----------------------------------------------------------#
box_loss_scale = 2 - y_true[l][..., 2:3] * y_true[l][..., 3:4] # (1,52,52,3,1)
# -----------------------------------------------------------#
# 计算Ciou loss
# -----------------------------------------------------------#
raw_true_box = y_true[l][..., 0:4]
ciou = box_ciou(pred_box, raw_true_box) #
ciou_loss = object_mask * box_loss_scale * (1 - ciou)
# ------------------------------------------------------------------------------#
# 如果该位置本来有框，那么计算1与置信度的交叉熵
# 如果该位置本来没有框，那么计算0与置信度的交叉熵
# binary_cross_entropy是二分类的交叉熵，实际是多分类softmax_cross_entropy的一种特殊情况，
# 当多分类中，类别只有两类时，即0或者1，即为二分类，二分类也是一个逻辑回归问题，也可以套用逻辑回归的损失函数。
# ------------------------------------------------------------------------------#
confidence_loss = object_mask * K.binary_crossentropy(object_mask, raw_pred[..., 4:5], from_logits=True) + \
(1 - object_mask) * K.binary_crossentropy(object_mask, raw_pred[..., 4:5],
from_logits=True) * ignore_mask
class_loss = object_mask * K.binary_crossentropy(true_class_probs, raw_pred[..., 5:], from_logits=True)
location_loss = K.sum(ciou_loss)#沿着指定轴计算张量之和
confidence_loss = K.sum(confidence_loss)
class_loss = K.sum(class_loss)
# -----------------------------------------------------------#
# 计算正样本数量，没有正样本的时候K.sum(K.cast(object_mask, tf.float32))=0，，num_pos取值为1
# -----------------------------------------------------------#
num_pos += tf.maximum(K.sum(K.cast(object_mask, tf.float32)), 1) # 返回最大值
loss += location_loss + confidence_loss + class_loss
# if print_loss:
# loss = tf.Print(loss, [loss, location_loss, confidence_loss, class_loss, K.sum(ignore_mask)], message='loss: ')
loss = loss / num_pos
sess = K.get_session()
loss = sess.run(loss)
location_loss = sess.run(location_loss)
object_mask_num = sess.run(K.sum(K.cast(object_mask, tf.float32)))
ignore_mask = sess.run(ignore_mask)
object_mask = sess.run(object_mask)
confidence_loss = sess.run(confidence_loss)
num_pos = sess.run(num_pos)
y_true = sess.run(y_true)
print('loss:', loss)
print('location_loss:', location_loss)
# print('ignore_mask',ignore_mask)
print('object_mask_num:', object_mask_num)
print('num_pos:', num_pos)4. yolov4的训练设置
import keras.backend as K
from keras.callbacks import EarlyStopping, ModelCheckpoint, TensorBoard
from keras.optimizers import Adam
from nets.yolo import get_train_model, yolo_body
from utils.callbacks import (ExponentDecayScheduler, LossHistory,
WarmUpCosineDecayScheduler)
from utils.dataloader import YoloDatasets
# from yolo_datasets_aug import YoloDatasets
from utils.utils import get_anchors, get_classes
'''
训练自己的目标检测模型一定需要注意以下几点：
1、训练前仔细检查自己的格式是否满足要求，该库要求数据集格式为VOC格式，需要准备好的内容有输入图片和标签
输入图片为.jpg图片，无需固定大小，传入训练前会自动进行resize。
灰度图会自动转成RGB图片进行训练，无需自己修改。
输入图片如果后缀非jpg，需要自己批量转成jpg后再开始训练。
标签为.xml格式，文件中会有需要检测的目标信息，标签文件和输入图片文件相对应。
2、训练好的权值文件保存在logs文件夹中，每个epoch都会保存一次，如果只是训练了几个step是不会保存的，epoch和step的概念要捋清楚一下。
在训练过程中，该代码并没有设定只保存最低损失的，因此按默认参数训练完会有100个权值，如果空间不够可以自行删除。
这个并不是保存越少越好也不是保存越多越好，有人想要都保存、有人想只保存一点，为了满足大多数的需求，还是都保存可选择性高。
3、损失值的大小用于判断是否收敛，比较重要的是有收敛的趋势，即验证集损失不断下降，如果验证集损失基本上不改变的话，模型基本上就收敛了。
损失值的具体大小并没有什么意义，大和小只在于损失的计算方式，并不是接近于0才好。如果想要让损失好看点，可以直接到对应的损失函数里面除上10000。
训练过程中的损失值会保存在logs文件夹下的loss_%Y_%m_%d_%H_%M_%S文件夹中
4、调参是一门蛮重要的学问，没有什么参数是一定好的，现有的参数是我测试过可以正常训练的参数，因此我会建议用现有的参数。
但是参数本身并不是绝对的，比如随着batch的增大学习率也可以增大，效果也会好一些；过深的网络不要用太大的学习率等等。
这些都是经验上，只能靠各位同学多查询资料和自己试试了。
'''
if __name__ == "__main__":
#--------------------------------------------------------#
# 训练前一定要修改classes_path，使其对应自己的数据集
#--------------------------------------------------------#
classes_path = 'model_data/voc_classes.txt'
#---------------------------------------------------------------------#
# anchors_path代表先验框对应的txt文件，一般不修改。
# anchors_mask用于帮助代码找到对应的先验框，一般不修改。
#---------------------------------------------------------------------#
anchors_path = 'model_data/yolo_anchors.txt'
anchors_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
#----------------------------------------------------------------------------------------------------------------------------#
# 权值文件的下载请看README，可以通过网盘下载。模型的 预训练权重 对不同数据集是通用的，因为特征是通用的。
# 模型的 预训练权重 比较重要的部分是 主干特征提取网络的权值部分，用于进行特征提取。
# 预训练权重对于99%的情况都必须要用，不用的话主干部分的权值太过随机，特征提取效果不明显，网络训练的结果也不会好
#
# 如果训练过程中存在中断训练的操作，可以将model_path设置成logs文件夹下的权值文件，将已经训练了一部分的权值再次载入。
# 同时修改下方的 冻结阶段 或者 解冻阶段 的参数，来保证模型epoch的连续性。
#
# 当model_path = ''的时候不加载整个模型的权值。
#
# 此处使用的是整个模型的权重，因此是在train.py进行加载的。
# 如果想要让模型从主干的预训练权值开始训练，则设置model_path为主干网络的权值，此时仅加载主干。
# 如果想要让模型从0开始训练，则设置model_path = ''，Freeze_Train = Fasle，此时从0开始训练，且没有冻结主干的过程。
# 一般来讲，从0开始训练效果会很差，因为权值太过随机，特征提取效果不明显。
#
# 网络一般不从0开始训练，至少会使用主干部分的权值，有些论文提到可以不用预训练，主要原因是他们 数据集较大 且 调参能力优秀。
# 如果一定要训练网络的主干部分，可以了解imagenet数据集，首先训练分类模型，分类模型的 主干部分 和该模型通用，基于此进行训练。
#----------------------------------------------------------------------------------------------------------------------------#
model_path = 'model_data/yolo4_weight.h5'
#------------------------------------------------------#
# 输入的shape大小，一定要是32的倍数
#------------------------------------------------------#
input_shape = [416, 416]
#------------------------------------------------------#
# Yolov4的tricks应用
# mosaic 马赛克数据增强 True or False
# 实际测试时mosaic数据增强并不稳定，所以默认为False
# Cosine_scheduler 余弦退火学习率 True or False
# label_smoothing 标签平滑 0.01以下一般 如0.01、0.005
#------------------------------------------------------#
mosaic = False
Cosine_scheduler = False
label_smoothing = 0
#----------------------------------------------------#
# 训练分为两个阶段，分别是冻结阶段和解冻阶段。
# 显存不足与数据集大小无关，提示显存不足请调小batch_size。
# 受到BatchNorm层影响，batch_size最小为2，不能为1。
#----------------------------------------------------#
#----------------------------------------------------#
# 冻结阶段训练参数
# 此时模型的主干被冻结了，特征提取网络不发生改变
# 占用的显存较小，仅对网络进行微调
#----------------------------------------------------#
Init_Epoch = 0
Freeze_Epoch = 50
Freeze_batch_size = 8
Freeze_lr = 1e-3
#----------------------------------------------------#
# 解冻阶段训练参数
# 此时模型的主干不被冻结了，特征提取网络会发生改变
# 占用的显存较大，网络所有的参数都会发生改变
#----------------------------------------------------#
UnFreeze_Epoch = 100
Unfreeze_batch_size = 4
Unfreeze_lr = 1e-4
#------------------------------------------------------#
# 是否进行冻结训练，默认先冻结主干训练后解冻训练。
#------------------------------------------------------#
Freeze_Train = True
#------------------------------------------------------#
# 用于设置是否使用多线程读取数据，1代表关闭多线程
# 开启后会加快数据读取速度，但是会占用更多内存
# keras里开启多线程有些时候速度反而慢了许多
# 在IO为瓶颈的时候再开启多线程，即GPU运算速度远大于读取图片的速度。
#------------------------------------------------------#
num_workers = 1
#----------------------------------------------------#
# 获得图片路径和标签
#----------------------------------------------------#
train_annotation_path = '2007_train.txt'
val_annotation_path = '2007_val.txt'
#----------------------------------------------------#
# 获取classes和anchor
#----------------------------------------------------#
class_names, num_classes = get_classes(classes_path)
anchors, num_anchors = get_anchors(anchors_path)
K.clear_session()
#------------------------------------------------------#
# 创建yolo模型
#------------------------------------------------------#
model_body = yolo_body((None, None, 3), anchors_mask, num_classes)
if model_path != '':
#------------------------------------------------------#
# 载入预训练权重
#------------------------------------------------------#
print('Load weights {}.'.format(model_path))
model_body.load_weights(model_path, by_name=True, skip_mismatch=True)
model = get_train_model(model_body, input_shape, num_classes, anchors, anchors_mask, label_smoothing)
#-------------------------------------------------------------------------------#
# 训练参数的设置
# logging表示tensorboard的保存地址
# checkpoint用于设置权值保存的细节，period用于修改多少epoch保存一次
# reduce_lr用于设置学习率下降的方式
# early_stopping用于设定早停，val_loss多次不下降自动结束训练，表示模型基本收敛
#-------------------------------------------------------------------------------#
logging = TensorBoard(log_dir = 'logs/')
checkpoint = ModelCheckpoint('logs/ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5',
monitor = 'val_loss', save_weights_only = True, save_best_only = False, period = 1)
if Cosine_scheduler:
reduce_lr = WarmUpCosineDecayScheduler(T_max = 5, eta_min = 1e-5, verbose = 1)
else:
reduce_lr = ExponentDecayScheduler(decay_rate = 0.94, verbose = 1)
early_stopping = EarlyStopping(monitor='val_loss', min_delta = 0, patience = 10, verbose = 1)
loss_history = LossHistory('logs/')
#---------------------------#
# 读取数据集对应的txt
#---------------------------#
with open(train_annotation_path) as f:
train_lines = f.readlines()
with open(val_annotation_path) as f:
val_lines = f.readlines()
num_train = len(train_lines)
num_val = len(val_lines)
if Freeze_Train:
freeze_layers = 249
for i in range(freeze_layers): model_body.layers[i].trainable = False
print('Freeze the first {} layers of total {} layers.'.format(freeze_layers, len(model_body.layers)))
#------------------------------------------------------#
# 主干特征提取网络特征通用，冻结训练可以加快训练速度
# 也可以在训练初期防止权值被破坏。
# Init_Epoch为起始世代
# Freeze_Epoch为冻结训练的世代
# UnFreeze_Epoch总训练世代
# 提示OOM或者显存不足请调小Batch_size
#------------------------------------------------------#
if True:
batch_size = Freeze_batch_size
lr = Freeze_lr
start_epoch = Init_Epoch
end_epoch = Freeze_Epoch
epoch_step = num_train // batch_size
epoch_step_val = num_val // batch_size
if epoch_step == 0 or epoch_step_val == 0:
raise ValueError('数据集过小，无法进行训练，请扩充数据集。')
model.compile(optimizer=Adam(lr = lr), loss={'yolo_loss': lambda y_true, y_pred: y_pred})
train_dataloader = YoloDatasets(train_lines, input_shape, anchors, batch_size, num_classes, anchors_mask, mosaic = mosaic, train = True)
val_dataloader = YoloDatasets(val_lines, input_shape, anchors, batch_size, num_classes, anchors_mask, mosaic = False, train = False)
print('Train on {} samples, val on {} samples, with batch size {}.'.format(num_train, num_val, batch_size))
model.fit_generator(
generator = train_dataloader,
steps_per_epoch = epoch_step,
validation_data = val_dataloader,
validation_steps = epoch_step_val,
epochs = end_epoch,
initial_epoch = start_epoch,
use_multiprocessing = True if num_workers > 1 else False,
workers = num_workers,
callbacks = [logging, checkpoint, reduce_lr, early_stopping, loss_history]
)
if Freeze_Train:
for i in range(freeze_layers): model_body.layers[i].trainable = True
if True:
batch_size = Unfreeze_batch_size
lr = Unfreeze_lr
start_epoch = Freeze_Epoch
end_epoch = UnFreeze_Epoch
epoch_step = num_train // batch_size
epoch_step_val = num_val // batch_size
if epoch_step == 0 or epoch_step_val == 0:
raise ValueError('数据集过小，无法进行训练，请扩充数据集。')
model.compile(optimizer=Adam(lr = lr), loss={'yolo_loss': lambda y_true, y_pred: y_pred})
train_dataloader = YoloDatasets(train_lines, input_shape, anchors, batch_size, num_classes, anchors_mask, mosaic = mosaic, train = True)
val_dataloader = YoloDatasets(val_lines, input_shape, anchors, batch_size, num_classes, anchors_mask, mosaic = False, train = False)
print('Train on {} samples, val on {} samples, with batch size {}.'.format(num_train, num_val, batch_size))
model.fit_generator(
generator = train_dataloader,
steps_per_epoch = epoch_step,
validation_data = val_dataloader,
validation_steps = epoch_step_val,
epochs = end_epoch,
initial_epoch = start_epoch,
use_multiprocessing = True if num_workers > 1 else False,
workers = num_workers,
callbacks = [logging, checkpoint, reduce_lr, early_stopping, loss_history])
本文章为转载内容，我们尊重原作者对文章享有的著作权。如有内容错误或侵权问题，欢迎原作者联系我们进行内容更正或删除文章。
