搭建SSD框架,下载解压即可
下载pascalvoc数据,自己的数据根据voc格式改写(图片的名称,不用拘泥于6位数字,其他命名也可以)数据集下载点击
解压后不要混合在一个文件夹下
VOCtrainval用来训练,VOCtest用来测试。
VOCtrainval 中JPEGImage文件夹中仅是训练和验证的图片,Main文件夹中仅是trainval.txt, train.txt, val.txt
VOCtest中JPEGImage文件夹中仅是测试图片,Main文件夹中仅是test.txt
自己的文件根据以上文件格式放置图片即可。
自己的数据根据voc格式改写(图片的名称,不用拘泥于6位数字,其他命名也可以)
文件重命名点击
标记自己的数据 ,这个过程枯燥,需要耐心。详情请点击,
生成txt文件,train.txt, trainval.txt, test.txt, val.txt(注意文件路径)
import os
import random
saveBasePath = r"./VOC2007/ImageSets" # txt文件保存目录
total_xml = os.listdir(r'./VOC2007/Annotations') # 获取标注文件(file_name.xml)
# 划分数据集为(训练,验证,测试集 = 49%,20%,30%)
trainval_percent = 0.7
train_percent = 0.7
tv = int(len(total_xml) * trainval_percent) # 70%训练-验证集的文件数目
tr = int(tv * train_percent) # 70%训练集的文件数目
# 打乱训练文件(洗牌)
trainval = random.sample(range(len(total_xml)), tv)
train = random.sample(trainval, tr)
print("train and val size", tv)
print("train size", tr)
ftrainval = open(os.path.join(saveBasePath, 'Main/trainval.txt'), 'w')
ftest = open(os.path.join(saveBasePath, 'Main/test.txt'), 'w')
ftrain = open(os.path.join(saveBasePath, 'Main/train.txt'), 'w')
fval = open(os.path.join(saveBasePath, 'Main/val.txt'), 'w')
for i in range(len(total_xml)): # 遍历所有 file_name.xml 文件
name = total_xml[i][:-4] + '\n' # 获取 file_name
if i in trainval:
ftrainval.write(name)
if i in train:
ftrain.write(name)
else:
fval.write(name)
else:
ftest.write(name)
ftrainval.close()
ftrain.close()
fval.close()
ftest.close()
VOCtrainval_06-Nov-2007\VOCdevkit\VOC2007\ImageSets\Main\
VOCtest_06-Nov-2007\VOCdevkit\VOC2007\ImageSets\Main\
SSD-Tensorflow/datasets/pascalvoc_common.py
# 注释原始的标签,添加自己的标签
VOC_LABELS = {
'none': (0, 'Background'),
'aeroplane': (1, 'Vehicle'),
'bicycle': (2, 'Vehicle'),
'bird': (3, 'Animal'),
'boat': (4, 'Vehicle'),
... ...
'Person': (15, 'Person'),
'pottedplant': (16, 'Indoor'),
'sheep': (17, 'Animal'),
'sofa': (18, 'Indoor'),
'train': (19, 'Vehicle'),
'tvmonitor': (20, 'Indoor'),
}
SSD-Tensorflow/datasets/pascalvoc_to_tfrecords.py
。image_data = tf.gfile.FastGFile(filename, 'rb').read()
;.jpg
格式,修改图片类型;SAMPLES_PER_FILES = 500(自定义)
意为:几个.xml转为一个tfrecords,如下图tf_convert_data.py
文件,依次点击:run、Edit Configuration
,在Parameters
中填入以下内容,再运行tf_convert_data.py
文件,在面板中得到成功信息,可以在tfrecords_文件夹下看到生成的.tfrecords文件;--dataset_name=pascalvoc
--dataset_dir=./VOC2007/
--output_name=voc_2007_train
--output_dir=./tfrecords_
或者在SSD-Tensorflow 文件夹下创建tf_conver_data.sh
运行。
#!/bin/bash
# 这是一个shell脚本,用于将pascal VOC数据集转换tfrecords数据
DATASET_DIR=./VOC2007/ # VOC数据保存的文件夹(VOC的目录格式未改变)
OUTPUT_DIR=./tfrecords_ # 保存tfrecords数据的文件夹
python ./tf_convert_data.py\
--dataset_name=pascalvoc\
--dataset_dir=${DATASET_DIR}\
--output_name=voc_2007_train\
--output_dir=${OUTPUT_DIR}
或者直接使用如下代码
"""
特别注意: path地址是否正确、要在主目录下提前创建“tfrecords_”文件夹
"""
import os
import sys
import random
import numpy as np
import tensorflow as tf
import xml.etree.ElementTree as ET # 操作xml文件
# 我的标签定义只有两类,要根据自己的图片而定
VOC_LABELS = {
'none': (0, 'Background'),
'aiaitie': (1, 'Product')
}
# 图片和标签存放的文件夹.
DIRECTORY_ANNOTATIONS = 'Annotations/'
DIRECTORY_IMAGES = 'JPEGImages/'
RANDOM_SEED = 4242 # 随机种子.
SAMPLES_PER_FILES = 3 # 每个.tfrecords文件包含几个.xml样本
# 生成整数型,浮点型和字符串型的属性
def int64_feature(value):
if not isinstance(value, list):
value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))
def float_feature(value):
if not isinstance(value, list):
value = [value]
return tf.train.Feature(float_list=tf.train.FloatList(value=value))
def bytes_feature(value):
if not isinstance(value, list):
value = [value]
return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))
# 图片处理
def _process_image(directory, name):
# Read the image file.
filename = directory + DIRECTORY_IMAGES + name + '.jpg'
image_data = tf.gfile.FastGFile(filename, 'rb').read()
# Read the XML annotation file.
filename = os.path.join(directory, DIRECTORY_ANNOTATIONS, name + '.xml')
tree = ET.parse(filename)
root = tree.getroot()
# Image shape.
size = root.find('size')
shape = [int(size.find('height').text),
int(size.find('width').text),
int(size.find('depth').text)]
# Find annotations.
bboxes = []
labels = []
labels_text = []
difficult = []
truncated = []
for obj in root.findall('object'):
label = obj.find('name').text
labels.append(int(VOC_LABELS[label][0]))
labels_text.append(label.encode('ascii')) # 变为ascii格式
if obj.find('difficult'):
difficult.append(int(obj.find('difficult').text))
else:
difficult.append(0)
if obj.find('truncated'):
truncated.append(int(obj.find('truncated').text))
else:
truncated.append(0)
bbox = obj.find('bndbox')
a = float(bbox.find('ymin').text) / shape[0]
b = float(bbox.find('xmin').text) / shape[1]
a1 = float(bbox.find('ymax').text) / shape[0]
b1 = float(bbox.find('xmax').text) / shape[1]
a_e = a1 - a
b_e = b1 - b
if abs(a_e) < 1 and abs(b_e) < 1:
bboxes.append((a, b, a1, b1))
return image_data, shape, bboxes, labels, labels_text, difficult, truncated
# 转化样例
def _convert_to_example(image_data, labels, labels_text, bboxes, shape,
difficult, truncated):
xmin = []
ymin = []
xmax = []
ymax = []
for b in bboxes:
assert len(b) == 4
# pylint: disable=expression-not-assigned
[l.append(point) for l, point in zip([ymin, xmin, ymax, xmax], b)]
# pylint: enable=expression-not-assigned
image_format = b'JPEG'
example = tf.train.Example(features=tf.train.Features(feature={
'image/height': int64_feature(shape[0]),
'image/width': int64_feature(shape[1]),
'image/channels': int64_feature(shape[2]),
'image/shape': int64_feature(shape),
'image/object/bbox/xmin': float_feature(xmin),
'image/object/bbox/xmax': float_feature(xmax),
'image/object/bbox/ymin': float_feature(ymin),
'image/object/bbox/ymax': float_feature(ymax),
'image/object/bbox/label': int64_feature(labels),
'image/object/bbox/label_text': bytes_feature(labels_text),
'image/object/bbox/difficult': int64_feature(difficult),
'image/object/bbox/truncated': int64_feature(truncated),
'image/format': bytes_feature(image_format),
'image/encoded': bytes_feature(image_data)}))
return example
# 增加到tfrecord
def _add_to_tfrecord(dataset_dir, name, tfrecord_writer):
image_data, shape, bboxes, labels, labels_text, difficult, truncated = \
_process_image(dataset_dir, name)
example = _convert_to_example(image_data, labels, labels_text,
bboxes, shape, difficult, truncated)
tfrecord_writer.write(example.SerializeToString())
# name为转化文件的前缀
def _get_output_filename(output_dir, name, idx):
return '%s/%s_%03d.tfrecord' % (output_dir, name, idx)
def run(dataset_dir, output_dir, name='voc_train', shuffling=False):
if not tf.gfile.Exists(dataset_dir):
tf.gfile.MakeDirs(dataset_dir)
path = os.path.join(dataset_dir, DIRECTORY_ANNOTATIONS)
filenames = sorted(os.listdir(path)) # 排序
if shuffling:
random.seed(RANDOM_SEED)
random.shuffle(filenames)
i = 0
fidx = 0
while i < len(filenames): # Open new TFRecord file.
tf_filename = _get_output_filename(output_dir, name, fidx)
with tf.python_io.TFRecordWriter(tf_filename) as tfrecord_writer:
j = 0
while i < len(filenames) and j < SAMPLES_PER_FILES:
sys.stdout.write(' Converting image %d/%d \n' % (i + 1, len(filenames))) # 终端打印,类似print
sys.stdout.flush() # 缓冲
filename = filenames[i]
img_name = filename[:-4]
_add_to_tfrecord(dataset_dir, img_name, tfrecord_writer)
i += 1
j += 1
fidx += 1
print('\nFinished converting the Pascal VOC dataset!')
# 原数据集路径,输出路径以及输出文件名,要根据自己实际做改动
dataset_dir = "C:/Users/Admin/Desktop/"
output_dir = "./tfrecords_"
name = "voc_train"
def main(_):
run(dataset_dir, output_dir, name)
if __name__ == '__main__':
tf.app.run()
datasets/pascalvoc_2007.py
修改训练数据shape:NUM_CLASSES = 类别数
;TRAIN_STATISTICS = {
'none': (0, 0),
'aeroplane': (238, 306), #238图片数, 306目标总数
'bicycle': (243, 353),
'bird': (330, 486),
'boat': (181, 290),
... ...
'sheep': (96, 257),
'sofa': (229, 248),
'train': (261, 297),
'tvmonitor': (256, 324),
'total': (5011, 12608), #5011 为训练的图片书,12608为目标总数
}
TEST_STATISTICS = {
'none': (0, 0),
'aeroplane': (1, 1),
'bicycle': (1, 1),
'bird': (1, 1),
... ...
'sheep': (1, 1),
'sofa': (1, 1),
'train': (1, 1),
'tvmonitor': (1, 1),
'total': (20, 20),
}
SPLITS_TO_SIZES = {
'train': 5011, # 训练数据量
'test': 4952, # 测试数据量
}
SPLITS_TO_STATISTICS = {
'train': TRAIN_STATISTICS,
'test': TEST_STATISTICS,
}
NUM_CLASSES = 20 # 类别,根据自己数据的实际类别修改(不包含背景)
nets/ssd_vgg_300.py
修改类别个数,根据自己训练类别数修改96 和97行:等于类别数+1; img_shape=(300, 300),
num_classes=21, #根据自己的数据修改为类别+1
no_annotation_label=21, #根据自己的数据修改为类别+1
eval_ssd_network.py
修改类别个数,修改66行的类别个数:等于类别数+1;tf.app.flags.DEFINE_integer(
'num_classes', 21, 'Number of classes to use in the dataset.')
train_ssd_network.py
tf.app.flags.DEFINE_integer(
'log_every_n_steps', 10,
'The frequency with which logs are print.')
tf.app.flags.DEFINE_integer(
'save_summaries_secs', 600,
'The frequency with which summaries are saved, in seconds.')
tf.app.flags.DEFINE_integer(
'save_interval_secs', 600,
'The frequency with which the model is saved, in seconds.')
tf.app.flags.DEFINE_float(
'gpu_memory_fraction', 0.9, 'GPU memory fraction to use.')
方案1 从vgg开始训练其中某些层的参数:
# 通过加载预训练好的vgg16模型,进行训练
# 通过 checkpoint_exclude_scopes 指定哪些层的参数不需要从vgg16模型里面加载进来
# 通过 trainable_scopes 指定哪些层的参数是需要训练的,未指定的参数保持不变,若注释掉此命令,所有的参数均需要训练
DATASET_DIR=./tfrecords_/ # 数据存放路径
TRAIN_DIR=./train_model/ # 训练生成模型的存放路径
CHECKPOINT_PATH=./checkpoints/vgg_16.ckpt # 加载预训练模型的路径
python ../train_ssd_network.py \
--train_dir=${TRAIN_DIR} \ # 训练生成模型的存放路径
--dataset_dir=${DATASET_DIR} \ # 数据存放路径
--dataset_name=pascalvoc_2007 \ # 数据名的前缀
--dataset_split_name=train \
--model_name=ssd_300_vgg \ # 加载的模型的名字
--checkpoint_path=${CHECKPOINT_PATH} \ # 所加载模型的路径
--checkpoint_model_scope=vgg_16 \ # 所加载模型里面的作用域名
--checkpoint_exclude_scopes=ssd_300_vgg/conv6,ssd_300_vgg/conv7,ssd_300_vgg/block8,ssd_300_vgg/block9,ssd_300_vgg/block10,ssd_300_vgg/block11,ssd_300_vgg/block4_box,ssd_300_vgg/block7_box,ssd_300_vgg/block8_box,ssd_300_vgg/block9_box,ssd_300_vgg/block10_box,ssd_300_vgg/block11_box \
--trainable_scopes=ssd_300_vgg/conv6,ssd_300_vgg/conv7,ssd_300_vgg/block8,ssd_300_vgg/block9,ssd_300_vgg/block10,ssd_300_vgg/block11,ssd_300_vgg/block4_box,ssd_300_vgg/block7_box,ssd_300_vgg/block8_box,ssd_300_vgg/block9_box,ssd_300_vgg/block10_box,ssd_300_vgg/block11_box \
--save_summaries_secs=60 \ # 每60s保存一下日志
--save_interval_secs=600 \ # 每600s保存一下模型
--weight_decay=0.0005 \ # 正则化的权值衰减的系数
--optimizer=adam \ # 选取的最优化函数
--learning_rate=0.001 \ # 学习率
--learning_rate_decay_factor=0.94 \ # 学习率的衰减因子
--batch_size=24 \ # 可以小一点,不然可能会报错(显存不够用)
--gpu_memory_fraction=0.9 # 指定占用gpu内存的百分比
方案2:从头开始训练自己的模型
#注释掉如下参数:
#CHECKPOINT_PATH=./checkpoints/vgg_16.ckpt 不提供初始化模型,让模型自己随机初始化权重,从头训练
#--checkpoint_path=${CHECKPOINT_PATH}
#--checkpoint_path=${CHECKPOINT_PATH}
#--checkpoint_model_scope=ssd_512_vgg
#--checkpoint_exclude_scopes=ssd_300_vgg/block10...
#--trainable_scopes=ssd_300_vgg/conv6...
#/bin/bash
DATASET_DIR=./tfrecords_/ # 数据存放路径
TRAIN_DIR=./train_model/ # 训练生成模型的存放路径
CUDA_VISIBLE_DEVICES=0 python ./train_ssd_network.py \
--train_dir=${TRAIN_DIR} \
--dataset_dir=${DATASET_DIR} \
--dataset_name=pascalvoc_2007 \
--dataset_split_name=train \
--model_name=ssd_300_vgg \
--save_summaries_secs=600 \
--save_interval_secs=600 \
--optimizer=adam \
--learning_rate_decay_factor=0.94 \
--batch_size=32 \
--gpu_memory_fraction=0.9
.tfrecords
文件。将测试图片转换为tfrecords#!/bin/bash
DATASET_DIR=./VOC2007/test_images/ # 测试图片目录(存放测试的图片)
OUTPUT_DIR=./tfrecords_/tfrecords/ # 测试图片的 .tfrecords文件
python ./tf_convert_data.py \
--dataset_name=pascalvoc \
--dataset_dir=${DATASET_DIR} \
--output_name=voc_2007_test \
--output_dir=${OUTPUT_DIR}
#!/bin/bash
DATASET_DIR=./tfrecords_/tfrecords/
EVAL_DIR=./ssd_eval_log/
CHECKPOINT_PATH=./train_model/model.ckpt-5000
python ./eval_ssd_network.py \
--eval_dir=${EVAL_DIR} \
--dataset_dir=${DATASET_DIR} \
--dataset_name=pascalvoc_2007 \
--dataset_split_name=test \
--model_name=ssd_300_vgg \
--checkpoint_path=${CHECKPOINT_PATH} \
--batch_size=1
notebooksssd_notebook.ipynb
来查看模型标注的图片。详情请点击ckpt_filename = "路径/自己训练的权重文件"
ZeroDivisionError: float division by zero
,详情如下:>> Converting image 117/504Traceback (most recent call last): #第117张标注文件有问题
File "D:/AI_target_detection/SSD-Tensorflow/tf_convert_data.py", line 59, in <module>
tf.app.run()
File "C:\Anaconda3\envs\AI_tensorflow_GPU\lib\site-packages\tensorflow\python\platform\app.py", line 125, in run
_sys.exit(main(argv))
File "D:/AI_target_detection/SSD-Tensorflow/tf_convert_data.py", line 54, in main
pascalvoc_to_tfrecords.run(FLAGS.dataset_dir, FLAGS.output_dir, FLAGS.output_name)
File "D:\AI_target_detection\SSD-Tensorflow\datasets\pascalvoc_to_tfrecords.py", line 223, in run
_add_to_tfrecord(dataset_dir, img_name, tfrecord_writer)
File "D:\AI_target_detection\SSD-Tensorflow\datasets\pascalvoc_to_tfrecords.py", line 182, in _add_to_tfrecord
_process_image(dataset_dir, name)
File "D:\AI_target_detection\SSD-Tensorflow\datasets\pascalvoc_to_tfrecords.py", line 121, in _process_image
bboxes.append((max(float(bbox.find('ymin').text) / shape[0], 0.1),
ZeroDivisionError: float division by zero
All bounding box coordinates must be in [0.0, 1.0]
原因及解决方法:标注数据集时鼠标多点了一下,没有任何标注,和标注框超出图片范围。
pascalvoc_to_tfrecords.py
114-119行将:
bboxes.append((float(bbox.find('ymin').text) / shape[0],
float(bbox.find('xmin').text) / shape[1],
float(bbox.find('ymax').text) / shape[0],
float(bbox.find('xmax').text) / shape[1]
))
修改为:
bboxes.append((max(float(bbox.find('ymin').text) / shape[0], 0.0),
max(float(bbox.find('xmin').text) / shape[1], 0.0),
min(float(bbox.find('ymax').text) / shape[0], 1.0),
min(float(bbox.find('xmax').text) / shape[1], 1.0)
))