Mask-RCNN(TensorFlow版)的项目地址:
https://github.com/matterport/Mask_RCNN
目录
准备数据集
标定样本预处理
开始训练
准备数据集
首先将你的数据集分为两类，一类为训练集（train），一类为验证集(val)。
然后是标注数据，这里使用我另一篇博客提到的标记工具。
对样本进行标定。
标定样本预处理
将标定号的图片放到datasets文件夹内，分train,val两个文件夹存放训练及测试样本数据集。每个文件夹图片对应有一个同名的.json标注信息文件。
由于原始图片是各种尺寸，而mask-rcnn样本需要归一化为统一尺寸。这里参考inspect_data.ipynb文件的处理方式。重新建立了一个train_sample_convert.ipynb文件。
其作用是将datasets目录下train,val两个文件夹里的图片尺寸归一化，并将标定信息改变对应到新图像上，将新图像及标定信息存储在samples文件夹内。
文件内容如下：
import os
import sys
import random
import visualize
#import model as modellib
import utils
import json
import skimage.draw
import numpy as np
from model import log
from config import Config
# 获取该project的根目录
ROOT_DIR = os.path.abspath('')
sys.path.append(ROOT_DIR)
#原始数据datasets文件夹
ORIDATA_DIR = os.path.join(ROOT_DIR, "datasets")
#原始数据文件夹下要转换的目录列表
oridirs = ['train', 'val']
#归一化后图像文件夹
NORDATA_DIR = os.path.join(ROOT_DIR,'samples')
class CoalMineConfig(Config):
"""Configuration for training on the toy shapes dataset.
Derives from the base Config class and overrides values specific
to the toy shapes dataset.
"""
# Give the configuration a recognizable name
NAME = "CoalMine"
# Use small images for faster training. Set the limits of the small side
# the large side, and that determines the image shape.
IMAGE_MIN_DIM = 512
IMAGE_MAX_DIM = 512
config = CoalMineConfig()
#config.display()
def mkdir(path):
# 引入模块
import os
# 去除首位空格
path=path.strip()
# 去除尾部 \ 符号
path=path.rstrip("\\")
# 判断路径是否存在
# 存在 True
# 不存在 False
isExists=os.path.exists(path)
# 判断结果
if not isExists:
# 如果不存在则创建目录
# 创建目录操作函数
os.makedirs(path)
print (path+'创建成功')
return True
else:
# 如果目录存在则不创建，并提示目录已存在
print (path+'目录已存在')
return False
class DatasetChange(utils.Dataset):
"""Generates the shapes synthetic dataset. The dataset consists of simple
shapes (triangles, squares, circles) placed randomly on a blank surface.
The images are generated on the fly. No file access required.
"""
def change_sample(self, dataset_dir, subset,config):
"""定义数据集有哪些类别.
dataset_dir: 数据集根路径.
subset: train or val
"""
# 加载训练集还是测试集
assert subset in ["train", "val"]
dataset_dir = os.path.join(dataset_dir, subset)
print("原始文件夹:",dataset_dir)
newdate_dir = os.path.join(NORDATA_DIR,subset)
print("转换后文件夹:",newdate_dir)
mkdir(newdate_dir)
sublist = os.listdir(dataset_dir)
imagelist = [];
jsonlist = [];
for i in range(0, len(sublist)):
path = os.path.join(dataset_dir, sublist[i])
if os.path.isfile(path) and path.lower().endswith(".jpg"):
jsonname = path[:-3]+"json"
if os.path.isfile(jsonname):
imagelist.append(sublist[i]);
jsonlist.append(sublist[i][:-3]+"json");
print("转换",len(imagelist),"张图片")
#坐标变换函数
change = lambda array, rate, step : array*rate+step+0.5
for i in range(len(imagelist)):
fo = open(os.path.join(dataset_dir, jsonlist[i]),encoding='utf-8')
text = fo.read()
fo.close()
if text.startswith(u'\ufeff'):
text = text.encode('utf8')[3:].decode('utf8')
annotations = json.loads(text)
if not annotations['regions']:
print(imagelist[i], "no regions")
continue
image_path = os.path.join(dataset_dir, imagelist[i])
new_path = os.path.join(newdate_dir, imagelist[i])
new_json = os.path.join(newdate_dir, jsonlist[i])
#按限定尺寸改变图像大小 并保存
image = skimage.io.imread(image_path)
if image.ndim != 3:
image = skimage.color.gray2rgb(image)
image, window, scale, padding = utils.resize_image(image, min_dim=config.IMAGE_MIN_DIM,
max_dim=config.IMAGE_MAX_DIM,padding=config.IMAGE_PADDING)
skimage.io.imsave(new_path, image)
#转变坐标值
for i in range(len(annotations['regions']) ):
xarray = np.array(annotations['regions'][i]['shape_attributes']['all_points_x'],float)
xarray = np.array( change(xarray, scale, padding[1][0]), int)
annotations['regions'][i]['shape_attributes']['all_points_x'] = xarray.tolist()
yarray = np.array(annotations['regions'][i]['shape_attributes']['all_points_y'],float)
yarray = np.array( change(yarray, scale, padding[0][0]), int)
annotations['regions'][i]['shape_attributes']['all_points_y'] = yarray.tolist()
with open(new_json, "w") as f:
json.dump(annotations, f,separators=(',',':'),ensure_ascii=False)
#转换样本图片
for dir in oridirs:
dataset_ori = DatasetChange()
dataset_ori.change_sample(ORIDATA_DIR, dir,config)执行完上面文件，可以在samples目录下得到归一化转换完成的train，val两个样本文件夹。
开始训练
训练文件仿照train_shapes.ipynb，重新新建一个train_sample.ipynb
代码如下：
import os
import sys
import random
import visualize
import model as modellib
import utils
import json
import skimage.draw
import numpy as np
from model import log
from config import Config
# 获取该project的根目录
ROOT_DIR = os.path.abspath('')
# 存放训练模型的logs路径
MODEL_DIR = os.path.join(ROOT_DIR, "logs")
# 加载预训练的COCO权重
COCO_MODEL_PATH = os.path.join(ROOT_DIR, "mask_rcnn_coco.h5")
sys.path.append(ROOT_DIR)
#samples文件夹下是归一化的训练集和验证集
BALLOON_DIR = os.path.join(ROOT_DIR, "samples")
class NewDateSetConfig(Config):
"""Configuration for training on the toy shapes dataset.
Derives from the base Config class and overrides values specific
to the toy shapes dataset.
"""
# Give the configuration a recognizable name
NAME = "NewDateSet"
# TITAN X有12gb显存可以一次装两张，因此设置为2
GPU_COUNT = 1
IMAGES_PER_GPU = 1
# 类别的数目（包括背景）
NUM_CLASSES = 1 + 2
# Use small images for faster training. Set the limits of the small side
# the large side, and that determines the image shape.
IMAGE_MIN_DIM = 128
IMAGE_MAX_DIM = 128
# 每一个epoch迭代多少次
STEPS_PER_EPOCH = 500
# Use smaller anchors because our image and objects are small
RPN_ANCHOR_SCALES = (8, 16, 32, 64, 128) # anchor side in pixels
# Reduce training ROIs per image because the images are small and have
# few objects. Aim to allow ROI sampling to pick 33% positive ROIs.
TRAIN_ROIS_PER_IMAGE = 32
# use small validation steps since the epoch is small
VALIDATION_STEPS = 5
# 学习率和动量，不同的优化器有不同的配置，而且不同框架之间的实现也不一样，因此需要自己尝试
LEARNING_RATE = 0.001
LEARNING_MOMENTUM = 0.9
# 权重衰减因子
WEIGHT_DECAY = 0.0001
config = NewDateSetConfig()
#config.display()
class NewDataSetDataset(utils.Dataset):
"""Generates the shapes synthetic dataset. The dataset consists of simple
shapes (triangles, squares, circles) placed randomly on a blank surface.
The images are generated on the fly. No file access required.
"""
def load_newdataset(self, dataset_dir, subset):
"""定义数据集有哪些类别.
dataset_dir: 数据集根路径.
subset: train or val
"""
# 添加类别的名称和ID号，在这里，为了简便，我只添加一类：‘car’，如果是多类，挨着添加即可，例如我还有‘cat’,‘dog’....
self.add_class("newdataset", 1, "stone")
self.add_class("newdataset", 2, "ironbolt")
#self.add_class("newdataset", 3, "dog")
# 加载训练集还是测试集?
assert subset in ["train", "val"]
dataset_dir = os.path.join(dataset_dir, subset)
imagelist = [];
jsonlist = [];
sublist = os.listdir(dataset_dir)
for i in range(0, len(sublist)):
path = os.path.join(dataset_dir, sublist[i])
if os.path.isfile(path) and path.lower().endswith(".jpg"):
jsonname = path[:-3]+"json"
if os.path.isfile(jsonname)
imagelist.append(sublist[i]);
jsonlist.append(sublist[i][:-3]+"json");
for i in range(len(imagelist)):
fo = open(os.path.join(dataset_dir, jsonlist[i]),encoding='utf-8')
text = fo.read()
fo.close()
if text.startswith(u'\ufeff'):
text = text.encode('utf8')[3:].decode('utf8')
annotations = json.loads(text)
#print(annotations)
# 加载图片
# 获取组成每个物体实例轮廓的多边形点的x、y坐标。
# 这些坐标保存在r['shape_attributes'中，（参见上面的json格式）
#polygons = [r['shape_attributes'] for r in a['regions'].values()]
polygons = [r['shape_attributes'] for r in annotations['regions']]
region = [r['region_attributes'] for r in annotations['regions']]
image_path = os.path.join(dataset_dir, annotations['filename'])
image = skimage.io.imread(image_path)
height, width = image.shape[:2]
self.add_image(
"newdataset",
image_id=annotations['filename'],
path=image_path,
width=width, height=height,
polygons=polygons, region=region)
def load_mask(self, image_id):
"""为图像生成实例mask.
Returns:
masks: 一个bool数组，每个实例一个mask,
其中每个mask为：[高, 宽, 数目]
class_ids: 每个mask对应的类别.
"""
image_info = self.image_info[image_id]
if image_info["source"] != "newdataset":
return super(self.__class__, self).load_mask(image_id)
# 将多边形转换为一个二值mask,[高, 宽, 数目]
info = self.image_info[image_id]
# print("这是P",info["height"], info["width"], len(info["polygons"]))
mask = np.zeros([info["height"], info["width"], len(info["polygons"])],
dtype=np.uint8)
for i, p in enumerate(info["polygons"]):
# 获取多边形内的像素索引并将其设置为1
#rr, cc = skimage.draw.rectangle((p['y'], p['x']), extent=(p['height'], p['width']))
rr, cc = skimage.draw.polygon(p['all_points_y'], p['all_points_x'])
mask[rr, cc, i] = 1
region = image_info["region"]
class_ids = np.array([self.class_names.index(s["type"]) for s in region])
# 返回mask和类别的信息
return mask, class_ids.astype(np.int32)
def image_reference(self, image_id):
"""Return the shapes data of the image."""
info = self.image_info[image_id]
if info["source"] == "newdataset":
return info["newdataset"]
else:
super(self.__class__).image_reference(self, image_id)
#加载训练集
dataset_train = NewDataSetDataset()
dataset_train.load_newdataset(BALLOON_DIR, "train")
dataset_train.prepare()
#打印出训练集信息
print("Train Image Count: {}".format(len(dataset_train.image_ids)))
print("Class Count: {}".format(dataset_train.num_classes))
for i, info in enumerate(dataset_train.class_info):
print("{:3}. {:50}".format(i, info['name']))
#从训练集中选出4张用与显示训练集是否正确
image_ids = dataset_train.image_ids[:2]
for image_id in image_ids:
image = dataset_train.load_image(image_id)
mask, class_ids = dataset_train.load_mask(image_id)
visualize.display_top_masks(image, mask, class_ids, dataset_train.class_names)
#加载验证集
dataset_val = NewDataSetDataset()
dataset_val.load_newdataset(BALLOON_DIR, "val")
dataset_val.prepare()
#打印出验证集信息
print("Val Image Count: {}".format(len(dataset_val.image_ids)))
print("Class Count: {}".format(dataset_val.num_classes))
for i, info in enumerate(dataset_val.class_info):
print("{:3}. {:50}".format(i, info['name']))
#从验证集中选出4张用与显示验证集是否正确
image_ids = np.random.choice(dataset_val.image_ids, 2)
for image_id in image_ids:
image = dataset_val.load_image(image_id)
mask, class_ids = dataset_val.load_mask(image_id)
visualize.display_top_masks(image, mask, class_ids, dataset_train.class_names)
#创建一个训练模型
model = modellib.MaskRCNN(mode="training", config=config,
model_dir=MODEL_DIR)
#设定权重的初始化方式，有imagenet，coco,last三种
init_with = "coco"
if init_with == "imagenet":
model.load_weights(model.get_imagenet_weights(), by_name=True)
elif init_with == "coco":
model.load_weights(COCO_MODEL_PATH, by_name=True,
exclude=["mrcnn_class_logits", "mrcnn_bbox_fc",
"mrcnn_bbox", "mrcnn_mask"])
elif init_with == "last":
model.load_weights(model.find_last()[1], by_name=True)
#设定训练参数，如学习率，epoch等
model.train(dataset_train, dataset_val, learning_rate=config.LEARNING_RATE, epochs=500, layers="heads")
# Save weights
# Typically not needed because callbacks save after every epoch
# Uncomment to save manually
model_path = os.path.join(MODEL_DIR, "mask_rcnn_cocoNew.h5")
model.keras_model.save_weights(model_path)等训练完成后，会在模型文件夹保存mask_rcnn_cocoNew.h5权重文件。
本文代码在ubuntu下运行验证，windows下运行可能遇到标定信息文件读取编码问题，可尝试修改文件encoding参数。
运行中出现什么问题，欢迎留言探讨！
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。
