语义分割 HybridTaskCascade 跑通DEMO 训练自己的数据集

火腿肠项目地址　https://github.com/qixuxiang/Hybrid_Task_Cascade

搭建环境

• Linux (tested on Ubuntu 16.04 and CentOS 7.2)
• Python 3.4+
• PyTorch 1.0
• Cython
• mmcv >= 0.2.2

首先把Pytorch1.0装好

编译

下载项目文件 确保777权限　不然会有编译失败

./compile.sh   

python setup.py develop

或者

python setup.py install

测试

mydemo.py

import mmcv
from mmcv.runner import load_checkpoint
from mmdet.models import build_detector
from mmdet.apis import inference_detector, show_result

cfg = mmcv.Config.fromfile('/media/yang/56c81da5-109b-42e1-9e51-0f9e1396fa9d/yang/Hybrid_Task_Cascade-master/configs/cascade_mask_rcnn_x101_64x4d_fpn_1x.py')
cfg.model.pretrained = None

# construct the model and load checkpoint
model = build_detector(cfg.model, test_cfg=cfg.test_cfg)
_ = load_checkpoint(model, 'checkpoints/cascade_mask_rcnn_x101_64x4d_fpn_20e_20181218-630773a7.pth')

# test a single image
img = mmcv.imread('demo/33823288584_1d21cf0a26_k.jpg')
result = inference_detector(model, img, cfg)
show_result(img, result)
'''
# test a list of images
imgs = ['test1.jpg', 'test2.jpg']
for i, result in enumerate(inference_detector(model, imgs, cfg, device='cuda:0')):
    print(i, imgs[i])
    show_result(imgs[i], result)
'''

权重下载地址　https://github.com/qixuxiang/Hybrid_Task_Cascade/blob/master/MODEL_ZOO.md　

作者试验了很多模型　什么都有　看自己情况下载

作者把权重放在了亚马逊云　下载起来很慢　要挂代理

５００Ｍ大概半小时左右

主要改３个地方　cfg路径　model路径　图片路径

python mydemo.py

训练自己的数据集

我是用coco-annotator标注的

格式和coco2014一样

但是这个模型需要coco2017格式的数据集

也就是多了一组图片　stuffmaps这个东西　

下面提供转换工具

https://github.com/nightrome/cocostuffapi　下载这个东西

2、进入PythonAPI/pycocotools/cocostuffhelper.py，改动：

(1) Line141:

from:

    labelMap = cocoSegmentationToSegmentationMap(coco, imgId, includeCrowd=includeCrowd)
to:

    labelMap = cocoSegmentationToSegmentationMap(coco, imgId, checkUniquePixelLabel=False, includeCrowd=includeCrowd)
(2) Line142 后面加上:

    labelMap = labelMap + 91
（3）进入 PythonAPI/ ：

make
4、使用 PythonAPI/cocostuff/cocoSegmentationToPngDemo.py ，设置好 annPath 为你自己的 coco/annotations/instances_train2014.json 路径；然后把 Line75-Line88 注释掉

４.5

中途报错：ValueError: invalid palette size，
解决方案：
引用cmap的时候改成 np.uint8(cmap).tolist()即可～

这句话加在cocostuffapi-master/PythonAPI/pycocotools/cocostuffhelper.py

大约153行　assert len(cmap) == 768, 'Error: Color map must have exactly 256*3 elements!'　前面

5、把cocostuff/cocoSegmentationToPngDemo.py 放到cocostuff这个文件夹外面

修改成自己的路径　下面是我的

#!/usr/bin/python

__author__ = 'hcaesar'

# Converts COCO segmentation .json files (GT or results) to one .png file per image.
#
# This script can be used for visualization of ground-truth and result files.
# Furthermore it can convert the ground-truth annotations to a more easily
# accessible .png format that is supported by many semantic segmentation methods.
#
# Note: To convert a result file to .png, we need to have both a valid GT file
# and the result file and set isAnnotation=False.
#
# The .png images are stored as indexed images, which means they contain both the
# segmentation map, as well as a color palette for visualization.
#
# Microsoft COCO Toolbox.      version 2.0
# Data, paper, and tutorials available at:  http://mscoco.org/
# Code written by Piotr Dollar and Tsung-Yi Lin, 2015.
# Licensed under the Simplified BSD License [see coco/license.txt]

import os
from pycocotools import mask
from pycocotools.cocostuffhelper import cocoSegmentationToPng
from pycocotools.coco import COCO
import skimage.io
import matplotlib.pyplot as plt

def cocoSegmentationToPngDemo(dataDir='coco/', dataTypeAnn='train2014', dataTypeRes='examples', \
        pngFolderName='export_png', isAnnotation=True, exportImageLimit=10000):
    '''
    Converts COCO segmentation .json files (GT or results) to one .png file per image.
    :param dataDir: location of the COCO root folder
    :param dataTypeAnn: identifier of the ground-truth annotation file
    :param dataTypeRes: identifier of the result annotation file (if any)
    :param pngFolderName: the name of the subfolder where we store .png images
    :param isAnnotation: whether the COCO file is a GT annotation or a result file
    :return: None
    '''

    # Define paths
    annPath = '%s/annotations/instances_%s.json' % (dataDir, dataTypeAnn)
    if isAnnotation:
        pngFolder = '%s/annotationss/%s' % (dataDir, dataTypeAnn)
    else:
        pngFolder = '%s/results/%s' % (dataDir, pngFolderName)
        resPath = '%s/results/stuff_%s_results.json' % (dataDir, dataTypeRes)

    # Create output folder
    if not os.path.exists(pngFolder):
        os.makedirs(pngFolder)

    # Initialize COCO ground-truth API
    coco = COCO(annPath)
    imgIds = coco.getImgIds()

    # Initialize COCO result
    if not isAnnotation:
        coco = coco.loadRes(resPath)
        imgIds = sorted(set([a['image_id'] for a in coco.anns.values()]))

    # Limit number of images
    if exportImageLimit < len(imgIds):
        imgIds = imgIds[0:exportImageLimit]

    # Convert each image to a png
    imgCount = len(imgIds)
    for i in range(0, imgCount):
        imgId = imgIds[i]
        imgName = coco.loadImgs(ids=imgId)[0]['file_name'].replace('.jpg', '')
        print('Exporting image %d of %d: %s' % (i+1, imgCount, imgName))
        segmentationPath = '%s/%s.png' % (pngFolder, imgName)
        cocoSegmentationToPng(coco, imgId, segmentationPath)

    # # Visualize the last image
    # originalImage = skimage.io.imread(coco.loadImgs(imgId)[0]['coco_url'])
    # segmentationImage = skimage.io.imread(segmentationPath)
    # plt.figure()
    # plt.subplot(121)
    # plt.imshow(originalImage)
    # plt.axis('off')
    # plt.title('original image')

    # plt.subplot(122)
    # plt.imshow(segmentationImage)
    # plt.axis('off')
    # plt.title('annotated image')
    # plt.show()

if __name__ == "__main__":
    cocoSegmentationToPngDemo()

然后执行

坑爹的是这个图像是彩色的　必须是单通道的灰度图才行

转换工具

import cv2
import numpy as np
import os    

IMAGE_DIR= '/media/yang/56c81da5-109b-42e1-9e51-0f9e1396fa9d/yang/Hybrid_Task_Cascade-master/data/coco/stuffthingmaps/val2014/'
out = '/media/yang/56c81da5-109b-42e1-9e51-0f9e1396fa9d/yang/Hybrid_Task_Cascade-master/data/coco/stuffthingmaps/out/'

file_names = next(os.walk(IMAGE_DIR))[2]
for x in range(len(file_names)):
    img = cv2.imread(os.path.join(IMAGE_DIR, file_names[x]))
    #img = cv2.imread(img_path)
    #获取图片的宽和高
    #width,height = img.shape[:2][::-1]
    # #将图片缩小便于显示观看
    # img_resize = cv2.resize(img,
    # (int(width*0.5),int(height*0.5)),interpolation=cv2.INTER_CUBIC)
    # cv2.imshow("img",img_resize)
    # print("img_reisze shape:{}".format(np.shape(img_resize)))

    #将图片转为灰度图
    img_gray = cv2.cvtColor(img,cv2.COLOR_RGB2GRAY)
    cv2.imwrite(out+str(file_names[x]),img_gray)

    #cv2.waitKey()

把这个图像包合并到之前的数据集

yang@yang:/media/yang/56c81da5-109b-42e1-9e51-0f9e1396fa9d/yang/maskrcnn/Mask_RCNN-master/datasets/coco$ tree
.
├── annotations
│   ├── instances_train2014.json
│   ├── instances_val2014.json
├── stuffthingmaps
│   ├── train2014
│   │   ├── 1000.png
...
│   └── val2014
│       ├── 10.png
...
└── val2014
    ├── 10.jpg
...
    └── 2.jpg
这个样子的

最后一步　训练

需要修改的地方有

Hybrid_Task_Cascade-master/configs/htc/htc_x101_64x4d_fpn_20e_16gpu.py这个文件中　因为我做语义分割　所有要用这个模型

第5行　填写预训练模型地址　我知道你没有　第一次运行会自动下载

第４４／５４／６４／７９行设为你自己的分类数

第１９１行设为你自己的图像尺寸

第２２４行记着修改学习率　不然会梯度爆炸

第２４３行　修改训练的Epoch数　默认的就没问题

第２４６行　存储log和模型的路径　有需要自己改　我没变

第２３９行取消注释　就会保存tfbroad文件

tensorboard --logdir event文件的上一级路径 要等一会才会显示

python tools/train.py --config configs/htc/htc_x101_64x4d_fpn_20e_16gpu.py 就可以了

意外中断　恢复训练

python tools/train.py --config configs/htc/htc_x101_64x4d_fpn_20e_16gpu.py --resume_from work_dirs/htc_x101_64x4d_fpn_20e/epoch_16.pth