MaskRCNN-Benchmark(Pytorch版本)训练自己的数据以及避坑指南

一、安装

首先要阅读官网说明的环境要求，千万不要一股脑直接安装，不然后面程序很有可能会报错！！！

PyTorch 1.0 from a nightly release. It will not work with 1.0 nor 1.0.1. Installation instructions can be found in https://pytorch.org/get-started/locally/

torchvision from master

cocoapi

yacs

matplotlib

GCC >= 4.9

OpenCV

# first, make sure that your conda is setup properly with the right environment
# for that, check that `which conda`, `which pip` and `which python` points to the
# right path. From a clean conda env, this is what you need to do

conda create --name maskrcnn_benchmark
conda activate maskrcnn_benchmark

# this installs the right pip and dependencies for the fresh python
conda install ipython

# maskrcnn_benchmark and coco api dependencies
pip install ninja yacs cython matplotlib tqdm opencv-python

# follow PyTorch installation in https://pytorch.org/get-started/locally/
# we give the instructions for CUDA 9.0
conda install -c pytorch pytorch-nightly torchvision cudatoolkit=9.0

export INSTALL_DIR=$PWD

# install pycocotools
cd $INSTALL_DIR
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
python setup.py build_ext install

# install apex
cd $INSTALL_DIR
git clone https://github.com/NVIDIA/apex.git
cd apex
python setup.py install --cuda_ext --cpp_ext

# install PyTorch Detection
cd $INSTALL_DIR
git clone https://github.com/facebookresearch/maskrcnn-benchmark.git
cd maskrcnn-benchmark

# the following will install the lib with
# symbolic links, so that you can modify
# the files if you want and won't need to
# re-build it
python setup.py build develop


unset INSTALL_DIR

# or if you are on macOS
# MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py build develop

一定要按上面的说明一步一步来，千万别省略，不然后面程序很有可能会报错！！！

二、数据准备

我要制作的原始数据格式是训练文件在一个文件（train），标注文件是csv格式，内容如下：

image

第一步，先把全部有标记的图片且分为训练集，验证集，分别存储在两个文件夹中，代码如下：

#!/usr/bin/env python
# coding=UTF-8
'''
@Description: 
@Author: HuangQinJian
@LastEditors: HuangQinJian
@Date: 2019-05-01 12:56:08
@LastEditTime: 2019-05-01 13:11:38
'''
import pandas as pd
import random
import os
import shutil

if not os.path.exists('trained/'):
    os.mkdir('trained/')

if not os.path.exists('val/'):
    os.mkdir('val/')

val_rate = 0.15

img_path = 'train/'
img_list = os.listdir(img_path)
train = pd.read_csv('train_label_fix.csv')
# print(img_list)
random.shuffle(img_list)

total_num = len(img_list)
val_num = int(total_num*val_rate)
train_num = total_num-val_num

for i in range(train_num):
    img_name = img_list[i]
    shutil.copy('train/' + img_name, 'trained/' + img_name)
for j in range(val_num):
    img_name = img_list[j+train_num]
    shutil.copy('train/' + img_name, 'val/' + img_name)

第二步，把csv格式的标注文件转换成coco的格式，代码如下：

#!/usr/bin/env python
# coding=UTF-8
'''
@Description: 
@Author: HuangQinJian
@LastEditors: HuangQinJian
@Date: 2019-04-23 11:28:23
@LastEditTime: 2019-05-01 13:15:57
'''
import sys
import os
import json
import cv2
import pandas as pd

START_BOUNDING_BOX_ID = 1
PRE_DEFINE_CATEGORIES = {}


def convert(csv_path, img_path, json_file):
    """
    csv_path : csv文件的路径
    img_path : 存放图片的文件夹
    json_file : 保存生成的json文件路径
    """
    json_dict = {"images": [], "type": "instances", "annotations": [],
                 "categories": []}
    bnd_id = START_BOUNDING_BOX_ID
    categories = PRE_DEFINE_CATEGORIES
    csv = pd.read_csv(csv_path)
    img_nameList = os.listdir(img_path)
    img_num = len(img_nameList)
    print("图片总数为{0}".format(img_num))
    for i in range(img_num):
        # for i in range(30):
        image_id = i+1
        img_name = img_nameList[i]
        if img_name == '60f3ea2534804c9b806e7d5ae1e229cf.jpg' or img_name == '6b292bacb2024d9b9f2d0620f489b1e4.jpg':
            continue
        # 可能需要根据具体格式修改的地方
        lines = csv[csv.filename == img_name]
        img = cv2.imread(os.path.join(img_path, img_name))
        height, width, _ = img.shape
        image = {'file_name': img_name, 'height': height, 'width': width,
                 'id': image_id}
        print(image)
        json_dict['images'].append(image)
        for j in range(len(lines)):
            # 可能需要根据具体格式修改的地方
            category = str(lines.iloc[j]['type'])
            if category not in categories:
                new_id = len(categories)
                categories[category] = new_id
            category_id = categories[category]
            # 可能需要根据具体格式修改的地方
            xmin = int(lines.iloc[j]['X1'])
            ymin = int(lines.iloc[j]['Y1'])
            xmax = int(lines.iloc[j]['X3'])
            ymax = int(lines.iloc[j]['Y3'])
            # print(xmin, ymin, xmax, ymax)
            assert(xmax > xmin)
            assert(ymax > ymin)
            o_width = abs(xmax - xmin)
            o_height = abs(ymax - ymin)
            ann = {'area': o_width*o_height, 'iscrowd': 0, 'image_id':
                   image_id, 'bbox': [xmin, ymin, o_width, o_height],
                   'category_id': category_id, 'id': bnd_id, 'ignore': 0,
                   'segmentation': []}
            json_dict['annotations'].append(ann)
            bnd_id = bnd_id + 1
    for cate, cid in categories.items():
        cat = {'supercategory': 'none', 'id': cid, 'name': cate}
        json_dict['categories'].append(cat)

    json_fp = open(json_file, 'w')
    json_str = json.dumps(json_dict, indent=4)
    json_fp.write(json_str)
    json_fp.close()


if __name__ == '__main__':
    # csv_path = 'data/train_label_fix.csv'
    # img_path = 'data/train/'
    # json_file = 'train.json'
    csv_path = 'train_label_fix.csv'
    img_path = 'trained/'
    json_file = 'trained.json'
    convert(csv_path, img_path, json_file)
    csv_path = 'train_label_fix.csv'
    img_path = 'val/'
    json_file = 'val.json'
    convert(csv_path, img_path, json_file)

第三步，可视化转换后的coco的格式，以确保转换正确，代码如下：

（注意：在这一步中，需要先下载 cocoapi , 可能出现的问题）

#!/usr/bin/env python
# coding=UTF-8
'''
@Description: 
@Author: HuangQinJian
@LastEditors: HuangQinJian
@Date: 2019-04-23 13:43:24
@LastEditTime: 2019-04-30 21:29:26
'''
from pycocotools.coco import COCO
import skimage.io as io
import matplotlib.pyplot as plt
import pylab
import cv2
import os
from skimage.io import imsave
import numpy as np
pylab.rcParams['figure.figsize'] = (8.0, 10.0)

img_path = 'data/train/'
annFile = 'train.json'

if not os.path.exists('anno_image_coco/'):
    os.makedirs('anno_image_coco/')


def draw_rectangle(coordinates, image, image_name):
    for coordinate in coordinates:
        left = np.rint(coordinate[0])
        right = np.rint(coordinate[1])
        top = np.rint(coordinate[2])
        bottom = np.rint(coordinate[3])
        # 左上角坐标, 右下角坐标
        cv2.rectangle(image,
                      (int(left), int(right)),
                      (int(top), int(bottom)),
                      (0, 255, 0),
                      2)
    imsave('anno_image_coco/'+image_name, image)


# 初始化标注数据的 COCO api
coco = COCO(annFile)

# display COCO categories and supercategories
cats = coco.loadCats(coco.getCatIds())
nms = [cat['name'] for cat in cats]
# print('COCO categories: \n{}\n'.format(' '.join(nms)))

nms = set([cat['supercategory'] for cat in cats])
# print('COCO supercategories: \n{}'.format(' '.join(nms)))

img_path = 'data/train/'
img_list = os.listdir(img_path)
# for i in range(len(img_list)):
for i in range(7):
    imgIds = i+1
    img = coco.loadImgs(imgIds)[0]
    image_name = img['file_name']
    # print(img)

    # 加载并显示图片
    # I = io.imread('%s/%s' % (img_path, img['file_name']))
    # plt.axis('off')
    # plt.imshow(I)
    # plt.show()

    # catIds=[] 说明展示所有类别的box，也可以指定类别
    annIds = coco.getAnnIds(imgIds=img['id'], catIds=[], iscrowd=None)
    anns = coco.loadAnns(annIds)
    # print(anns)
    coordinates = []
    img_raw = cv2.imread(os.path.join(img_path, image_name))
    for j in range(len(anns)):
        coordinate = []
        coordinate.append(anns[j]['bbox'][0])
        coordinate.append(anns[j]['bbox'][1]+anns[j]['bbox'][3])
        coordinate.append(anns[j]['bbox'][0]+anns[j]['bbox'][2])
        coordinate.append(anns[j]['bbox'][1])
        # print(coordinate)
        coordinates.append(coordinate)
    # print(coordinates)
    draw_rectangle(coordinates, img_raw, image_name)

三、文件配置

在训练自己的数据集过程中需要修改的地方可能很多，下面我就列出常用的几个：

修改maskrcnn_benchmark/config/paths_catalog.py中数据集路径：

class DatasetCatalog(object):
    # 看自己的实际情况修改路径！！！
    # 看自己的实际情况修改路径！！！
    # 看自己的实际情况修改路径！！！
    DATA_DIR = ""
    DATASETS = {
        "coco_2017_train": {
            "img_dir": "coco/train2017",
            "ann_file": "coco/annotations/instances_train2017.json"
        },
        "coco_2017_val": {
            "img_dir": "coco/val2017",
            "ann_file": "coco/annotations/instances_val2017.json"
        },
        # 改成训练集所在路径！！！
        # 改成训练集所在路径！！！
        # 改成训练集所在路径！！！
        "coco_2014_train": {
            "img_dir": "/data1/hqj/traffic-sign-identification/trained",
            "ann_file": "/data1/hqj/traffic-sign-identification/trained.json"
        },
        # 改成验证集所在路径！！！
        # 改成验证集所在路径！！！
        # 改成验证集所在路径！！！
        "coco_2014_val": {
            "img_dir": "/data1/hqj/traffic-sign-identification/val",
            "ann_file": "/data1/hqj/traffic-sign-identification/val.json"
        },
        # 改成测试集所在路径！！！
        # 改成测试集所在路径！！！
        # 改成测试集所在路径！！！
        "coco_2014_test": {
            "img_dir": "/data1/hqj/traffic-sign-identification/test"
        ...

config下的配置文件：

由于这个文件下的参数很多，往往需要根据自己的具体需求改，我就列出自己的配置（使用的是e2e_faster_rcnn_X_101_32x8d_FPN_1x.yaml，其中我有注释的必须改，比如 NUM_CLASSES）：

INPUT:
  MIN_SIZE_TRAIN: (1000,)
  MAX_SIZE_TRAIN: 1667
  MIN_SIZE_TEST: 1000
  MAX_SIZE_TEST: 1667
MODEL:
  META_ARCHITECTURE: "GeneralizedRCNN"
  WEIGHT: "catalog://ImageNetPretrained/FAIR/20171220/X-101-32x8d"
  BACKBONE:
    CONV_BODY: "R-101-FPN"
  RPN:
    USE_FPN: True
    BATCH_SIZE_PER_IMAGE: 128
    ANCHOR_SIZES: (16, 32, 64, 128, 256)
    ANCHOR_STRIDE: (4, 8, 16, 32, 64)
    PRE_NMS_TOP_N_TRAIN: 2000
    PRE_NMS_TOP_N_TEST: 1000
    POST_NMS_TOP_N_TEST: 1000
    FPN_POST_NMS_TOP_N_TEST: 1000
    FPN_POST_NMS_TOP_N_TRAIN: 1000
    ASPECT_RATIOS : (1.0,)
  FPN:
    USE_GN: True
  ROI_HEADS:
    # 是否使用FPN
    USE_FPN: True
  ROI_BOX_HEAD:
    USE_GN: True
    POOLER_RESOLUTION: 7
    POOLER_SCALES: (0.25, 0.125, 0.0625, 0.03125)
    POOLER_SAMPLING_RATIO: 2
    FEATURE_EXTRACTOR: "FPN2MLPFeatureExtractor"
    PREDICTOR: "FPNPredictor"
    # 修改成自己任务所需要检测的类别数+1
    NUM_CLASSES: 22
  RESNETS:
    BACKBONE_OUT_CHANNELS: 256
    STRIDE_IN_1X1: False
    NUM_GROUPS: 32
    WIDTH_PER_GROUP: 8
DATASETS:
  # paths_catalog.py文件中的配置，数据集指定时如果仅有一个数据集不要忘了逗号（如：("coco_2014_val",)）
  TRAIN: ("coco_2014_train",)
  TEST: ("coco_2014_val",)
DATALOADER:
  SIZE_DIVISIBILITY: 32
SOLVER:
  BASE_LR: 0.001
  WEIGHT_DECAY: 0.0001
  STEPS: (240000, 320000)
  MAX_ITER: 360000
  # 很重要的设置，具体可以参见官网说明：https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/README.md
  IMS_PER_BATCH: 1
  # 保存模型的间隔
  CHECKPOINT_PERIOD: 18000
# 输出文件路径
OUTPUT_DIR: "./weight/"

如果只做检测任务的话，删除 maskrcnn-benchmark/maskrcnn_benchmark/data/datasets/coco.py 中 82-84这三行比较保险。

image
maskrcnn_benchmark/engine/trainer.py 中第 90 行可设置输出日志的间隔（默认20，我感觉输出太频繁，看你自己）

四、模型训练

单GPU

官网给出的是：

python /path_to_maskrcnn_benchmark/tools/train_net.py --config-file "/path/to/config/file.yaml"

但是这个默认会使用第一个GPU，如果想指定GPU的话，可以使用以下命令：

# 3是要使用GPU的ID
CUDA_VISIBLE_DEVICES=3 python /path_to_maskrcnn_benchmark/tools/train_net.py --config-file "/path/to/config/file.yaml"

如果出现内存溢出的情况，这时候就需要调整参数，具体可以参见官网：内存溢出解决

多GPU

官网给出的是：

export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS /path_to_maskrcnn_benchmark/tools/train_net.py --config-file "path/to/config/file.yaml" MODEL.RPN.FPN_POST_NMS_TOP_N_TRAIN images_per_gpu x 1000

但是这个默认会随机使用GPU，如果想指定GPU的话，可以使用以下命令：

# --nproc_per_node=4 是指使用GPU的数目为4
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4  /path_to_maskrcnn_benchmark/tools/train_net.py --config-file "path/to/config/file.yaml"

遗憾的是，多GPU在我的服务器上一直运行不成功，还请大家帮忙解决！！！

问题地址：Multi-GPU training error

五、模型验证

修改 config 配置文件中 WEIGHT: "../weight/model_final.pth"（此处应为训练完保存的权重）
运行命令：

CUDA_VISIBLE_DEVICES=5 python tools/test_net.py --config-file "/path/to/config/file.yaml" TEST.IMS_PER_BATCH 8

其中TEST.IMS_PER_BATCH 8也可以在config文件中直接配置：

TEST:
  IMS_PER_BATCH: 8

六、模型预测

修改 config 配置文件中 WEIGHT: "../weight/model_final.pth"（此处应为训练完保存的权重）
修改demo/predictor.py中 CATEGORIES ，替换成自己数据的物体类别（如果想可视化结果，没有可以不改，可以参考demo/下面的例子）：

class COCODemo(object):
    # COCO categories for pretty print
    CATEGORIES = [
        "__background",
        ...
    ]

新建一个文件 demo/predict.py（需要修改的地方已做注释）

#!/usr/bin/env python
# coding=UTF-8
'''
@Description:
@Author: HuangQinJian
@LastEditors: HuangQinJian
@Date: 2019-05-01 12:36:04
@LastEditTime: 2019-05-03 17:29:23
'''
import os

import matplotlib.pylab as pylab
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from PIL import Image

from maskrcnn_benchmark.config import cfg
from predictor import COCODemo
from tqdm import tqdm

# this makes our figures bigger
pylab.rcParams['figure.figsize'] = 20, 12

# 替换成自己的配置文件
# 替换成自己的配置文件
# 替换成自己的配置文件
config_file = "../configs/e2e_faster_rcnn_R_50_FPN_1x.yaml"

# update the config options with the config file
cfg.merge_from_file(config_file)
# manual override some options
cfg.merge_from_list(["MODEL.DEVICE", "cuda"])


def load(img_path):
    pil_image = Image.open(img_path).convert("RGB")
    # convert to BGR format
    image = np.array(pil_image)[:, :, [2, 1, 0]]
    return image

# 根据自己的需求改
# 根据自己的需求改
# 根据自己的需求改
coco_demo = COCODemo(
    cfg,
    min_image_size=1600,
    confidence_threshold=0.7,
)

# 测试图片的路径
# 测试图片的路径
# 测试图片的路径
imgs_dir = '/data1/hqj/traffic-sign-identification/test'
img_names = os.listdir(imgs_dir)

submit_v4 = pd.DataFrame()
empty_v4 = pd.DataFrame()

filenameList = []

X1List = []
X2List = []
X3List = []
X4List = []

Y1List = []
Y2List = []
Y3List = []
Y4List = []

TypeList = []

empty_img_name = []

# for img_name in img_names:
for i, img_name in enumerate(tqdm(img_names)):
    path = os.path.join(imgs_dir, img_name)
    image = load(path)
    # compute predictions
    predictions = coco_demo.compute_prediction(image)
    try:
        scores = predictions.get_field("scores").numpy()
        bbox = predictions.bbox[np.argmax(scores)].numpy()
        labelList = predictions.get_field("labels").numpy()
        label = labelList[np.argmax(scores)]

        filenameList.append(img_name)
        X1List.append(round(bbox[0]))
        Y1List.append(round(bbox[1]))
        X2List.append(round(bbox[2]))
        Y2List.append(round(bbox[1]))
        X3List.append(round(bbox[2]))
        Y3List.append(round(bbox[3]))
        X4List.append(round(bbox[0]))
        Y4List.append(round(bbox[3]))
        TypeList.append(label)
        # print(filenameList, X1List, X2List, X3List, X4List, Y1List,
        #       Y2List, Y3List, Y4List, TypeList)
        print(label)
    except:
        empty_img_name.append(img_name)
        print(empty_img_name)

submit_v4['filename'] = filenameList
submit_v4['X1'] = X1List
submit_v4['Y1'] = Y1List
submit_v4['X2'] = X2List
submit_v4['Y2'] = Y2List
submit_v4['X3'] = X3List
submit_v4['Y3'] = Y3List
submit_v4['X4'] = X4List
submit_v4['Y4'] = Y4List
submit_v4['type'] = TypeList

empty_v4['filename'] = empty_img_name

submit_v4.to_csv('submit_v4.csv', index=None)
empty_v4.to_csv('empty_v4.csv', index=None)

运行命令：

CUDA_VISIBLE_DEVICES=5  python demo/predict.py

七、结束语

若有修改maskrcnn-benchmark文件夹下的代码，一定要重新编译！一定要重新编译！一定要重新编译！

MaskRCNN-Benchmark(Pytorch版本)训练自己的数据以及避坑指南

2. 更多精彩内容，欢迎前往我的 CSDN