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:
{}
.format( .join(nms)))

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

# 载入并显示图片
# 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。

• 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在我的伺服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

1. 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！

更多内容欢迎关注博客是大黄鸭 BLOG。

推荐阅读：