Pytorch mmdetection2.0 installation training test (coco training set)

Abstract

I use it. google Server for colab For installation , Very convenient , open vpn You can use it for free k80 The graphics card , One member can use p100, Basically stable , It can meet the training needs . If it's your own computer, pay attention to the version of my installation package ,pytorch1.5 torchvision0.6 cuda yes 10.1 edition

install mmdetection2.0

The server comes with many installation packages , It is very convenient ,

Look at me colab The overall layout of , You should download the file to drive My Drive Can only be saved under , This downloaded thing will be attached to the cloud disk . Then start typing commands on the right

%cd drive/My\ Drive/

This step is to download mmdet To this file directory , It will be preserved forever

!git clone https://github.com/open-mmlab/mmdetection.git

The command in this line is to download in the current folder mmdet Package installed

!pip install -r requirements/build.txt
!pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"
!pip install -v -e .

Enter three lines of commands to download the corresponding package and compile , stay colab Every time you shut down and start up, you need to reinstall the compiler , No need to download again . Every time 5 About minutes , fast .

!pip install mmcv

In the download mmcv Even if the download is complete

from mmdet.apis import inference_detector, init_detector, show_result_pyplot
config = 'configs/cascade_rcnn/cascade_mask_rcnn_r101_fpn_1x_coco.py'
checkpoint = 'https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/cascade_rcnn/cascade_rcnn_r101_fpn_1x_coco/cascade_rcnn_r101_fpn_1x_coco_20200317-0b6a2fbf.pth'
model = init_detector(config, checkpoint, device='cuda:0')
img = 'demo/demo.jpg'
result = inference_detector(model, img)
show_result_pyplot(model, img, result, score_thr=0.3)

Run the test , Just run it directly .

Modify the package with its own version , Otherwise, the test will report an error

!pip install pillow==7.1.2
!pip install -U numpy==1.17.0

Run these two lines of code each time

Training coco Training set

Modify category mmdet core evalution class_name.py

The category needs to be changed to your own

modify configs Under the _bass_ models for example cascade-rcnn

modify num_classes= Number of your own categories , You don't need to add one , My is 44 Categories , Namely 44. There are three num_classes Just modify all .

Data sets are still converted into coco Training set of , Unwanted segment part

Generate json file

import os
import cv2
import json
import xml.dom.minidom
import xml.etree.ElementTree as ET

data_dir = './data' # Root file , It includes image The folder and box Folder （ Modify this path according to your own situation ）

image_file_dir = os.path.join(data_dir, 'image')
xml_file_dir = os.path.join(data_dir, 'box')

annotations_info = {
    'images': [], 'annotations': [], 'categories': []}

categories_map = {
    'holothurian': 1, 'echinus': 2, 'scallop': 3, 'starfish': 4}

for key in categories_map:
    categoriy_info = {
    "id":categories_map[key], "name":key}
    annotations_info['categories'].append(categoriy_info)

file_names = [image_file_name.split('.')[0]
              for image_file_name in os.listdir(image_file_dir)]
ann_id = 1
for i, file_name in enumerate(file_names):
    print(i)
    image_file_name = file_name + '.jpg'
    xml_file_name = file_name + '.xml'
    image_file_path = os.path.join(image_file_dir, image_file_name)
    xml_file_path = os.path.join(xml_file_dir, xml_file_name)

    image_info = dict()
    image = cv2.cvtColor(cv2.imread(image_file_path), cv2.COLOR_BGR2RGB)
    height, width, _ = image.shape
    image_info = {
    'file_name': image_file_name, 'id': i+1,
                  'height': height, 'width': width}
    annotations_info['images'].append(image_info)

    DOMTree = xml.dom.minidom.parse(xml_file_path)
    collection = DOMTree.documentElement

    names = collection.getElementsByTagName('name')
    names = [name.firstChild.data for name in names]

    xmins = collection.getElementsByTagName('xmin')
    xmins = [xmin.firstChild.data for xmin in xmins]
    ymins = collection.getElementsByTagName('ymin')
    ymins = [ymin.firstChild.data for ymin in ymins]
    xmaxs = collection.getElementsByTagName('xmax')
    xmaxs = [xmax.firstChild.data for xmax in xmaxs]
    ymaxs = collection.getElementsByTagName('ymax')
    ymaxs = [ymax.firstChild.data for ymax in ymaxs]

    object_num = len(names)

    for j in range(object_num):
        if names[j] in categories_map:
            image_id = i + 1
            x1,y1,x2,y2 = int(xmins[j]),int(ymins[j]),int(xmaxs[j]),int(ymaxs[j])
            x1,y1,x2,y2 = x1 - 1,y1 - 1,x2 - 1,y2 - 1

            if x2 == width:
                x2 -= 1
            if y2 == height:
                y2 -= 1

            x,y = x1,y1
            w,h = x2 - x1 + 1,y2 - y1 + 1
            category_id = categories_map[names[j]]
            area = w * h
            annotation_info = {
    "id": ann_id, "image_id":image_id, "bbox":[x, y, w, h], "category_id": category_id, "area": area,"iscrowd": 0}
            annotations_info['annotations'].append(annotation_info)
            ann_id += 1

with  open('./data/annotations.json', 'w')  as f:
    json.dump(annotations_info, f, indent=4)

print('--- Sorted annotation file ---')
print(' Number of all pictures ：',  len(annotations_info['images']))
print(' Number of all indications ：',  len(annotations_info['annotations']))
print(' Number of all categories ：',  len(annotations_info['categories']))

modify mmdet datasets coco.py file

The with seg Part of the code line is commented out

modify mmdet core evaluation class_name.py file

Change this code into your own category

The data is placed as follows , there custom You can change it to coco There is no need to modify the code that loads the data

!python tools/train.py configs/cascade_rcnn/cascade_rcnn_r50_fpn_1x_coco.py --gpus 1

The running code is almost the same ,

!python tools/test.py configs/cascade_rcnn/cascade_rcnn_r50_fpn_1x_coco.py  checkpoints/grid.pth --eval bbox

The evaluation command here is coco Evaluation command of training set , Follow voc Different .

colab Use

Here is entering the cloud disk , It is convenient to upload your own data here

Here to mmdetection Next, you can right-click to upload data . Only 15 individual g You can open another member to expand memory .

summary

Use colab The server is very convenient . At least it can be installed very quickly mmdet Use . The package inside may be too old and too new , All need debugging . Download the package quickly . I'm here, too colab Trained in efficientdet The Internet .