Preface

Code: https://github.com/dog-qiuqiu/FastestDet

Algorithm introduction: https://zhuanlan.zhihu.com/p/536500269

Training environment: Win11

1. Data set construction

The data set format is the same as Darknet Yolo, and each picture corresponds to a txt tag file. The label format is also based on Darknet Yolo's data set label format: "category cx cy wh", where category is the category subscript, cx, cy are the coordinates of the center point of the normalized label box, and w and h are the coordinates of the normalized label box. Width and height, examples of .txt tag file content are as follows:

2 0.2841796875 0.3815104166666667 0.068359375 0.257825
1 0.43505859375 0.55078125 0.0966796875 0.122395833333333

The data set file structure is as follows:

The format of train.txt and val.txt is:

E:\cv_code\FastestDet-main\datasets\images\train\aug_0_10.jpg
E:\cv_code\FastestDet-main\datasets\images\train\aug_0_100.jpg
E:\cv_code\FastestDet-main\datasets\images\train\aug_0_101.jpg

Code examples to generate train.txt and val.txt:

import os
import glob

# Define the directory containing the dataset
train_dir = r'E:\cv_code\FastestDet-main\datasets\images\train'
val_dir = r'E:\cv_code\FastestDet-main\datasets\images\val'

# Get a list of all file paths in the dataset directory
train_files = glob.glob(os.path.join(train_dir, '*.jpg'))
val_files = glob.glob(os.path.join(val_dir, '*.jpg'))

# Write the file paths to a .txt file
with open('datasets/train.txt', 'w') as file:
    for path in train_files:
        file.write(path + '\n')

with open('datasets/val.txt', 'w') as file:
    for path in val_files:
        file.write(path + '\n')

2. Configuration file settings

1..Yaml configuration file introduction

Take ./configs/coco.yaml as an example:

DATASET:
  TRAIN: 'E:\cv_code\FastestDet-main\datasets\train.txt'       # train.txt
  VAL: 'E:\cv_code\FastestDet-main\datasets\val.txt'           # val.txt
  NAMES: "configs/coco.names"                                  # label name
MODEL:
  NC: 4                                                        # number of classes
  INPUT_WIDTH: 352                                             # input width
  INPUT_HEIGHT: 352                                            # input height
TRAIN:
  LR: 0.001                                                    # learning rate
  THRESH: 0.25                                                 # threshold
  WARMUP: true                                                 # warmup
  BATCH_SIZE: 96                                               # batch size
  END_EPOCH: 10                                                # train epoch
  MILESTIONES:                                                 # Declining learning rate steps
    - 100
    - 200
    - 250

Where coco.names is the label name of your data set

person
bicycle
car
motorbike
.......
toothbrush

2. Start training

The pre-trained model is: shufflenetv2.pth

Specify your configuration file:

        parser.add_argument('--yaml', type=str, default='configs/test.yaml', help='.yaml config')

Start training:

python3 train.py

3. Model evaluation

python3 eval.py --yaml configs/coco.yaml --weight weights/weight_AP05:0.253207_280-epoch.pth

4. Export onnx for testing

python3 test.py --yaml configs/coco.yaml --weight weights/weight_AP05:0.253207_280-epoch.pth --img data/3.jpg --onnx

Summarize:

The above is what I will talk about today. This article only briefly introduces the use of FastestDet. Come on, friends, go practice it.

FastestDet trains its own data set