/dcn-4-heads-yolov5

Add dcn and a 4 heads model config file in 'models' folder to process objects in different size

Primary LanguagePython

dcn-4-heads-yolov5

Add dcn and a 4 heads model config file in 'models' folder to process objects in different size

Keypoint

  1. how to use github to download code and find issues
  2. how to use code/folders downloaded ()
  3. knowledge of python/pytorch/numpy etc. (how to code deeplearning model in python language)
  4. how to use colab/azure to run code and setup environment for yolov5
  5. knowlegde of deeplearning and YOLO

Github

  • You can find the source code from YOLOv5
  • you can try to explore problems in issues section
  • use git clone to get the newest code

How to use ultralytics/YOLOv5 code?

Ⅰ. Folders and files

1. data

dataset.yaml

  • The yaml files in this folder aim to configure dataset used in training/testing/ process. You can change the dictionary in these files to make use of your own dataset.

  • Note: directory of images and labels should be orgnized in example way, such as coco.yaml

class names in dataset.yaml

  • items in this list should be in the sequence of label number from 0 in Label txt files

hyp.yaml

  • Hyperparameters setting
  • scracth is used to training from the begining or to retrain a model
  • finetune is used to training from pretrained model or to finetune models

2. models

  • You can configure yolov5 model here.

yolov5s/m/l/x.yaml

  • Adjust model depth and with on the top of the file opened
  • Models are modified in modular approch (which modules are arranged in model)

common.py

  • Modules are coded here, such as C3, Conv, SPP etc. used in model's yaml (what are modules)

yolo.py

  • Model construction based on common.py and yolov5.yaml is completed here (how the models are construct in details)
  • When moudles added to common.py, yolo.py should also be modified somewhere concerned.

3. utils

  • tool files

general.py

loss.py

  • to build targets of input, then compute loss
  • loss functions are defined here

plots

  • to plot figures of train/val results in results.txt
  • to plot metrics

metrics

  • to define how to compute precision/Recall/mAP ....

4. weights

  • This folder contains weight files for yolov5 models, the files inside in .pt format and will be download automatically if not existed when you use scripts to train. The .pt files are general in deep learning codes, you may see .h5 or .pth weight files in other deep learning codes.

5.runs

  • train/test/detect results saved here, figures contained

6.requirements.txt

  • python packages list required

7.train.py test.py detect.py

  • train.py: train a model and save in .pt file format
  • test.py: to test a model trained on val/test dataset, and obtain the metrics/figures
  • detect.py: use a trained model to detect pictures/video/camera data, and save results

Good blogs:

usage of YOLOv5 code from download to train your own dataset

analysis of train.py for YOLOv5

analysis of test.py for YOLOv5

how YOLOv5 algorithm work logically

What is YOLO?

YOLO models are one-stage object detection methods formulated by deep learning, whose name from "you only look once". You can learn yolo from this website: zhihu-yolo series

Suggested learning steps of YOLOv5

  1. knowledge about neural network (weights, linear and non-linear functions, optimizer, loss functions)
  2. knowlege about object detection (mAP, one stage and two stage method)
  3. knowlege about steps of construct a model in python and how to train (what's train/valitation/test)