LOST yolo v3 tf2

About

Adaptions of pylessons repo for yolo training (https://github.com/pythonlessons/TensorFlow-2.x-YOLOv3)

Description

This Repo is used to train a tiny YOLO V4 model to use it with lost_semi_auto_pipes Repo. The description and the default options are only to train a tiny YOLO V4 model. To train an other model look at the pylessons link above.

Installation

First, clone or download this GitHub repository and install requirements.

pip install -r ./requirements.txt

Data

To train the model you need:

• annotation data as parquet file
• train classes as json file

You get both data after the first iteration of the lost semi auto pipe.

directory structure:

.
├── Images
├── anno_data
│   ├── anno_data_0.parquet
│   ├── anno_data_1.parquet
├── checkpoints
│   ├── checkpoint
│   ├── yolov4_custom_Tiny.data-00000-of-00001
│   └── yolov4_custom_Tiny.index
├── log
├── mAP
│   ├── ground-truth
│   └── results.txt
├── model_data
│   ├── coco
│   │   ├── coco.names
│   │   └── coco_names.json
│   ├── anno_data.json
│   ├── model_test.txt
│   ├── model_train.txt
│   ├── yolov3-tiny.weights
│   └── yolov4-tiny.weights
├── yolov3
│   ├── __ init __.py
│   ├── configs.py
│   ├── dataset.py
│   ├── utils.py
│   ├── yolov3.py
│   └── yolov4.py
├── Changelog.md
├── LICENSE
├── README.md
├── convert_parquet_to_yolo.py
├── convert_to_pb.py
├── evaluate_mAP.py
├── requirements.txt
└── train.py

Edit the configs

To train your model, edit the config.py

• line 40: Path and name of the train classes json file.
• line 41: Path and name of model training data. This file and directory is creating
  by convert_parquet_to_yolo.py and will call to train the model
• line 42: Path of the image dataset
• line 43: Path of annotation data parquet file. Create this folder for all your annotation parquet files.
• line 51: If it is the first training of the model set it true else false
• line 52: If it is the first training of the model set it false else true
• line 56: Choose the number of training epochs.
• line 61: Path and name of model test data. This file and directory is creating
  by convert_parquet_to_yolo.py and will call to validate the model after training
• line 76: Path of your model repo. The target folder has to be the name of your model.

Training process

1. download anno_data.json and anno_data.parquet from data export of "tiny_yolo_triton_sia_loop" pipeline
2. load anno_data.json into your choosen directory
   • you only have to update the anno data json if there is a change since the last training
3. load or add your anno_data.parquet into your created directory
4. run convert_parquet_to_yolo.py
   • builds a list of all anno data parquet in the directory
   • train.txt and test.txt will created
5. run train.py
   • train the model and save best model at the checkpoints directory
   • saves only best model according validation loss
6. run convert_to_pb.py
   • save the model as a TensorFlow model
   • use the weights from checkpoint
   • check the model directory for older versions
   • save the new model version with one counter higher as the version befor