Tflite model trainer

Features

• Automatic download of open images dataset (+ conversion to voc format)
• Training of EfficientDet for object detection
• Save fine-tuned model in tflite model

---

Getting started

Requirements

• Python = 3.8
• Linux OS (or wsl2)

Environment

The following pipeline of commands was tested on windows subsystem for linux (wsl) - ubuntu and python 3.8

1. Install python3.8 and virtualenv tool.

2. Create and activate virtual environment:

virtualenv --python=python3.8 env

Check if activated using python --version -> should print python version 3.8.x

3. Upgrade pip

python -m pip install -U pip

4. Install requirements:

pip install -r requirements.txt

5. Install required tools:

sudo apt-get install libsndfile1
sudo apt install libportaudio2

6. Install python packages in consecutive order:

pip install -q --use-deprecated=legacy-resolver tflite-model-maker
pip install -q opencv-python-headless==4.1.2.30
pip uninstall -y tensorflow && pip install -q tensorflow==2.8.0
pip install numpy==1.20.3
pip install oidv6-to-voc
pip install boto3==1.24.91
pip install botocore==1.27.91

---

Use

1. Download images:

python open_images_downloader.py --max-images=500 --class-names "Tennis ball" --data=data/dataset

2. Convert images:

oidv6-to-voc ./data/dataset/sub-train-annotations-bbox.csv -d ./data/dataset/class-descriptions-boxable.csv --imgd ./data/dataset/train --outd ./data/dataset/annotations_train

oidv6-to-voc ./data/dataset/sub-validation-annotations-bbox.csv -d ./data/dataset/class-descriptions-boxable.csv --imgd ./data/dataset/validation --outd ./data/dataset/annotations_validation

oidv6-to-voc ./data/dataset/sub-test-annotations-bbox.csv -d ./data/dataset/class-descriptions-boxable.csv --imgd ./data/dataset/test --outd ./data/dataset/annotations_test

3. Run train.py script.