ML_Object_Detection

• This tool assists you in training your object detection model and converting it to a TFLite model, which can be easily deployed on your device.
• The training framework utilizes TensorFlow Object Detection API on TensorFlow 2 and TensorFlow 1.
• The notebooks simplify the complicated installation steps and provide an easy-to-use approach for data preparation, training, and conversion.

1. First step

1. Install virtual env

• If you haven't installed NuEdgeWise, please follow these steps to install Python virtual environment and choose NuEdgeWise_env.
• Skip if you have done.

2. Install the Visual C++ 2015 build tools

• Log in to https://my.visualstudio.com/Downloads (You will need a free Microsoft account).
• Enter 'Build Tools' in the search bar.
• Select 'Visual Studio 2015 Update 3' on the left side.
• Click on 'DVD' under 'Visual C++ Build Tools...' and initiate the download.
• Follow the installation steps to complete the installation process.

3. Install git windows version

• Git is required during the 5th step.

4. Download this git folder

• git clone https://github.com/OpenNuvoton/ML_tf2_object_detection_nu.git
• Or you can download the zip file directly

5. Object Detection API installation

• Open Miniforge or your python environment with administrator privileges and select the NuEdgeWise_env. Utilize setup_objdet_tf2.ipynb to install the remaining packages.

6. TF1 Object Detection API installation (Optional)

• If you wish to use ssd_mobileNetv3, TensorFlow 1 is required, so you need to create a new Python virtual environment not using NuEdgeWise_env.
• First, refer to create_conda_env_tf1.ipynb, and then proceed to setup_objdet_tf1.ipynb.

---

2. Work Flow

1. data prepare

• Open create_data.ipynb located in image_dataset. Ensure that you execute it in a TensorFlow 2 environment.
• This process will handle the downloading of open-source images or labeling your customized images to create a training dataset.
• The tutorial for this process is provided within the create_data.ipynb notebook.

2. train & tflite model creating

• Open the workspace folder.
• Users need to choose either the TensorFlow 1 or TensorFlow 2 environment based on their training model.
• This process will handle training, mAP evaluation, and TFLite conversion.
• The tutorial for this process is provided within the notebook.

tensorflow2

• Use train_evl_monitor_tf2.ipynb for an easy-to-use user interface that facilitates training, evaluation, and monitoring.
• Use convert_tflite_tf2.ipynb for an easy-to-use user interface that aids in converting to TFLite format.

[alternative way]:

• Open train_cmd_tf2.ipynb.
• Google's object detection support models: TensorFlow 2 Detection Model Zoo

tensorflow1

• Open train_cmd_tf1.ipynb. This should be excuted in tf1 env (Check create_conda_env_tf1.ipynb & setup_objdet_tf1.ipynb to create tf1 env).
• Google's object detection support models: TensorFlow 1 Detection Model Zoo

3. evaluation & test

• Open test_tflite.ipynb located in the workspace folder. Make sure to execute it in a TensorFlow 2 environment.
• Evaluate the results of the TFLite model using your own dataset.
• The tutorial for this process is provided within the test_tflite.ipynb notebook.

3. Inference code

• The MPU example code: MA35D1_machine_learning