How to Re-train AI model

This page explains how to re-train the AI model of AI Applications using RZ/V AI Transfer Learning Tool.

Supported version: RZ/V AI Applications v4.00

Introduction

RZ/V AI Transfer Learning Tool (TLT) can re-train AI model used in RZ/V AI Applications with different datasets.

RZ/V AI Transfer Learning Tool has 5 Sections.

• PROJECT DETAILS: Create new project or load an existing project.
• ANNOTASTION: Create your own training data by labeling images.
• DATASET: Load the dataset to be used for training.
• TRAINING: Train the model on the selected dataset.
• INFERENCE: Test the trained model.

Supported Functions of AI Applications

Function Name of AI Application	TLT Project Name	AI Model	Task Type	Download Dataset Name *1	Target product in AI Application
					RZ/V2L	RZ/V2H
01_Head_count	01_Head_count_yolov3	YOLOv3	Object detection	Head Count(140MB)	✔	✔
02_Line_crossing_object_counting	02_Line_crossing_object_counting_tinyyolov3	TinyYOLOv3	Object detection	Line Cross Object Counting(141MB)	✔	-
	02_Line_crossing_object_counting_yolov3	YOLOv3	Object detection	Line Cross Object Counting(141MB)	-	✔
04_Safety_helmet_vest_detection	04_Safety_helmet_vest_detection_yolov3	YOLOv3	Object detection	Safety Helmet Vest Detection(3.4GB)	✔	-
07_Animal_detection	07_Animal_detection_yolov3	YOLOv3	Object detection	Animal Detection(1.83GB)	✔	✔
09_Human_gaze_detection	09_Human_gaze_detection_resnet18	ResNet18	Classification	Human gaze detection resnet18(17GB)*2	✔	✔
	09_Human_gaze_detection_tinyyolov2	TinyYOLOv2	Object detection	Human gaze detection(4.43GB)*2	✔	✔
11_Head_count_topview	11_Head_count_topview_yolov3	YOLOV3	Object detection	Head Count Topview(1.8GB)	✔	✔
12_Hand_gesture_recognition_v2	12_Hand_gesture_recognition_v2_resnet_18	ResNet18	Classification	Hand Gesture Recognition Resnet(21.9GB)	✔	-
	12_Hand_gesture_recognition_v2_tinyyolov3	TinyYOLOv3	Object detection	Hand Gesture Recognition Tiny Yolo(21.9GB)	✔	-
	12_Hand_gesture_recognition_yolov3	YOLOv3	Object detection	Hand gesture recognitionyolov3(160MB)	-	✔
13_Car_ahead_departure_detection	13_Car_ahead_departure_detection_tinyyolov3	TinyYOLOv3	Object detection	Coco	✔	-
Q01_footfall_counter	Q01_footfall_counter_tinyyolov3	TinyYOLOv3	Object detection	Coco	✔	-
	Q01_footfall_counter_yolov3	YOLOV3	Object detection	Coco	-	✔
Q03_smart_parking	Q03_smart_parking	Custom model	Classification	Smart Parking(414MB)	✔	-
Q04_fish_classification	Q04_fish_classification_resnet34	ResNet34	Classification	Fish Classification(1.57GB)	✔	-
Q06_expiry_date_detection	Q06_expiry_date_detection_tinyyolov3	TinyYOLOv3	Object detection	Expiry Date Detection(1.32GB)	✔	-
	Q06_expiry_date_detection_yolov3	YOLOv3	Object detection	Expiry Date Detection(1.32GB)	-	✔
Q07_plant_disease_classification	Q07_plant_disease_classification_resnet34	ResNet34	Classification	Plant Disease(1.34GB)	✔	-
Q08_object_counter	Q08_object_counter_animal_tinyyolov3	TinyYOLOv3	Object detection	Object Counter Animal(6.97GB)	✔	-
	Q08_object_counter_animal_yolov3	YOLOv3	Object detection	Object Counter Animal(6.97GB)	-	✔
	Q08_object_counter_coco_tinyyolov3	TinyYOLOv3	Object detection	Coco	✔	-
	Q08_object_counter_coco_yolov3	YOLOv3	Object detection	Coco	-	✔
	Q08_object_counter_vehicle_tinyyolov3	TinyYOLOv3	Object detection	Object Counter Vehicle(1.5GB)	✔	-
	Q08_object_counter_vehicle_yolov3	YOLOv3	Object detection	Object Counter Vehicle(1.5GB)	-	✔
Q09_crack_segmentation	Q09_crack_segmentation_unet	UNet	Segmentation	Crack Segmentation(0.99GB)	✔	✔
Q10_suspicious_person_detection	Q10_suspicious_person_detection_tinyyolov3	TinyYOLOv3	Object detection	Suspicious Person Detection(1.57GB)	✔	-
	Q10_suspicious_person_detection_yolov3	YOLOv3	Object detection	Suspicious Person Detection(1.57GB)	-	✔
Q11_fish_detection	Q11_fish_detection_tinyyolov3	TinyYOLOv3	Object detection	Fish Detection(700MB)	✔	-
	Q11_fish_detection_yolov3	YOLOv3	Object detection	Fish Detection(700MB)	-	✔

*1: Datasets will be downloaded automatically as needed.
*2: This file is temporarily unavailable for automatic download. To obtain the file, please contact Renesas Technical Support.

Unsupported Functions of AI Applications

The following Functions of AI Applications are NOT supported.

• 03_Elderly_fall_detection
• 05_Age_gender_detection
• 06_Face_recognition_spoof_detection
• 08_Hand_gesture_recognition
• 10_Driver_monitoring_system
• Q02_face_authentication
• Q05_suspicious_activity

Installation

To start using RZ/V AI Transfer Learning Tool, PC that can display the desktop with the following environment is recommended.

Hardware

Description	Specification
Linux OS	Ubuntu 20.04 and above.
Processor	Intel(R) Xeon(R) and above.
GPU	Any GPU with CUDA.
Storage	At least 100GB of free space is necessary.

Software

Description	Specification
Python	3.10.5
Docker	27.0.3

Other software not listed above will be installed automatically when you install the tool.

Installation procedures

This chapter describes the installation procedure when using RZ/V AI Transfer Learning Tool alone.

Note If you use AI Navigator of e² studio, the installation procedure is different, so please refer to AI Navigator Quick Start Guide.

1. If you have not yet obtained RZ/V AI Transfer Learning Tool v4.00, click on the link below to download it.
   
   Download Link
   
   
2. Put the downloaded tar file into a Linux PC and extract it using the following command in a terminal.
   

   tar -xzvf RTK0EF0178Z04000SJ_rzv-ai-tlt.tar.gz

   
   Make sure the following folders and files are generated after extracting the tar file.

   Tool File Structure

   Folder or File Name	Description
   rzv_ai_tlt_v4.00/	Working directory.
   docs/	User's Manual.

3. Change to tlt_backend directory and install the tool.
   

   cd RTK0EF0178Z04000SJ_rzv-ai-tlt/rzv_ai_tlt_v4.00/tlt_backend
   ./install_docker.sh

   
   
4. Build the docker image and start the docker container.
   

   ./launch_tlt_service.sh

   
   
5. Change to rzv_ai_tlt_v4.00 directory and launch the tool.
   

   cd ../
   ./start_rzv_ai_tlt_gui.sh

   
   
   
   If the popup is "Docker container is up and running", click "OK" to start.
   
   Note If the docker container is not running, close the GUI and start the docker container with the following command.
   

   sudo docker restart tltdoc

   Once the container is started, please start again from step 4.

Re-train the AI model

Once the installation is complete, you can retrain the model.

Step 1: Create the TLT project

1. Select PROJECT DETAILS.
   
   
   
   
2. After selecting UPLOAD, upload the YAML file in UPLAOD YAML.
   The YAML files for existing projects are located at rzv_ai_tlt_v4.00/tlt_backend/assets/(any YAML file of TLT project)/config.yaml.
   This time, we uploaded rzv_ai_tlt_v4.00/tlt_backend/assets/Q03_smart_parking/config.yaml.
   
   
   
   
3. Set the project name and click SUBMIT SELECTION.
   (The project name must be different from any existing project.)
   If successful, Successfully created the task!! will be displayed.
   
   
   
   

Step 2: Create the custom dataset

In ANNOTATION, you can create custom datasets only for object detection.



This time, since Q03_smart_parking is a classification model, we will skip ANNOTATION.
Please refer to the user's manual for how to use ANNOTATION.

Step 3: Load the dataset

1. Select DATASET.
   
   
   
   
2. Click Download Dataset and Select a dataset from the list.
   This time, we use Smart Parking.
   
   
   
   
3. Click the Download button to download the dataset.
   (After it reaches 100%, it will take some time to unpack the downloaded file, depending on the size of the dataset.)
   
   
   
   
4. Click SELECT DIRECTORY to select the dataset directory, and click NEXT.
   The downloaded dataset is located at rzv_ai_tlt_v4.00/tlt_backend/datasets/.
   This time, we use rzv_ai_tlt_v4.00/tlt_backend/datasets/Q03_smart_parking/.
   
   
   
   
5. Set the image size to input to the model and select Augmentation.
   This time, leave the defaults and click NEXT.
   
   
   
   
6. Set the dataloader parameters according to your PC environment.
   This time, leave the defaults and click FINISH.
   
   
   
   
7. After confirming the settings, click SUBMIT SELECTION.
   If the load is successful, the dataset configuration will be displayed.
   
   
   
   

Step 4: Re-train the model

1. Select TRAINING.
   
   
   
   
2. Check the model and set the optimizer parameters.
   Click SUBMIT SELECTION and the learning log will be displayed.
   This time, leave everything as default.
   
   
   
   
3. When Training Completed!!! pop-up appears, click OK to complete the training.
   
   
   
   The trained model is stored in the following directory.
   rzv_ai_tlt_v4.00/tlt_backend/assets/(Your TLT prject name).
   

Step 5: Test the trained model

1. Select INFERENCE
   
   
   
   
2. Click UPLOAD IMAGE to select a test image.
   This time, we use rzv_ai_tlt_v4.00/tlt_backend/datasets/Q03_smart_parking/valid/occupied/0.jpg.
   
   
   
   
3. Click PREDICT to test.
   Since Q03_smart_parking is a classification that predicts whether it is "occupied" or "empty", the result is displayed as a string.
   
   
   
   

Tips

Next steps

Once the model has been trained, the next step is to convert it for DRP-AI using DRP-AI TVM.
For instructions on how to use DRP-AI TVM, please see DRP-AI TVM GitHub.
If you use DRP-AI TVM with AI Navigator, sample code for DRP-AI TVM conversion scripts for each model is provided.

Uninstallation

For instructions on how to uninstall TLT, please refer to the user's manual.

Back to Home >