Training a P1 Image Classification Model

What makes Computer Vision useful is the fact that models can be trained to classify user-defined objects.

In this guide, the instructions will guide you on how to train your very own Precision-Level 1 (P1) Image Classification computer vision model which could then used to deploy as a modular ROS2 package.

Collect Training Data

A neural network model, much like a student, requires training data to know what are the intended objects it is meant to detect. In the case of Computer Vision, these training data refers to images of user-defined objects. If the user intention is to train a model to distinguish between cats and dogs, the training data ought to be photos of cats and dogs.

Therefore, please take as many as photos or images that suit the following criteria:

Criteria No.	Description
1	A training image should only show one class of object.
2	The lighting condition in the image should be similar to the physical setup of the camera.
3	A training image should be at least of size, 224 by 224.

The recommended number of images to capture is: 30 per object class.

Eg.

30 different photos of cats
30 different of dogs

Label Training Data

Arranging Data

Among the three precision levels, labelling a P1 training image dataset is the easiest. It is only a matter of arranging files in the following file structure:

insert_your_training_here_dataset/
   |_train
     |_name_of_object_class_1
       |_<photos of object_class_1>
     |_name_of_object_class_2
       |_<photos of object_class_2>
     |_ ...
   |_val
     |_name_of_object_class_1
       |_<photos of object_class_1>
     |_name_of_object_class_2
       |_<photos of object_class_2>
     |_ ...

Eg.

p1_catsdogs_dataset/
   |_train
     |_cat
       |_<photos of cats>
     |_dog
       |_<photos of dogs>
   |_val
     |_cat
       |_<photos of cats>
     |_dog
       |_<photos of dogs>

Create Custom Label List

Once you have arranged your file like as shown above, create a .txt label list.

Eg.

cat
dog

Ensure that the object names are sorted in lexicographical order.

Note

P2 and P3 training datasets will adopt a slightly different label list format.

Train Model

Follow the instructions below to get started training and generating your P1 ONNX model.

1. Double-click on easy_perception_deployment.desktop on your Desktop. This file is generated following instructions under the Setup section.

A window labelled, easy_perception_deployment, will appear.

2. Click on button labelled, Train, will appear.

A window labelled, Train, will appear.

3. Click on button labelled, P1.

You have selected Precision Level 1 training framework.

4. Click on DropDown Menu, Set Model and select the model you wish to run.

There are currently only 6 pretrained model available in the PyTorch Model Zoo. Each has its own pros and cons.

[Tips on choosing]:
# If you wish to prioritize speed over accuracy,
Choose [mobilenet].
# If you wish to prioritize accuracy over speed,
Choose [densenet].

5. Click on button labelled, Choose Label List and select the .txt label list file you have created under Create Custom Label List.

6. Click on button labelled, Choose Dataset and select the training dataset folder you have created under Arranging Data.

7. Click on button labelled, Validate Dataset to verify if your dataset is in the correct prescribed file structure.

If the dataset is properly arranged, the button labelled, Train will now be highlighted.

8. Click on button labelled, Train to start training.

The final .onnx trained model will be written to /data/model/ directory in the package and timestamped with the date in which it was trained in the filename.