README.md

Training Folder Structure

The training folder is structured as followed:

.
├── logs
├── models
├── README.md
├── synth_image_creator.py
├── databuilder.py
├── experiment_log.py
├── clsf_training.py
├── complete_process.sh
└── train_only.sh

training_logs : Logs generated after model training based on the info defined in experiment_log.py. This path is defined by the log_dir parameter in the train.sh file and must be created manually, if not already exists

models : Best model generated after model training . This path is defined by the model_dir parameter in the train.sh file and must be created manually, if not already exists

synth_image_creator.py : Processing file used to convert raw image data to background replaced training data.

databuilder.py : python script for structuring background removed images into label-wise folders.

experiment_log.py : defines how logs are stored in the log directory.

clsf_training.py : main classifier training file. All the used hyperparameters are internally defined.

train_only.sh : standalone ViT training code, on backgroudn replaced and structured imaged dataset.

complete_process.sh : combines background replacement, folder structuring and training code in the same pipeline.

Training Procedure

To run the whole training from scratch, use the complete_process.sh script. The script has three stages.

1. Preprocessing raw image data from AiCity-Track 4 to replace their background with a programmatically simulated tray background

2. Relocate the background-replace images into folders according to their labels.

3. Run the training script and save the best model in a defined folder.

Here's an example script we have used :

python3 synth_image_creator.py \
--source_dir dataset/Auto-retail-syndata-release/syn_image_train \
--target_dir dataset/Auto-retail-syndata-release/bgr_images \
--segmentation_dir dataset/Auto-retail-syndata-release/segmentation_labels


python3 databuilder.py --data_dir dataset/Auto-retail-syndata-release/bgr_images 


python3 clsf_training.py \
--data_dir dataset/Auto-retail-syndata-release/ \
--log_dir training_logs \
--model_dir models

To only run training on already preprocessed and formatted dataset, use the training_only.sh and set the arguments accordingly.