OFSNet_training

Training process for the OFSNet

This repository contains the code for fine-tuning the neural network, OFSNet (Optical Flow-based Segmentation Network) based on the collected and prepared data. For the data collection and preparation see the data_collection. The trained OFSNet is used in the obstacle-detection repository.

How to use

The code in this repository requires the obstacle-detection repository to be present at the target machine already as it contains the definitions for the OFSNet.

Clone the repository so it is in the same directory as the obstacle-detection repository:

git clone https://github.com/karolyartur/OFSNet_training

Make sure that the obstacle-detection repository is at the latest commit of the deploy branch (most up-to-date network definitions) and that the resulted directory structure looks like this:

│  
├── obstacle-detection
├── OFSNet_training

After this, create a folder named data under OFSNet_training and make two subdirectories of the data folder called train and valid.

The resulted file structure should look like this:

│  
├── obstacle-detection
├── OFSNet_training
        │
        ├── data
        │   │
        │   ├── train
        │   ├── valid
        │
        ├── .gitignore
        ├── fine-tuning.py
        ├── README.md     

After preparing the datasets with the data_collection repository, move the generated datasets (the individual folders, not the whole datafolder) into the train or the valid folders. When both folders are ready it is time to train the OFSNet.

Before running the fine-tuning of the network, make sure that the virtual environment used for the obstacle detection package is activated, then type:

python fine-tuning.py

The training of the network will start. There are three parameters that can be set by changing their value inside the python script.

• Batch size: The number of inputs in a training batch (20 by default)
• Number of iterations: The number of maximum iterations after which the training stops (2000 by default)
• Number of training steps after saving and validation: By default after every 500 training step the model is saved and validated

Results

After the training the results are written in a newly created folder named fine_tuning_result in the root of the repository. This folder contains the fine-tuned neural network definition files and the saved training statistics.

In order to visualize the training statistics use TensorBoard with the logdir parameter set to the fine_tuning_result folder.

After the fine-tuning, the updated model files can be copied into the obstacle-detection package for use. This way, the OFSNet definition used for the obstacle detection can be fine-tuned for new environments.