TRAFFIC SIGN CLASSIFICATION

This project builds a machine learning model using the python language and leveraging on a convolutional neural network for classification of 58 distinct traffic signs. Each image is converted to a tensor within Python script "traffic.py" to aid training.

Prerequisites

The ML operations within this repo was implemented using Azure machine learning (AML) studio and Azure Devops. A more general approach is to implement the CI/CD with docker but due to technical issues during the creation of this project, Azure DevOps was adopted. The following prerequisites are required to make this repository work:

• Azure subscription
• Owner or contributor access to the Azure subscription
• Azure DevOps account
• Create an Azure resource group
• Azure workspace within that resource group
• Create a compute instance/cluster running in AML studio

Data Preparation

• The data for this project was downloaded from the Kaggle site. It contains 58 distinct traffic signs resulting in 58 labels to be classified
• The data in the DATA folder is split into two seperate folders, 'train' and 'val', holding 80% and 20% respectively, of the total set.
• Folders 'train' and 'val' are then imported to the python script, taffic.py.

Python script

• The traffic.py file contains the python code used to develop this model.
• The requirements for this file can be installed by:

pip install requirements.txt

• Within this script, the PyTorch framework is adopted. It is a widely utilised framework for building and training neural networks and can be imported by:

    import torch as torch

Neural Network - The Model

• A pretrained CNN, RESNET-50 is used to train the model via transfer learning - output size is adjusted to meet the number of classes intended for classification. It contains 50 layers as the name connotes; 48 Convolutional layers, 1 MaxPool and 1 Average Pool layer.

    resnet50 = models.resnet50(pretrained=True)

• For faster re-training of the model, support for NVIDIA GPU's is integrated within the traffic.py script in line 71 as:

resnet50 = resnet50.to('cuda:0')  for NVIDIA GPU

Note that it needs to be uncommented for it to function.

• The Adam optimizer is used with this network and its learning rate is kept at a standard value of 0.001 but can be adjusted

optimizer = optim.Adam(resnet50.parameters(), lr=0.001)

• The model is then saved in folder named 'models'

CI/CD Pipeline

• For integrating and deploying this model automatically, the azure DevOps tool is utilised
• A script named 'start_experiment.py' defines the environment, model, Script run configuration and retrieves the python script, 'traffic.py', for retraining the model
• Also wihin this script, the azure compute instance created is defined as well as the workspace, subscription ID, and resoure-group are specified. This can always be modified.

compute_name = 'hackcompute22'

experiment_name = 'my-experiment' 
environment_name = 'my-environment'

model_name = 'my-model'
model_path = 'outputs/model.pkl'


source_directory = '.'

script_path = 'Hack Partners/traffic.py'


subscription_id = '4f67948b-2ff9-49ee-bf1f-90c32dc7545e'
resource_group  = 'cloud-shell-storage-westeurope'
workspace_name  = 'Hack_Workspace'

• The Yml file, azure-pipelines.yml then runs this script, 'start_experiment.py'
• Within this file, the GitHub branch where the pipeline retrieves and runs within is specified - 'main'. In the long run, this could be changed. The python 3.9 version is also specified to be used and the Azure mnachine learning sdk is installed.
• The pipeline file triggers the experiment file to run thereby re-training the model in the traffic.py script at intervals when a push to the traffic.py file is made.

update file