Description

Objective
– Continue gaining experience with PyTorch and helper libraries
– Understand the VOC Object Detection Dataset
– Train and evaluate the SSD neural network architecture
– Perform an ablation study testing a different base network and learning rate schedule
– Learn the Non-Maximum Suppression (NMS) algorithm

Resources and Instructions Environment Setup:
We recommend using Google Colab to complete this assignment.
1. Create a folder called “ece495_assignment4” within your Google Colab “Colab Notebooks” folder.
2. Upload the assignment ipynb, utils.py and json files to the Google Colab “ece495_assignment4” folder
3. Open the assignment
• Runtime -> change runtime type
• Set hardware accelerator to GPU
•

Assignment:
2. Ablation study on using a different network base
• Model A: Train and evaluate the SSD network with the default VGG base.
• Model B: Implement the ResNetBase class. Then train and evaluate this model.
3. Ablation study on updating the learning rate
• Model C: Train and evaluate the SSD network with the default VGG base but also with a PyTorch learning rate scheduler.
4. Answer 2 questions on the differences from the NMS pseudo code described in the lectures / tutorial and the implemented version in the code.

Deliverable HTML output:
In the Jupyter notebook, go to File > Download as > HTML (.html) Submit a ZIP file containing the HTML output. Please follow the naming convention of your zip file: a4_<user_id>.zip

Marking Scheme

This assignment is out of 10 marks.
• 2 marks for training and evaluation output for the SSD network with the default VGG base
• 3 marks for implementing the ResNet base and displaying the training and evaluation output
• 3 marks for adding the learning rate scheduler and the training and evaluation output for the SSD network with the default VGG base
• 1 mark for the model with the learning rate scheduler to get at least 50% mAP
• 1 mark for the NMS questions

Policies
Collaboration
You can discuss the problem with peers, but you must design and implement your own solution independently.

Use of online resources