Feedback

[jmacey]

Please add a README.md file with the project ideas in it and how you are going to approach it so I can begin to give feedback.

[RoryMabina]

I have updated the Readme.md file

[jmacey]

What sort of objects are you going to try and tag? Are you going to do some transfer learning to use your own custom items etc?

Try to use marimo over jupyter if you can. The note book is using os to scan folders etc, as mentioned in class pathlib is best practice for this now.

[RoryMabina]

The CIFAR-100 dataset, which includes 100 categories such animals, cars, domestic goods, and natural objects, will be recognised and tagged by the system. I will also map these objects into more general semantic tag groupings, such as "animal," "vehicle," "nature," and "misc." Additionally, I will use transfer learning by fine-tuning a pretrained ResNet-18 model and comparing its performance to a smaller CNN that I trained from scratch. In order to assess generalisation, I will test the system on real-world photos rather than building a proprietary dataset. As advised, I'm transferring the work to Marimo and, in accordance with best practices, I've swapped out the old folder scanning code with pathlib.

[jmacey]

Good start to the project and it is good to see the code being developed. What are the initial results like?

[RoryMabina]

The initial results are positive. With consistently improving loss and accuracy curves, the custom CNN achieves test accuracy of about 48%. The confusion matrix and incorrectly identified samples show that it struggles with fine-grained distinctions but does well on more general categories like cars and animals because of its pretrained ImageNet features, the ResNet-18 transfer learning model performs noticeably better, converging more quickly and generalising more successfully. Usually, its mistakes have a stronger logical connection to the actual labels. Overall, the comparison verifies the project's direction and effectively shows the benefit of transfer learning for this.

[jmacey]

try to setup the project to use uv rather than pip (as it is much easier to work cross platform and in the labs this way).

You can do uv init followed by uv sync then a uv pip install -e Requirments.txt once and this will do it. Also Torch and Torchvision were comented out of the Requirments.txt It was also missing torchinfo

At present it doesn't fully run for me as it has issues with not finding the models directory FileNotFoundError: [Errno 2] No such file or directory: 'models/training_stats.npz'

[RoryMabina]

Thanks for the feedback. I’ve fully migrated the project to use uv with a pyproject.toml file (replacing Requirements.txt), ensuring torch, torchvision, and torchinfo are correctly locked for cross-platform support. To fix the FileNotFoundError, I updated src/train.py to automatically create the models/ directory if it’s missing and force pushed the pre trained weights and stats so the notebook runs immediately after a uv sync.