Week 2 (Jun 09 - Jun 15) | Sakhineti Praveena

Meeting with Mentors

We discussed about the progress made in the previous week where I have refactored dataset and model classes. I also raised an issue I encountered during testing: in one of the examples, the inference output appeared distorted compared to earlier results. David mentioned he would look into it on his end(which was fixed later that week through Issue #314).

Another topic was the current file structure. Right now, both the dataset and model directories have a lot of files, and David suggested we may consider restructuring them later as the project evolves. Overall, the feedback from mentors was positive, and we moved ahead with planning goals for the coming week.

This Week’s To-Do

Add support for COCO dataset
Add support for Torch detection models (Partially done)
Explore Streamlit if time permits

Progress

COCO Dataset Support

I started by creating coco.py, which includes a CocoDataset class derived from ImageDetectionDataset.
I was spending some time to check if I should use pycocotools or write our own parsing logic. Some GitHub issues pointed out that the official pycocotools is not Windows-friendly.
We found a workaround using pycocotools-windows for Windows machines. After discussing with mentors, we decided to handle installation conditionally. David also confirmed by installing the library in his windows machine. I have handled the installation in pyproject.toml like shown in the code below using this snippet:

 pycocotools = { version = "^2.0.7", markers = "sys_platform != 'win32'" }
pycocotools-windows = { version = "^2.0.0.2", markers = "sys_platform == 'win32'" }

Within the CocoDataset class:
- _build_ontology() maps COCO category IDs to the framework’s format.
- _build_dataframe() prepares the structure used by the base class.
- read_annotation() processes annotations for each image based on image_id.

Torch Model Refactor

I renamed torch.py to torch_segmentation.py and updated the imports across all example scripts accordingly.
Then I created torch_detection.py to support detection models.
Taking torch_segmentation.py as a reference, I modified:
- data_to_device()
- get_data_shape()
- get_computational_cost()
to account for the different input format used in detection: List[Dict[str, Tensor]].
I wasn’t sure if unsqueeze_data() and the CustomResize class were necessary for detection, so I skipped them for now.
I was unwell on Tuesday and Wednesday and couldn’t make as much progress as I had planned. I couldn’t fully finish torch_detection.py this week. However, I’ve committed the partial changes in the draft PR #312.

Going ahead

Next week, I’ll aim to:

Complete the Torch detection model integration
Add the metrics evaluation part for detection
Add an end-to-end example with COCO + Torch detection