Progress Board

[eddogola]

Moving onto the composable multi-stage finetuning approach as described here.

Given the following model choices,

• Object detection
  • FCOS
  • SSD
  • Yolov8(w/ seg baked in)
• Segmentation
  • UNet
  • Squeeze UNet

I have decided to proceed with an initial architecture of FCOS(one pass, no anchors) and Squeeze UNet, which I'd initially implemented here

[eddogola]

Mon Feb 26, 2026, In office

Plan

• implement FPN, FCN layers (of the yolact model) in nn.Module s.t. they can be ONNX'd
• hopefully fully write the yolact model
• Note: this porting is being done mostly using the author's implementation

Pull Request: #10

TODO: fix bugs; test: inference runs successfully on natural image and can train unfrozen wrapper layers

[eddogola]

Tue Feb 27, 2026, In office

Plan

• Figured I have a knowledge in regards to the intricacies of FCN and FPN; so, I am going through said papers to deepen my understanding
• fix bugs from yesterday; wrap all up in working yolact model

Pull Request: #10

[eddogola]

Thur Apr 25, 2026

• moved from colab to local because of its ephemeral nature -- can't keep track of code in addition to notebook without getting into setting up google drive and github together, too cumbersome. Will, therefore, revert to local version control. Only concern is training speed, hence feedback loop length, but hinging my hopes on Apple Silicon's MPS device which Torch now supports.
• wrote initial implementations of regression layers
  Currently
• writing supporting code to freeze relevant parts of the network
• adding regression modules into the model's forward pass and ensuring torch.grad is True for just them and false for the rest
• fix bugs from yesterday; wrap all up in working yolact model
  TODO/urgent:
• lab data setup: The model expects the data in a very specific format so I should write a Dataset class that provides just that, without changing any model implementation details.
  Concerns
• not thinking about NMS or post processing for now, should get to it later.

Pull Request: #10