GSoC - Coding Period Week 5

Work Done This Week (July 5th to July 12th)

• Refactored the training loop for the nucleus segentation model, this enabled faster prototyping and experimentation.
• Integrated Optuna for automated hyperparameter optimization. Ran 200 optuna trials to find training hyperparams.
• Optuna samples hyperparams from a given range, which in this case was -
  • Learning rate: 0.5e-3 to 20e-3
  • Batch Size: 8 to 64
• Each trial trained the model on 10% of available data for 3 epochs, and returned the resulting IOU score.
• The hyperparams from the best optuna trial is shown in the image below

• As a recap, the model details are :

  • Type - Feature Pyramid Network
  • Backbone - ResNet-18

• Trained the model for 50 epochs, the training metrics are showcased below.

• The model seems to perform well, even when the inputs are super noisy/underexposed. The GIF below shows model predictions on 2D slices of 3D data at one point in time.

• More inference examples:

• Feel free to check out this work:
  • Github - Link
  • NBViewer - Link

Planned:

• Build a GUI for the online demo.
• Add this model to the DevoLearn library.
• Swap out the existing DevoLearn cell membrane segmentation model with the new upgraded model.