callumtilbury/bubble-distill-v2

🫧 Microbubble Distilled Student Model (v2)

TinyBubbleNet — a 389K parameter distilled student for microbubble segmentation.

Property	Value
Architecture	Depthwise-separable U-Net (Cellpose-compatible)
Params	388,580
Size	~4.6 MB
base_ch	16
Input	Grayscale microscopy images
Output	4 channels: dY, dX, cell_prob, dist_transform

Training Summary

• Teacher: Cellpose-SAM-FT (300M params, 1.1GB) — fine-tuned on 2-3 annotated microbubble images
• Dataset: 50 lab images from callumtilbury/microbubble-images
• Pseudo-labels: Generated by teacher on all 50 images
• Student: TinyBubbleNet(base_ch=16, depthwise=True)
• Epochs: 300 (best at epoch 282)
• Best val loss: 0.00083
• Training time: ~30 min on NVIDIA A10G

Pipeline

Cellpose-SAM-FT → pseudo-labels on 50 lab images → TinyBubbleNet distillation

Usage

import torch
from model import TinyBubbleNet

model = TinyBubbleNet(in_channels=1, base_ch=16, out_channels=4)
checkpoint = torch.load("best_model.pt", map_location="cpu")
model.load_state_dict(checkpoint["model_state_dict"])

# For inference, see:
# https://huggingface.co/callumtilbury/bubble-distill

Files

• best_model.pt — trained model weights (epoch 282, val_loss=0.00083)
• model_and_train.py — full training script with model, loss, dataset definitions

Why This Matters

Model	Params	Size	Inference @ 256²
Cellpose-SAM	~300M	1.1 GB	~100 ms
TinyBubbleNet	389K	~1.5 MB	~3 ms

~750× smaller, ~33× faster — specialized for your exact lab domain.

When image style changes, re-annotate 2-3 images, re-finetune the teacher, regenerate pseudo-labels, and re-train the student in ~30 minutes.

---

Distilled from Cellpose-SAM-FT using pseudo-labels on microbubble-images.