CircleNet Hourglass: reconstructed MoNuSeg
hourglass_coco_best.pth is the two-stack Hourglass checkpoint trained on the
reconstructed MoNuSeg splits.
Use it with CircleNet, not a Transformers pipeline.
COCO initialization; Adam, batch size 4, learning rate 0.00025, 50 epochs. Lowest validation loss selected epoch 33. Standard test evaluation on 140 patches gave 44.2 AP / 82.1 AP50 / 43.3 AP75.
This was a single RTX 3090 run with modern PyTorch/CUDA. New crops, splits and
XML-derived labels differ from the historical benchmark; these are not
reproduced paper scores. No DLA checkpoint or rotation-consistency result is supplied.
test_previews.jpg shows model predictions, not ground-truth annotations.
Checkpoint SHA-256:
2744597be28071b738243ac9f3113b16b700537becde47e7646bc4f552bb0db7
Usage, training details and limitations. CircleNet code is MIT; training data are CC BY-NC-SA 4.0. This rehosting grants no additional rights to third-party code, weights or data.