Watershed End-to-end (WN = DN + WTN)

qAp / sartorius_cell_instance_segmentation_kaggle

Solution for Sartorius Cell Instance Segmentation Kaggle

0 stars 0 forks source link

Watershed End-to-end (WN = DN + WTN) #10

Open qAp opened 2 years ago

qAp commented 2 years ago

End-to-end = DN + WTN

https://github.com/min2209/dwt/tree/master/E2E

qAp commented 2 years ago

[x] How to load pretrained DN and WTN from lightning checkpoints.

qAp commented 2 years ago

End-to-end fine-tuning: We cascaded the pre-trained models for the DN and WTN and fine-tuned the complete model for 20 epochs using the RGB image and semantic segmentation output of PSPNet as input, and the ground truth distance transforms as the training target. We use a batch size of 3, constant learning rate of 5e-6, and a L2 weight penalty of 1e-6.

qAp commented 2 years ago

Loading pretrained DN and WTN before training, WN's metric and loss are improved.

qAp commented 2 years ago

[ ] Re-train after DN has been re-trained using semseg[..., [0]] + semseg[..., [1]].

qAp commented 2 years ago

[x] Build inference script.