search

histocartography / hact-net

MIT License
70 stars 14 forks source link

Val Loss increases #2

Closed csccsccsccsc closed 3 years ago

csccsccsccsc commented 3 years ago

When I train HACT-Net (or the cell graph / tissue graph net), the validation loss keeps increasing and the accuracy keeps unchanged (very low).

guillaumejaume commented 3 years ago

I would need more information to help you here,

csccsccsccsc commented 3 years ago

1) I train on the BRACSv1 dataset (".../BRACS_RoI/previous_versions/Version1_MedIA").

2) I try the pre-trained model, but I got: Test weighted F1 score 0.4220916765247371 Test accuracy 0.46166134185303515

3) I have tried learning rate 0.1 / 0.001 / 0.0005 / 0.0001 / 0.00001. I only wait for around 10 epochs. In my experiments, the "val weighted F1 score" and the "val accuracy" keeps unchanged (usually around 0.04527802507158154 and 0.16220735785953178).

csccsccsccsc commented 3 years ago

Will this caused by the different graph features? Could you send me your processed graph features?

csccsccsccsc commented 3 years ago

I found that there exist some empty images in the website: train/5_DCIS/BRACS_1478_DCIS_24.png, train/5_DCIS/BRACS_1478_DCIS_30.png, train/5_DCIS/BRACS_295_DCIS_31.png train/6_IC/BRACS_280_IC_14.png val/6_IC/BRACS_1638_IC_13.png

guillaumejaume commented 3 years ago

Thanks for reporting the issue with empty images. We're working on it, and will fix it as soon as possible.

guillaumejaume commented 3 years ago

You can download the preprocessed cell, tissue and hact graphs for the BRACS dataset here:

Or by downloading this zip file that includes the test cell graphs:

guillaumejaume commented 3 years ago

Empty images updated on the FTP server, thanks for reporting the issue.