search

qubvel-org / segmentation_models.pytorch

Semantic segmentation models with 500+ pretrained convolutional and transformer-based backbones.
https://smp.readthedocs.io/
MIT License
9.58k stars 1.67k forks source link

Can't get more than 57% accuracy on different models #391

Closed VEXstar closed 3 years ago

VEXstar commented 3 years ago

Hello everyone, I'm new to machine learning, and as a start I decided to try this library. The problem is that I tried several models, encoders, loss functions, but none of the combinations gave more than 57% accuracy in validation. However, on the training sample, the accuracy reached 97%. As a dataset, I use multiple chest CT scans (2k slices in total) with matching masks. Masks have a binary format (0-background, 1-mask) Images are stored in a grayscale. I'm sure the problem is not related to this library, but I hope for help.

AlphonsG commented 2 years ago

Hello @VEXstar, did you resolve this?