A PyTorch-based library for working with 3D and 2D convolutional neural networks, with focus on semantic segmentation of volumetric biomedical image data
MIT License
160
stars
27
forks
source link
Various improvements by @Optiligence and @mdraw #40
Summary (see individual commit messages for more details):
Rebase and integrate new multi-class/multi-label-specific code by @Optiligence
Includes new experimental losses, support for masking labels for loss and metric calculation, more detailed tensorboard logging, training code changes etc.
Most of it is integrated through new options and checks, with minimal to no changes to previous default behavior
Refactoring incompatible/intrusive changes into a Trainer subclass (elektronn3.training._trainer_multi.TrainerMulti)
Fixes for tensorboard image plots (better alpha blending and more consistent color mapping), additional overlay plots. Note that as a side effect this changes the class-to-color mapping in tensorboard for overlay plots.
Remove experimental autoencoder pretraining feature from UNet due to BC issues
Write more concise log messages, less warning spam
Disable custom "adaptive" convolution operators in UNet by default because they are no longer needed and have increased the model code complexity
Minor improvements for transforms, add transform support in elektronn3.inference.Predictor
Summary (see individual commit messages for more details):
UNet
due to BC issuesUNet
by default because they are no longer needed and have increased the model code complexityelektronn3.inference.Predictor