Loss Curves

Loss measures how well the network predicts masked pixels — lower is better.

The loss chart shows two curves

Train Loss (red)

Error on the patches used for learning. Should decrease steadily as the model learns.

Val Loss (teal/cyan)

Error on held-out patches not used for training. Shows how well the model generalizes.

The colors follow the module theme (train uses the primary red, validation uses the info teal) and adapt automatically to light and dark mode.

Both curves decrease together, then flatten. A small gap between them is normal.

Training loss keeps decreasing but validation loss starts rising. The model is memorizing training data. Early stopping will typically catch this.

Both losses stay high and don't decrease much. Try longer training, larger model, or different learning rate.

Losses jump around erratically. Try reducing learning rate or increasing batch size.

Note: Unlike segmentation, there's no Dice score for denoising. Loss is the primary metric — visual inspection of results is important.