TensorBoard Explanation
If you have enable TensorBoard logging in the training settings, a TensorBoard page will be open up in your Web Browser, most likely with http://localhost:6006
Only appears if you have chosen to enable visualising training progress on the fly.
It shows the output of the current network’s prediction of the input image.
Since TensorBoard doesn’t support output 3D image stack, it only shows the top-layer of the input image prediction.
Learn Rate: The current learning rate.
VRAM Usage (MB): Video memory usage obtained using “torch.cuda.max_memory_allocated()”. The actual usage reported by your GPU could be slightly higher, due to the complexity of memory management.
Train Dice: The Dice Score between the network’s prediction and the ground truth label. Higher is better. Range from 0 to 1, 0 is the worst match, while 1 is the perfect match.
Train Loss: Essentially 1 - Train Dice. Note that Train Loss often goes up and down more frequently than Val Loss. This is due to the data augmentation used in training, which could sometimes generate hard-to-label samples.
Train Sensitivity: A.K.A True Positive Rate. Generally goes down at the first half of training, then slowly climbs back at the second half.
Train Specificity: A.K.A True Negative Rate. Generally climbs up steadily throughout the training.
Same as above, but is for the validation samples.
Since a network could over-fit on the training samples, the stats of training samples may not represent network’s performance on unseen data.
The network does not get trained on validation, so the stats obtained with validation samples are more representational of the network’s performance.
In general, you should aim for a validation dice score of over 0.8 for a satisfying segmentation result.