Frequently Asked Questions (FAQ)
❓ When I try to train the model there is a freeze/crash?
This can happen for several reasons, and we recommend posting on image.sc to get help troubleshooting deep learning dependencies and GPU drivers.
One common issue we've noticed is that training with the wrong model type can lead to problems. For example, in the screenshot below, the labels are drawn for semantic segmentation, but issues can occur if you accidentally select an instance segmentation framework instead.
Solution: Make sure you choose the right framework for your specific problem!
- Semantic Segmentation: Use frameworks like "Monai UNET Model" or "Pytorch Semantic Model"
- Instance Segmentation: Use frameworks like "CellPose Instance Framework" or "Micro-sam Instance Framework"
If the freeze/crash persists after verifying your framework choice, please share details about your setup on the forum, and the community will be happy to help! 😊
❓ I can't find my training and validation augmentations?
Sometimes there's confusion between labels and patches (augmentations). Let's clarify the difference!
Labels:
- These are the annotations you manually drew and marked with a bounding box
- Saved in the labels directory
- Validation labels won't be so indicated in the labels folder structure
Patches (Augmentations):
- These are the hundreds of small, augmented images generated from your labels
- Created when you run the augmentation step
- Saved in the patches directory
- when the validation labels are used to generate patches, those patches will have 'validation' appended to them.
Where to find validation data:
If you marked some labels as "validation" (by pressing the v key), they won't be marked as such in your labels directory structure. However, after you run augmentation, you'll find them marked 'validation' in the patches folder! Look for subdirectories like:
- input_validation0 for validation images
- ground truth_validation0 for validation labels
The patches directory is where all your training and validation augmentations live, ready for the model to learn from! 🎉