...
- Patch size: Together with "Batch size", will change how much data is passed to the GPU per batch. A patch of 256x256 with a batch size of 16 works best for our workstation (the one next to the spinning disk microscope)
- Number of training epochs: Increase this if your loss is still decreasing after 100 epochs.
- → Jump to 5. if you already have a trained model
- Choose a name for the model
- Select the path where to save the model
- "Number of training epochs" and "Batch size" can be left as default (if there is a memory error, for the training, the batch size must be reduced).
- Click on "Start training" and go for a break. This should last 10-15 min.
Image Removed
4. Look at the result with "Preview Result" and use the navigation pane to observe the before/after
...
- Steps per epochs: You can increase this parameter if you want your network to train on a larger portion of data for each epoch.
- Train fraction: Indicates which portion of the dataset is used for training, the rest being used for validation. Increase this to 0.99 if you have a large dataset and you want to avoid the validation step to take too long (there is one validation step per epoch)
- Neighborhood radius: N2V parameter. Read more about it in their paper https://arxiv.org/abs/1811.10980
Image Added
4. Apply a trained model to a dataset by going to "Apply N2V"
Image Added
- By default, the program will upload the images in the same dataset by adding the suffix "_N2V" to the image name
- Images can be uploaded during the processing. Otherwise, they can be uploaded when the processing is done with "Upload to omero"
- Result display can be toggled off to make the process faster.
5. Apply the model to the dataset and upload results to OMERO → Start with this step if you already have a trained model for your images