4.1 Answers#

How many pixels are masked per patche?

It depends on the patch size, but the configuration method has a parameter called n2v_perc_pix which by default is equal to 0.198%. With a patch size of 64 pixels, that leads to: \(64^2 * 0.2 / 100 = 8\) pixels.

What are the dimensions of the training set? What is the last dimension?

Default is (512, 64, 64, 1). The channel dimension is at the end. If your images are multichannel, they should end up there. But careful! The channel dimension is necessary due to the network architecture. If you have multiple channels, they will be aggregated and the network will learn noise model and image features from all channels together. That is usually not a good idea.

The reason to have the channel dimension is to be compatible with RGB images.

The configuration is printed to the console on the notebook. Yes yes, it is the very long one-liner that starts with N2VConfig.

Here are a few questions:

1. What do you think "means" and "stds" are and what they are used for?
2. What is the value of the UNet depth? And what is that?
3. Can you find the loss name?

Here are the answers:

1. The means and stds are the means and standard deviations of the training dataset pixel values. They are used to normalize the patches. This allows to have pixel values in similar ranges across patches, leading to more stable training.
2. The UNet depth is 2 by default. This is the number of downsampling and upsampling that the UNet performs. Remember the "U" shape.
3. The loss is mse, which stands for mean square error.