Noise2Void validation convenience functions

[jdeschamps]

Problem

Since N2V is self-supervised, it is not clear to many users how to properly validate the results. In the docs, we highlight the importance to examine the following:

• autocorrelation to inspect structured noise in the data
• residuals after training to see whether any spatial information has been removed, spatial information being not 0-mean as opposed to Poisson or readout noise.

Currently, the autocorrelation is not compatible with stacks (#298), and there is no convenience function to compute the residuals or examine their statistics.

Finally, residuals could be logged during training (idea raised by @conradkun).

Potential solution

• Fix Autocorrelation function compatible with stacks #298
• Compute residuals in Improved training tracking: compute metrics, save predictions #294
• Explore window-averaging the residuals to explore areas with non-0-mean statistics, validate on poorly trained models and models with good performances on the benchmarks.

Linked issues

#294: log prediction (potentially residuals) during training
#298: extend autocorrelation function to stacks