Atypical background inferences

[cleterrier]

HI, I've made a trained model for PAINT with far-red imagers, but when using it to fit the experiment (from which the simulation parameters were measured), I get strange background inferences. Sometimes it's just a single value at 144 photons for all localizations, whereas I trained with 50 to 400 background, as seen on the param_run_647.yaml file here

For the other channel (red imager), I get a broader distribution of background values but sometimes they are clipped at 144, whereas I trained with 50 to 250 background, as seen on the param_run_561.yaml file here

What happens here? Seems to me that the network is incapable of inferring backgrounds above 150, despite being trained on larger values?

[cleterrier]

Wait, I just realized there's a
bg_max: 144.0
line in the param_run.yaml files.

Why is this? Why is it not overridden by the background range specified from processing the experiments with SMAP and getting the values from there?

Similarly, the line
phot_max: 13000
could potentially create problems for PAINT as sometimes we get high photon numbers.

So... should I redo both trainings (547 and 561) after changing these lines to something like
bg_max: 400.0
phot_max: 30000
to avoid these problems, or are they just used at fitting?

@ASpeiser @Haydnspass

[cleterrier]

Also I've found this which is helpful to explain some of the yaml parameters:
https://github.com/TuragaLab/DECODE/blob/67b8331758202c2d87d4501186e21be0971e018b/decode/utils/reference_files/reference.yaml

[Haydnspass]

@cleterrier You do not need to use the training notebook. But make sure you don't use an old param_run.yaml and rather an old param_friendly.yaml because this is the set of parameters you typically need to change.
The parameters are filled up by the reference param at training launch if they are not present in your specified param file; i.e. if you provide Scaling parameters in your yaml when launching a training these parameters will not be overwritten.

[cleterrier]

OK I see. If I want to have additional adjustments to parameters that are not in the param_friendly.yaml file, I can add them to this file and they will be used when generating the full file? I'm thinking about the learning rate for example

[Haydnspass]

Yes exactly :)

[cleterrier]

After putting the right values for bg_max and retraining it works much better! PAINT really has more background :)