The training dataset for custom DECODE model

[Hankeun-Lee]

Hi, I am a researcher at an university and first of all, thank you for your outstanding works!
I am very interested in applying DECODE model to one of our problems, which is to localize nanoparticles from an image acquired by a microscope we developed.
Since the dataset is different from that of yours, I decided to rebuilt the model following your code.
In our dataset, we do not have frames, but rather single images at a given time.
So, I decided to apply a single UNet with multi-head outputs, including intensity, background, prediction probability, and the uncertainties.
It seems like the output image of the uncertainties and prediction probability closely follow the output of your original model.
However, I am having trouble getting the right intensity output.
In my dataset, the intensity of a PSF ranges from 2 to 4, but I can see that you applied sigmoid to the intensity output, which forces the values between 0 to 1.
So my questions are

1. what are the values of your intensity labels? are they normalized between 0 and 1 during the PSF analysis?
2. My implementation of the DECODE model seems to overfit to training data. It seems like you included dropout in the model but i am not sure if the dropout layer is actually used for training.

Any valuable insight is highly appreciated. Thank you in advance for your time.

[Haydnspass]

Dear @Hankeun-Lee,
I appreciate your interest in DECODE. Indeed, we rescale the predictions (including the intensity) to roughly achieve a predicted target value in the 0...1 regime. You can rescale by a constant factor after the model prediction. The scaling factor is either auto-determined or set by the user in the config file.

2. Currently, we don't use dropouts since we simulate the training data on the fly and only use them once.

Best,
Lucas

[Hankeun-Lee]

Thank you Haydnspass. Your answer helped me a lot!