I changed my workflow

[huangchieh]

I spent lots of time making the augmentations method on the fly. To make this work, a trained generator needs to be added when training. I intended to make it user friendly. Image to add one line of code to use the trained model.But it’s painful for me.

Then I will leave this idea there at this point. And make the offline dataset at first. And generate new datasets in the previous format. In this way no existing code needs to be changed.

If it works, I may come back to this on-the-fly idea.

[huangchieh]

Todo list:

• Find the code of Real-Simulation classifier and make it a regression model. See SimulationRealClassifier
  • Add noise to the training and testing steps of machine expert. -- Suggested by Niko
  • See the failed classification. -- Suggested by Adam

• Open the black box using CAM machine learning. -- By Phillippo.
• By changing the $\lambda$, generate a set of fake AFM images. This Fake-AFM-Generator is needed here. You have done most of job in this folder (CycleGanAFM).
• Find the candidate
• Zip the data in previous format. (/scratch/phys/project/sin/AFM_Hartree_DB/AFM_sims/striped/Water-Au111-FB)

[huangchieh]

SIN group meeting 03 suggestions:

1. More paired images can be obtained from create more PPAFM images for given AFM. (Nian)
2. More AFM images can be supported from Nan, the experimentalist. (Nian)
   Nan provided some example AFM images I can play around (/scratch/phys/project/sin/Jie/Github/CycleGanAFM/largeDataSet/2022-Alkyne-Porphines).
3. Only learn noise, rather than the whole noised image. The noised image = y_0 + \Delta y + noise, where \Delta y is the output of Gan. Noise can be obtained by analyzing the background of image. As for how to output the noise image, I can reference to resnet, auto-encoder. (Niko)
4. I can try another simplified version of data augmentation, by just adding experimental noise from the previous points. (Niko)
5. As for the inconsistency problem, I can try to select two or three slices from all the slice. (Niko)
6. As for the scale problem, Joakim suggested two papers, link1 and link2

Kiitos paljon!