search

jacobjma / PyQSTEM

A Python interface to the electron microscopy simulation program QSTEM
GNU General Public License v3.0
60 stars 30 forks source link

Column Heights #14

Open mragon2 opened 5 years ago

mragon2 commented 5 years ago

Hi, I read in your paper that the model is able to output the heights of the atomic columns, but when I try to load an image and run it through the CNN, the only output is the confidence map. The code for the calculation of the column heights has already been implemented? If so, how can I see the heights ? Thanks.

jacobjma commented 5 years ago

Hi, are you talking about this repo? https://gitlab.com/schiotz/NeuralNetwork_HRTEM

The column height algorithm is not currently implemented in a public repository. We have not evaluated the algorithm experimentally, but we are continuing to develop the method.

mragon2 commented 5 years ago

Good evening, Yes I am working on this model and reading this paper https://arxiv.org/abs/1802.03008. Ok, thank you very much for the update. So at the moment the published algorithm is only able to find the peaks right ? Thank you also for the solution of the issue I posted on the github.

Regards, Marco

On Tue, Nov 27, 2018 at 1:50 PM Jacob Madsen notifications@github.com wrote:

Hi, are you talking about this repo? https://gitlab.com/schiotz/NeuralNetwork_HRTEM

The column height algorithm is not currently implemented in a public repository. We have not evaluated the algorithm experimentally, but we are continuing to develop the method.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/jacobjma/PyQSTEM/issues/14#issuecomment-442193331, or mute the thread https://github.com/notifications/unsubscribe-auth/AqeA9EiqIbmuqwPSdsOetgmPnupYSmC2ks5uzZecgaJpZM4Yya6F .

jacobjma commented 5 years ago

Yes, the released implementation of the algorithm can only identify peaks. The reliability of the column height identification is still uncertain and needs more verification.

If you are committed to trying to make it work, I could share some of the experimental code. However, you are in for a potentially very big project.

I am currently just focusing on the peak finding, as this has been proven on experimental data.

mragon2 commented 5 years ago

Dear Jacob, thank you very much for your reply. I have a last question about the model. Are the images generated during training and visible in the folder 'debug' representative of the ground truth obtained through the Gaussian distribution? I tried to debug the file 'labels.py', to understand how the labels are defined and I found a good similarity between the numerical values of the labels and the values on the color bar of the generated pictures. I have a matrix of zero labels for the large part of the pixels, and values around 1 for the pixels where the atomic column should be located, same as the picture generated when running the file 'ktrain.py'. Moreover, I guess that the pictures show the same column height for each cluster since a small rotation is applied in the generation of the images in the file 'make_cluster_training_data_100.py'?. I tried to impose a large rotation and I have clusters with different values in the color bar. I guess this is because a rotation results in seeing columns with different number of atoms from the perspective of the xy plane.

Thanks, Marco

On Wed, Nov 28, 2018 at 3:08 PM Jacob Madsen notifications@github.com wrote:

Yes, the released implementation of the algorithm can only identify peaks. The reliability of the column height identification is still uncertain and needs more verification.

If you are committed to trying to make it work, I could share some of the experimental code. However, you are in for a potentially very big project.

I am currently just focusing on the peak finding, as this has been proven on experimental data.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/jacobjma/PyQSTEM/issues/14#issuecomment-442605218, or mute the thread https://github.com/notifications/unsubscribe-auth/AqeA9Ct8powRw25brnj109cmyNxolqvdks5uzvtPgaJpZM4Yya6F .