Retraining for new microsccope #76
Description
From Ben:
On another note, as I mentioned we recently upfitted a manual microscope with an XY stage, Z focus drive, and a cube for brightfield reflected imaging. The new system has lower quality optics, and a lower quality camera with different pixel resolution, unfortunately. When I originally tried to analyze the dataset, it missed about half of the apartments and couldn't find any markers in some images, so I had to retrain a marker model for it, which I uploaded to your GS bucket at the location:
Buckets / celldom / models / r0.9
However, even after I did that I found that it only corrected identified 90% of the apartments, so I think there may be an issue with either the new marker model of the digit classifier, but I'm not sure which one is the problem. I have never trained a single digit classifier before, and I'm not sure whether the code is not finding the markers or is incorrectly calling digits. If you have a chance, could you take a look at the dataset and output file in the folder:
Buckets / celldom / datasets / 20190415-NoCell-ML01-Imaging-Test
I also put the config files and other details if that helps you troubleshoot. This is actually a good exercise, as the imaging conditions are very different between the two systems, which if we are able to solve will make the software more robust. If you'd like me to upload the cross marker training data as well, I can do that.