Hi Daniel,

Thank you for the response! I am going to try multiple zoom levels to see if I can get that to work. My initial concern was running multiple at different zoom levels, zoom_level 21 (on it’s own) wasn’t training (MCC going to Inf) but zoom_level 20 (on it’s own) worked well. It is also possible the resolution of my image simply doesn’t get to the resolution equivalent for z21; I couldn’t make out why z21 wasn’t working giving me similar results when training and if there was a possible explanation.

I agree that training on the CPU is time consuming, I will try to get my hand on a GPU in the future. Thanks again for sharing this awesome project!

Hi Daniel,

Thank you for sharing this, kudos for the great instructions. I had a question regarding training with a different zoom level. I’ve been using a qgis add-in called QTiles to create my tiles on in-house imagery. I have not been successful getting the correct rasterio version to use your tiler.py unfortunately.

I’ve been successful running trainings and doing predictions at zoom level 20 on my CPU, however I have been running into issues at zoom level 21. In a nutshell, the MCC for the training and validation epochs go to Infinity (“Inf”). If I understand correctly, √ (TP +FP) * (TP + FN) * (TN + FP) * (TN + FN) must be equal to zero. I believe the issue lies in the zoom level but I am not 100% sure how to interpret the lack of conversion. I’ve change different parameters such as the the way I create the slippy tile, the way the “y” folders are split between training/validation/evaluation (randomized vs linear cut at the end) and the zoom level is what I suspect is causing issues. Initially I suspected blank tiles but now I am not so sure. I was wondering what your experience was with different levels (I read the manilla diary, in which he successfully runs it on zoom_level = 21).

Do you know what could be causing the lack of convergence for my MCC? I killed the training runs after a couple of epochs, could the lack of convergence be due to my lack of patience and further epochs could converge?

Thanks for all the work on this!