Closed daniel-j-h closed 2 years ago
When running the script we now also get metrics for how bad the error is in meters.
python@ab66a7fc082f:~/app$ ./main.py data/1202213002.tif
params: 15649
step: 10, loss: 0.49938, mean: 404m, std: 295m, min: 0, max: 1627m
step: 20, loss: 0.49717, mean: 402m, std: 296m, min: 0, max: 1632m
step: 30, loss: 0.49476, mean: 401m, std: 295m, min: 0, max: 1636m
[..]
step: 1980, loss: 0.00290, mean: 30m, std: 24m, min: 0, max: 537m
step: 1990, loss: 0.00288, mean: 30m, std: 24m, min: 0, max: 538m
step: 2000, loss: 0.00287, mean: 30m, std: 23m, min: 0, max: 537m
Meaning: abs(predicted - actual)
for all pixels has an error mean of 30 meters, std dev of 23 meters, minimum of 0 meter, maximum of 537 meter. This is with the tiny model with ~15k parameters, that you see visualized above.
Maybe interesting to you @sowmyay. Closing here because not actionable.
https://user-images.githubusercontent.com/527241/158652652-e6e790c3-31d0-4234-8647-278bd204a1aa.mp4
https://user-images.githubusercontent.com/527241/158652661-a4b318e1-1edd-4e6e-a6da-08e1056e36e5.mp4
https://user-images.githubusercontent.com/527241/158652666-1b4d4a8d-e8cc-4a8e-a857-c8799d7f6554.mp4
https://user-images.githubusercontent.com/527241/158652674-915c0447-54d3-44b4-a5c1-9c99daca426f.mp4