Closed chuckwondo closed 4 months ago
@nemo794 and @arthurduf, I did some investigation into this issue, and as I suspected, the difference in outputs between NASA and ESA (including the reason for the exception described in #7) is due to differences in Python package dependencies.
Here's what I did to confirm this:
Dockerfile
to install conda
and install dependencies from environment.yml
instead of using pip
to only install a couple of dependencies.np.linalg.pinv
back to np.linalg.inv
, which would cause the ESA version to fail again, if the change in dependencies did not have an impact.-156 18.8 -154.7 20.3
, which is the bbox that caused the exception in ESA (per #7).cksum
between the NASA and ESA dem.tif
files, and found them to be identical.Therefore, I suggest that in order to be confident that running an algorithm in both systems with the same inputs will produce the same outputs, we must use the same dependencies and install them with the same dependency manager in both environments.
@arthurduf, is there any reason you chose not to use conda
in the Dockerfile
? Is there anything in the ESA environment that would prevent you from doing so?
Due to #7, it was discovered that, given the same inputs, there are discrepancies between the results produced by running the algorithm in the NASA and ESA environments.
In addition to using this algorithm to gather profiling metrics from the 2 environments, we also want to make sure we can produce the same results.
See attached slides detailing @nemo794's initial investigation into the output discrepancies: NASA-ESA get-dem discrepancy.pptx