Closed dbuscombe-usgs closed 3 weeks ago
I was able to fix this here https://github.com/yotarazona/scikit-eo/issues/5#issuecomment-2136179777
I would suggest this way of importing keras to avoid those import errors
Hi @dbuscombe-usgs, thanks a lot for open this issue!. After some trial and error the package is running well now. tensorflow
, rasterio
and other required dependencies will be installed when scikit-eo
is installed. Additionally, scikit-eo
was tested on multiple machines in order to make sure everything is working out. Please see issue number #5.
Hi @yotarazona , I'm moving on with my side of the review, and I have a question about this.
Right now it installs everything, which is the most straightforward way, however, tensorflow makes the installation quite heavy.
I've seen in the past packages that have the possibility of using neural networks but where it's not a strict requirement treat it as an optional dependency, that can be installed if you do something like pip install scikeo[neural-nets]
.
I think it could improve the user experience, as long as it's documented in the installation instructions, although it depends a lot on what you expect the average user to be doing!
ping https://github.com/openjournals/joss-reviews/issues/6692
Tensorflow is not a listed dependency for your package. See https://github.com/yotarazona/scikit-eo/issues/6. This means that notebook 11 fails
I have tensorflow installed, and can verify using python and ipython terminals, as well as jupyter. I can import
import tensorflow as tf
with no errorHowever, when I attempt
from scikeo.deeplearning import DL
, I get the following errorModuleNotFoundError: No module named 'tensorflow.keras'
It appears the packaging issues I've already identified are at play here (i.e., these are all parts of the same problem). In this case, it seems to me that the from
scikeo.deeplearning
module is improperly linked against tensorflow, or perhaps the conda environment I had to create to run these codes was not done so optimally? For example, I regularly have to install tensorflow in some specific order.The packages I maintained usually spend quite a significant amount of time ensuring that all the dependencies are installable together in a stable way, usually tested on multiple machines. I recommend that here, coming up with a conda or virtual environment recipe that is known to work for all the provided functionality of this software.