emdgroup / tnmf Goto Github PK

...after the repo became public and GitHub pages have been set up

I am getting some flake8 errors locally that do not show up in the Github actions:

./build/lib/tnmf/tests/test_init.py:10:5: F403 'from tnmf import *' used; unable to detect undefined names
    from tnmf import *
    ^
./build/lib/tnmf/tests/test_init.py:10:5: F401 'tnmf.*' imported but unused
    from tnmf import *
    ^
1     F401 'tnmf.*' imported but unused
1     F403 'from tnmf import *' used; unable to detect undefined names
2

Currently, the inhibition gradient is computed in a way that it suppresses neighboring activations. Intuitively, we might instead want to suppress activations that lead to overlapping atoms in the reconstruction. Thus, another inhibition mode might make sense, that ensures pairwise orthogonality of the resulting partial reconstructions.

...to prevent smeared out Activations.

See https://academia.stackexchange.com/a/172780, https://citation-file-format.github.io, https://github.com/citation-file-format/citation-file-format/blob/1.1.0/README.md#software-citation-metadata-required

example: https://github.com/emdgroup/brain_waves_for_planning_problems/blob/main/CITATION.cff

Need to think about how to handle even atom sizes.

Presumably, the convolution, that is used to compute the inhibition gradient, does not consider the reconstruction mode. Potentially, this is not correct: At least intuitively, I would expect that for circular boundary conditions an activation at a border should suppress activations at the opposite border.

@AdrianSosic : What do you think?

The current setting how the different fitting interfaces (batch, minibatch, stream) are implemented has several problems:

• The variable naming is not consistent / misleading but cannot be changed (e.g. the streaming method also processes minibatches, but the corresponding kwargs cannot be renamed as such because the higher level fit function uses the names to differentiate between the methods).
• Switching between the different functions is not straightforward (e.g. it requires adding an additional algorithm kwarg), which requires if-else logic in the tests.

A potential solution might be to create a higher-level abstract Algorithm class that has the current MiniBatchAlgorithm as a subclass together with others that cover the batch and streaming scenario.

Should be analyzed and monitored.

The streamlit sharing demo simply crashes when exceeding the memory resources of the provided machine. Maybe we can identify during runtime if the demo is executed locally or in the cloud. In the latter case, we could then show a hint to the user that this is the "online version" of the demo, which is running with limited resources, and prevent the execution for settings that would lead to a crash.

Currently, we would not recognize if these fail, e.g. due to interface incompatibilities.

Easiest might be to create dummy unit tests that set matplotlib into non-interactive mode and run the examples just to check that nothing fails horribly (no result verification).