Improved robusteness of matching algorithm by taking into account lambdas as well
I tested with several random seeds
test_pythia_14m_build_graph_jacobian_stochastic has been failing for me with the new requirements in #327.
bumps the atol for get_rib_acts from 0 to 1e-12.
287 , tests have been xfailed for a while (and were vacuously true before)
test basis directly in addition to edges
raise truncation threshold, as the directions with very small lambdas are often very noisy (but we don't care about them).
don't include output layer in modular mlp test. For some reason the results are much more different when this is the layer that is/isn't rotated than when it's layers.2.
Removed functional edge and float32 tests (this is a test about basis similarity, I don't see why we would care about functional edges).
More robust tests
Description
Fixes a couple of problems with RIB tests:
329
test_pythia_14m_build_graph_jacobian_stochastichas been failing for me with the new requirements in #327.get_rib_actsfrom 0 to 1e-12.287 , tests have been xfailed for a while (and were vacuously true before)
outputlayer in modular mlp test. For some reason the results are much more different when this is the layer that is/isn't rotated than when it'slayers.2.Related Issue
Closes #329 . Closes #287 .
Does this PR introduce a breaking change?
No