You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Monte Carlo dropout appears to silently fail with the current release of pytorch-lightning.
To Reproduce
See this colab workbook. TCNModel instances with different dropout keywords produce identical CV results and inferences if given the same random_state, even with dropout set to 0.99, which should catastrophically underfit. I was able to reproduce this behaviour on my laptop using this python environment obtained by running pip install darts in a fresh Python 3.11 venv.
Expected behavior
The dropout keyword should impact model results. Extremely high dropout should result in severe underfitting. I was able to obtain this expected behavior in this environment by downgrading pytorch-lightning to 2.1.2. This can also be observed in colab
The text was updated successfully, but these errors were encountered:
eschibli
changed the title
[BUG] Monte Carlo dropout does not work with pytorch-lightning 2.2.1
[BUG] Monte Carlo dropout does not work with pytorch-lightning >=2.2.0
Apr 8, 2024
Monte Carlo dropout appears to silently fail with the current release of pytorch-lightning.
To Reproduce
See this colab workbook.
TCNModel
instances with differentdropout
keywords produce identical CV results and inferences if given the samerandom_state
, even withdropout
set to 0.99, which should catastrophically underfit. I was able to reproduce this behaviour on my laptop using this python environment obtained by runningpip install darts
in a fresh Python 3.11 venv.Expected behavior
The
dropout
keyword should impact model results. Extremely high dropout should result in severe underfitting. I was able to obtain this expected behavior in this environment by downgrading pytorch-lightning to 2.1.2. This can also be observed in colabThe text was updated successfully, but these errors were encountered: