Skip to content

Fix export issue - json now in agreement with pt

Nicole Schulte requested to merge export_fix into master

Fix in the exporting of the model

Since the function in load_state_dict changed to model_arch_optim, the function assign_weights needs to change as well. There one needs to remove the get_normed_weights (implemented in the package source code) function, otherwise the weights get normed twice. Now the pt prediction on a vector agrees with the json for ALL lipschitz constants

For bookkeeping:

  • The error made itself visible, when testing the pytorch model as well as the json model on a given vector
  • for Lambda = 1, one could not see the issue and everything was in agreement
  • for Lamda != 1, inconsistencies arose between the predictions which worsend with higher values of lambda

Merge request reports