Skip to content

FNO inverse pipeline fails with InitialConditionInterp when initial_step > 1 #100

Description

@Mattia-Marras

Description

The FNO inverse pipeline appears to fail when using InitialConditionInterp with initial_step > 1.

Reproduction

Run the following command:

python models/inverse/train.py +args=config_1DCFD ++args.filename=1D_CFD_Rand_Eta1.e-8_Zeta1.e-8_periodic_Train.hdf5 ++args.model_name=FNO ++args.base_path=../data/ ++args.initial_step=5 ++args.t_train=20 ++args.reduced_resolution=16 ++args.reduced_resolution_t=5 ++args.reduced_batch=200 ++args.batch_size=1 ++args.num_workers=0 ++args.num_channels=3 ++args.inverse_model_type=InitialConditionInterp ++args.inverse_epochs=20 ++args.inverse_learning_rate=0.01 ++args.in_channels_hid=4 ++args.num_samples_max=10 ++args.inverse_verbose_flag=true

Observed behavior

The script fails before the inverse optimization starts with:

UnboundLocalError: local variable 'model' referenced before assignment

After fixing the loader initialization, the same command proceeds further but fails during the FNO forward pass with:

RuntimeError: mat1 and mat2 shapes cannot be multiplied (64x4 and 16x20)

Expected behavior

The inverse pipeline should support FNO checkpoints trained with initial_step > 1 without raising shape mismatch errors.

Additional information

The second error suggests that the inverse reconstruction and the FNO forward model expect different input representations for the initial condition when initial_step > 1.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions