Rename interpolator to temporal downscaler by dnerini · Pull Request #171 · MeteoSwiss/evalml

dnerini · 2026-06-06T19:08:51Z

In anemoi, the task formerly called "interpolator" has been officially renamed to "temporal downscaler" (see anemoi training docs). evalml uses the anemoi model type concept throughout its config schema, workflow rules, Python code, and config files. This rename keeps evalml aligned with upstream anemoi terminology.

frazane

LGTM, ready to merge if experiment and showcase both work on the example config!

MicheleCattaneo

This look good to me. I guess it's mostly a refactor of names, correct?
The only main doubt I have is regarding the comment I made for the start/end/step formatting of the forecasting setting.

jonasbhend

I still get an error when running the temporal-downscaler config:

2026-06-09 08:38:33 INFO GRIB template summary:
2026-06-09 08:38:33 INFO   - SamplesTemplates: FI, PMSL, PS, QV, T, TD_2M, T_2M, U, U_10M, V, V_10M
2026-06-09 08:38:33 INFO Post processor: ExtractMask(lam_0/cutout_mask, points=1147980/slice(np.int64(1147980), None, None))
2026-06-09 08:38:33 INFO Post processor: AssignMask(mask=global/cutout_mask, points=542080/542080, fill_value=nan)
2026-06-09 08:38:35 INFO GRIB template summary:
2026-06-09 08:38:35 INFO   - SamplesTemplates: FI, PMSL, PS, QV, T, TD_2M, T_2M, U, U_10M, V, V_10M
2026-06-09 08:38:36 INFO Temporal downscaling step (2025-03-01 00:00:00): 1 second.
Traceback (most recent call last):
  File "/user-environment/bin/anemoi-inference", line 12, in <module>
    sys.exit(main())
             ^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/__main__.py", line 38, in main
    cli_main(__version__, __doc__, COMMANDS)
  File "/user-environment/lib/python3.12/site-packages/anemoi/utils/cli.py", line 266, in cli_main
    cmd.run(args)
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/commands/run.py", line 60, in run
    runner.execute()
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runners/temporal_downscaler.py", line 166, in execute
    for state_idx, states in enumerate(
                             ^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runner.py", line 232, in run
    yield from self.prepare_output_state(
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runner.py", line 276, in prepare_output_state
    for state in output:
                 ^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runners/temporal_downscaler.py", line 290, in forecast
    y_pred = self.predict_step(self.model, input_tensors_torch)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runner.py", line 379, in predict_step
    return model.predict_step(input_tensors_torch, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/interface/__init__.py", line 241, in predict_step
    return self.model.predict_step(**predict_kwargs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/models/base.py", line 332, in predict_step
    y_hat = self.forward(x, model_comm_group=model_comm_group, grid_shard_shapes=grid_shard_shapes, **kwargs)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/models/interpolator.py", line 187, in forward
    x_data_latent, x_latent = self.encoder[dataset_name](
                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 593, in forward
    x_dst = super().forward(
            ^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 483, in forward
    return self.mapper_forward_with_edge_sharding(**kwargs_forward)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 384, in mapper_forward_with_edge_sharding
    out_dst[dst_chunk] = maybe_checkpoint(
                         ^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/utils.py", line 52, in maybe_checkpoint
    return checkpoint(func, *args, **kwargs, use_reentrant=False)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/_compile.py", line 32, in inner
    return disable_fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 632, in _fn
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/utils/checkpoint.py", line 496, in checkpoint
    ret = function(*args, **kwargs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 321, in run_processor_chunk_edge_sharding
    x_src_chunk, x_dst_chunk, _, _ = self.pre_process(
                                     ^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 570, in pre_process
    x_src = self.emb_nodes_src(x_src)
            ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/linear.py", line 125, in forward
    return F.linear(input, self.weight, self.bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: mat1 and mat2 shapes cannot be multiplied (41092x81 and 158x1024)
srun: error: nid002828: task 0: Exited with exit code 1
srun: Terminating StepId=226333.0

jonasbhend · 2026-06-09T06:11:45Z

    mem_mb_per_cpu: 1800
    runtime: "1h"
    gpus: 0
+    account: s83


Should we add this to all configs?

MicheleCattaneo · 2026-06-09T07:10:00Z

I still get an error when running the temporal-downscaler config:

2026-06-09 08:38:33 INFO GRIB template summary:
2026-06-09 08:38:33 INFO   - SamplesTemplates: FI, PMSL, PS, QV, T, TD_2M, T_2M, U, U_10M, V, V_10M
2026-06-09 08:38:33 INFO Post processor: ExtractMask(lam_0/cutout_mask, points=1147980/slice(np.int64(1147980), None, None))
2026-06-09 08:38:33 INFO Post processor: AssignMask(mask=global/cutout_mask, points=542080/542080, fill_value=nan)
2026-06-09 08:38:35 INFO GRIB template summary:
2026-06-09 08:38:35 INFO   - SamplesTemplates: FI, PMSL, PS, QV, T, TD_2M, T_2M, U, U_10M, V, V_10M
2026-06-09 08:38:36 INFO Temporal downscaling step (2025-03-01 00:00:00): 1 second.
Traceback (most recent call last):
  File "/user-environment/bin/anemoi-inference", line 12, in <module>
    sys.exit(main())
             ^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/__main__.py", line 38, in main
    cli_main(__version__, __doc__, COMMANDS)
  File "/user-environment/lib/python3.12/site-packages/anemoi/utils/cli.py", line 266, in cli_main
    cmd.run(args)
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/commands/run.py", line 60, in run
    runner.execute()
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runners/temporal_downscaler.py", line 166, in execute
    for state_idx, states in enumerate(
                             ^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runner.py", line 232, in run
    yield from self.prepare_output_state(
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runner.py", line 276, in prepare_output_state
    for state in output:
                 ^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runners/temporal_downscaler.py", line 290, in forecast
    y_pred = self.predict_step(self.model, input_tensors_torch)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/inference/runner.py", line 379, in predict_step
    return model.predict_step(input_tensors_torch, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/interface/__init__.py", line 241, in predict_step
    return self.model.predict_step(**predict_kwargs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/models/base.py", line 332, in predict_step
    y_hat = self.forward(x, model_comm_group=model_comm_group, grid_shard_shapes=grid_shard_shapes, **kwargs)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/models/interpolator.py", line 187, in forward
    x_data_latent, x_latent = self.encoder[dataset_name](
                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 593, in forward
    x_dst = super().forward(
            ^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 483, in forward
    return self.mapper_forward_with_edge_sharding(**kwargs_forward)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 384, in mapper_forward_with_edge_sharding
    out_dst[dst_chunk] = maybe_checkpoint(
                         ^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/utils.py", line 52, in maybe_checkpoint
    return checkpoint(func, *args, **kwargs, use_reentrant=False)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/_compile.py", line 32, in inner
    return disable_fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 632, in _fn
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/utils/checkpoint.py", line 496, in checkpoint
    ret = function(*args, **kwargs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 321, in run_processor_chunk_edge_sharding
    x_src_chunk, x_dst_chunk, _, _ = self.pre_process(
                                     ^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/anemoi/models/layers/mapper.py", line 570, in pre_process
    x_src = self.emb_nodes_src(x_src)
            ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/user-environment/lib/python3.12/site-packages/torch/nn/modules/linear.py", line 125, in forward
    return F.linear(input, self.weight, self.bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: mat1 and mat2 shapes cannot be multiplied (41092x81 and 158x1024)
srun: error: nid002828: task 0: Exited with exit code 1
srun: Terminating StepId=226333.0

This issue comes from the checkpoint being trained with a wrong default value from anemoi training.
multistep_in used to be 1 by default and was not used by anemoi training, only by inference. It should be set to 2, which was done for another checkpoint, with runid f9279244ed6f4c458597bdcf335ab36f. This value is both in the metadata and a field in the model's object (questionable design).
@hlasco did manually fix this checkpoint (eb534e4c9ca5408e8b83ace2f268bb29), if I recall correctly. Do you still have it?

dnerini · 2026-06-09T08:22:09Z

@jonasbhend @MicheleCattaneo thanks, I reverted back to the previous checkpoint see 039a9da

In anemoi, the task formerly called "interpolator" has been officially renamed to "temporal downscaler" (see [anemoi training docs](https://anemoi.readthedocs.io/projects/training/en/latest/modules/train.html#available-tasks)). evalml uses the anemoi model type concept throughout its config schema, workflow rules, Python code, and config files. This rename keeps evalml aligned with upstream anemoi terminology.

Rename interpolator to temporal downscaler

6afde17

dnerini requested review from MicheleCattaneo and frazane June 6, 2026 19:08

dnerini added 3 commits June 6, 2026 21:25

Merge branch 'main' into rename-interpolator-downscaler

b2c3259

Update config for downscaler

5d70477

Merge branch 'main' into rename-interpolator-downscaler

c3d416b

frazane approved these changes Jun 7, 2026

View reviewed changes

dnerini added 2 commits June 7, 2026 11:46

Fix config

b5e7ef4

Merge branch 'main' into rename-interpolator-downscaler

9c87aca

MicheleCattaneo reviewed Jun 8, 2026

View reviewed changes

Comment thread workflow/tools/config.schema.json Outdated

Comment thread workflow/tools/config.schema.json

dnerini added 4 commits June 8, 2026 21:24

Merge branch 'main' into rename-interpolator-downscaler

eec0b5b

Use same checkpoint used by runml

646c8d2

Update description

c0cb007

Revert checkpoint update

039a9da

jonasbhend reviewed Jun 9, 2026

View reviewed changes

dnerini added 2 commits June 9, 2026 10:18

Handle 2- and 3- digit leadtimes

96d5dd0

Update config schema

32b2e39

dnerini merged commit 0c2ee3b into main Jun 9, 2026
4 checks passed

dnerini deleted the rename-interpolator-downscaler branch June 9, 2026 08:22

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Rename interpolator to temporal downscaler#171

Rename interpolator to temporal downscaler#171
dnerini merged 12 commits into
mainfrom
rename-interpolator-downscaler

dnerini commented Jun 6, 2026

Uh oh!

frazane left a comment

Uh oh!

MicheleCattaneo left a comment •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

jonasbhend left a comment

Uh oh!

jonasbhend Jun 9, 2026

Uh oh!

MicheleCattaneo commented Jun 9, 2026

Uh oh!

dnerini commented Jun 9, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

dnerini commented Jun 6, 2026

Uh oh!

frazane left a comment

Choose a reason for hiding this comment

Uh oh!

MicheleCattaneo left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

jonasbhend left a comment

Choose a reason for hiding this comment

Uh oh!

jonasbhend Jun 9, 2026

Choose a reason for hiding this comment

Uh oh!

MicheleCattaneo commented Jun 9, 2026

Uh oh!

dnerini commented Jun 9, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

MicheleCattaneo left a comment •

edited

Loading