---

name: fedformer description: "Routes FEDformer training, comparison, and sweep workflows for the FEDformer subrepo, including Fourier and Wavelets variants, model-selection knobs, and the bundled CUDA smoke check." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

FEDformer

Use this route when the task is about the FEDformer subrepo entry point FEDformer/run.py, the FEDformer model stack in FEDformer/exp, FEDformer/models, FEDformer/layers, FEDformer/utils, or the FEDformer sweep scripts.

Natural triggers:

• FEDformer
• Fourier
• Wavelets
• mode_select
• modes
• L
• base
• cross_activation
• compare FEDformer against Autoformer, Informer, or Transformer within this subrepo

What this sub-skill covers

• The native CLI in FEDformer/run.py.
• The FEDformer architecture variants implemented in FEDformer/models/FEDformer.py.
• The exp, layer, and utility trees that support training, testing, and prediction.
• Dataset loading and shape alignment for the FEDformer entry point.
• The sweep patterns encoded in FEDformer/scripts/LongForecasting.sh and FEDformer/scripts/LookBackWindow.sh.
• CUDA-only smoke validation for the Fourier or Wavelets path.

What it does not cover

• The repo-wide Linear, DLinear, and NLinear benchmark routes.
• Statistical baseline workflows.
• Pyraformer workflows.
• Maintainer-only or release-only tasks.

If the user really means the root long-forecasting benchmark, the statistical baselines, or Pyraformer, route them to the sibling route for that workflow instead of staying here.

Start here

Read these bundled files first:

• scripts/run_fedformer.py for a safe wrapper that resolves a checkout and builds a FEDformer command.
• scripts/smoke_fedformer.py for a tiny CUDA forward pass on synthetic data.
• references/cli-reference.md for the FEDformer CLI flags and the arguments that matter most.
• references/model-overview.md for Fourier vs. Wavelets, embedding choices, and comparison guidance.
• references/data-layout.md when the CSV format, split rule, or channel counts are unclear.
• references/workflows.md for one-off runs, sweep recipes, and comparison loops.
• references/troubleshooting.md when the run fails, the parser defaults look suspicious, or the dataset layout does not match.

Environment expectations

This route is verified for a CUDA build of PyTorch 1.9.0+cu111 on GPU hardware. Wavelets also needs sympy, einops, and scipy in addition to the standard FEDformer stack.

Common decisions

• Use --model FEDformer for the FEDformer family.
• Use --version Fourier when you want mode selection over frequency bins.
• Use --version Wavelets when you want the multiresolution branch with L, base, and cross_activation.
• Use --model Autoformer, Informer, or Transformer only when the comparison is still anchored in this subrepo.
• Keep the dataset, lengths, and feature mode fixed when comparing model families.
• Leave the --moving_avg parser default alone unless you are programmatically building a real Python list.

Key runtime outputs

A successful run usually writes checkpoints and evaluation artifacts under the current working directory, including:

• checkpoints/<setting>/checkpoint.pth
• results/<setting>/metrics.npy
• results/<setting>/pred.npy
• results/<setting>/true.npy
• test_results/<setting>/

Troubleshooting shortcut

Before debugging a failed FEDformer run, check these first:

1. CUDA is available in the active environment.
2. --model is set explicitly; the parser default is not a valid FEDformer family member.
3. root_path, data_path, features, and channel counts match the dataset.
4. You are not relying on the pred-only or test-only path without checking its caveats.
5. The version-specific knobs match the selected --version.

See references/troubleshooting.md for the full failure map.