ancs21's picture
DropoutTS reproduction bundle
90e244e verified
|
Raw
History Blame Contribute Delete
1.28 kB

DropoutTS reproduction bundle

Reproduction of DropoutTS: Sample-Adaptive Dropout for Robust Time Series Forecasting (arXiv:2601.21726, OpenReview 7sksHLUvhH) for the Hugging Face "Reproducing ICML 2026" challenge.

Logbook: https://huggingface.co/spaces/ancs21/repro-dropoutts Paper code: https://github.com/CityMind-Lab/DropoutTS

What's here

  • smoke_claim4.py — local check of Claim 4a/4b (param count = 2*num_features+2; eval-mode no-op).
  • modal_repro.py — Modal GPU pipeline: Informer +/- DropoutTS on synthetic sweep (Claim 1), ETTh2 (Claim 2), and the Selective Learning combo (Claim 5). Also times training for Claim 4c.
  • analyze_claim1.py — computes MSE/MAE improvements + the Claim 4c timing.
  • claim1_results.json, claim2_ETTh2_results.json, claim5_results.json — raw run outputs.
  • claim1_table.csv, claim1_plot.html — per-cell synthetic results + figure.

Outcome

Claims 3, 4a, 4b verified. Claim 2 (ETTh2) reproduced (up to +59% MSE). Claim 1 not reproduced under default config (mean -7.5%). Claim 4c contradicted (1.3x slower, not faster). Claim 5 contradicted (combo underperformed Selective Learning alone). See the logbook for details.

Ran on Modal A10G GPUs, single seed, default hyperparameters. ~51 training runs.