---

name: training-and-evaluation description: "Train, validate, test, smoke-check, and interpret Informer2020 forecasting runs." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Training and evaluation

Use this sub-skill for the forecasting run loop: choose the model and attention, set sequence lengths and feature mode, launch train/validate/test cycles, read checkpoints and results, and compare benchmark presets.

What it covers

• Model family: informer vs informerstack.
• Attention: prob vs full.
• Feature modes: M, S, MS.
• Lengths: seq_len, label_len, pred_len.
• Training controls: itr, train_epochs, batch_size, patience, learning_rate, lradj, AMP, and multi-GPU.
• Outputs: checkpoints, prediction arrays, metrics, and future-prediction artifacts.
• Benchmark presets distilled from the shipped ETTh1 / ETTh2 / ETTm1 / WTH run families.

Route out

• Custom CSV shape, date extension for future prediction, target / cols handling, and dataset validation: ../custom-data-and-prediction/SKILL.md.
• Tiny fixture generation and smoke orchestration: ../../scripts/make_tiny_forecast_csv.py, ../../scripts/check_forecast_csv.py, ../../scripts/run_forecasting_smoke.py.

Fast facts

• The source CLI always trains first, then tests once per itr; do_predict appends future prediction after test.
• Built-in dataset presets for ETTh1 / ETTh2 / ETTm1 / ETTm2 / WTH / ECL / Solar override data_path, target, enc_in, dec_in, and c_out.
• informerstack uses s_layers; informer uses e_layers.
• --distil and --mix are disable flags: adding them turns those features off.
• checkpoints/<setting>/checkpoint.pth holds the best model; results/<setting>/ holds metrics.npy, pred.npy, true.npy, and real_prediction.npy when prediction is enabled.
• For CPU smoke, hide CUDA or use the smoke helper's CPU backend option; --use_gpu False is not a reliable CPU switch.

Read next

• references/workflows.md
• references/cli-reference.md
• references/troubleshooting.md