Tune a Transformer¶
Runs a small Ray Tune search for the transformer_embedding method
with RoPE positional encoding.
Command¶
Flags:
-n 50— 50 Ray Tune trials. Bump to200for paper-grade numbers.-wl 90— 90-day window (the canonical choice; the sweep varies it at evaluation time).-pe rope— Rotary Positional Embedding. The paper reports all three PE variants.
Expected runtime¶
On one A100, ~2 hours for 50 trials. On an RTX 4090 expect ~2.5 hours.
Scale linearly with -n.
What it writes¶
swissrivernetwork/benchmark/outputs/ray_results/<run_id>/— one subdirectory per trial with the best checkpoint and Ray metadata.- Weights & Biases run, if
WANDB_API_KEYis set.
After tuning¶
Evaluate with:
evaluate writes the wl=90 tables; sweep writes
dump/predictions/<path_extra_keys>-evalwl{W}/ for every eval window
length.
Matching the paper
To reproduce the paper's Transformer numbers exactly, run all three
PE variants (rope, sinusoidal, learnable) with -n 200, then
run the sweep.