Sourced from whest-starterkit @ aaa3882.

Stage 1: Iterate Locally (Just flopscope)

← Tutorial

Ladder: 1 · 2 · 3 · 4 · 5

"Just flopscope" means: no whest CLI required. You run python estimator.py and the bundled local_engine.py constructs an MLP, calls your predict() inside a flopscope.BudgetContext, and sweeps Monte-Carlo sample counts to print a FLOPs-vs-MSE table. The whestbench.BaseEstimator and whestbench.MLP types you'll see imported are just the shared dataclasses — they don't pull in the harness.

Iterate here until predict() converges, then climb to Stage 2 to confirm the contract.

🚀 Run it

uv run python estimator.py

You should see a table like:

--- Your estimator ---
MLP: width=256 depth=8 seed=0

 n_samples | sampling_flops | estimator_flops |        MSE
----------------------------------------------------------
        10 |     10,551,808 |               0 |   0.706450
       100 |    105,499,648 |               0 |   0.718928
     1,000 |  1,054,978,048 |               0 |   0.739125
    10,000 | 10,549,762,048 |               0 |   0.735092
   100,000 | 105,497,602,048 |               0 |   0.737172

The stub predict() returns all zeros, so estimator_flops is 0 and the MSE plateaus at the variance of the true outputs — once you put real math in predict(), both columns come alive and the MSE should shrink roughly as 1/sqrt(n_samples) (Monte Carlo converging to your estimator's answer).

Edit predict()

Open estimator.py. The body of predict() returns all zeros — replace it with your idea. The template already imports flopscope as flops and flopscope.numpy as fnp, so any array op you write through fnp (or via Python operators on fnp arrays) is FLOP-counted automatically. Re-run; the MSE column tells you how close you are, and estimator_flops shows what your math cost.

Compare against a baseline

uv run python estimator.py --baseline mean_propagation

This loads examples/02_mean_propagation.py and runs both estimators on the same MLP.

✅ Expected outcome

You're ready for Stage 2 once your estimator's MSE is comfortably below the zeros floor and estimator_flops stays under whatever budget you'd target downstream (Stage 3's grader default is 6.8e10).

✅ When you're ready

Move on to Stage 2: validate the contract.