Align Iris amplitude encoding benchmark between PennyLane and QDP

[rich7420]

Related Issues

Changes

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Why

Align the Iris amplitude encoding benchmark between the pure PennyLane baseline and the QDP pipeline so we can directly compare full training behavior with only the encoding step changed.

How

Updated pennylane_baseline/iris_amplitude.py and qdp_pipeline/iris_amplitude.py to share the same data loading, CLI options, and training loop, differing only in encoding (get_angles vs QDP QuantumDataLoader + StatePrep).
Added --data-file, optimizer/early-stop/trials flags, and consistent logging to both scripts to support reproducible comparisons.

please try it

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[rich7420]

results in my local environment
our throuphput is better

uv run python benchmark/encoding_benchmarks/pennylane_baseline/iris_amplitude.py --data-file benchmark/encoding_benchmarks/pennylane_baseline/data/iris_classes1and2_scaled.txt --optimizer nesterov --lr 0.01 --layers 6 --trials 3 --iters 80 --early-stop 0 2>&1
Iris amplitude baseline (PennyLane) — 2-class variational classifier
  Data: official file (2 features): benchmark/encoding_benchmarks/pennylane_baseline/data/iris_classes1and2_scaled.txt → L2 norm → get_angles  (n=100; 2-class Iris = 100 samples)
  Iters: 80, batch_size: 5, layers: 6, lr: 0.01, optimizer: nesterov

  Trial 1:
    Compile:   0.0121 s
    Train:     1.4786 s
    Train acc: 1.0000  (n=75)
    Test acc:  1.0000  (n=25)
    Throughput: 270.5 samples/s

  Trial 2:
    Compile:   0.0101 s
    Train:     1.5911 s
    Train acc: 1.0000  (n=75)
    Test acc:  1.0000  (n=25)
    Throughput: 251.4 samples/s

  Trial 3:
    Compile:   0.0102 s
    Train:     1.8057 s
    Train acc: 1.0000  (n=75)
    Test acc:  1.0000  (n=25)
    Throughput: 221.5 samples/s

  Best test accuracy:  1.0000  (median: 1.0000, min: 1.0000, max: 1.0000)
  → Target ≥0.9 achieved.

uv run python benchmark/encoding_benchmarks/qdp_pipeline/iris_amplitude.py --data-file benchmark/encoding_benchmarks/pennylane_baseline/data/iris_classes1and2_scaled.txt --optimizer nesterov --lr 0.01 --layers 6 --trials 3 --iters 80 --early-stop 0 2>&1
Iris amplitude (QDP encoding) — 2-class variational classifier
  Data: official file (2 features): benchmark/encoding_benchmarks/pennylane_baseline/data/iris_classes1and2_scaled.txt → QDP amplitude  (n=100; 2-class Iris = 100 samples)
  Iters: 80, batch_size: 5, layers: 6, lr: 0.01, optimizer: nesterov

  Trial 1:
    QML device: cpu
    Compile:   0.0117 s
    Train:     1.2235 s
    Train acc: 0.9867  (n=75)
    Test acc:  1.0000  (n=25)
    Throughput: 326.9 samples/s

  Trial 2:
    QML device: cpu
    Compile:   0.0081 s
    Train:     1.2546 s
    Train acc: 1.0000  (n=75)
    Test acc:  1.0000  (n=25)
    Throughput: 318.8 samples/s

  Trial 3:
    QML device: cpu
    Compile:   0.0081 s
    Train:     1.3195 s
    Train acc: 0.9867  (n=75)
    Test acc:  1.0000  (n=25)
    Throughput: 303.1 samples/s

  Best test accuracy:  1.0000  (median: 1.0000, min: 1.0000, max: 1.0000)
  → Target ≥0.9 achieved.

[ryankert01]

Nice

[ryankert01]

Why did you set qubits = 2?

[ryankert01]

Oh, It might be we want to keep the qubit small for a clear speedup result.

[ryankert01]

lgtm, I think if we are going to add multiple dataset, we can abstract many things in common as a followup (not necessary) and utility to store results in a csv or something.