mem: use np.full() instead of np.ones() * scalar in YoloX preprocess

[KRRT7]

Replace np.ones(shape, dtype=np.uint8) * 114 with np.full(shape, 114, dtype=np.uint8) in yolox.preprocess().

np.ones() * scalar allocates two arrays: the ones array and a temporary for the multiplication result. np.full() does it in a single allocation — same result, half the memory.

Benchmark

Azure Standard_D8s_v5 — 8 vCPU Intel Xeon Platinum 8473C, 32 GiB RAM, Python 3.12.12

Timing — preprocess (1700x2200x3 letter-size image at 200 DPI)

Ref	Min	Median	Mean	OPS	Rounds
44da8d85541d (base)	2.1ms	2.8ms	2.7ms	365.03 ops/s	124
6a085c621f1c (head)	2.0ms	2.5ms	2.6ms	377.49 ops/s	98
Speedup	1.04x	1.11x	1.03x	1.03x

Memory (peak per test)

Ref	Peak Memory	Allocations	Delta
44da8d85541d (base)	13.4 MiB	54
6a085c621f1c (head)	13.4 MiB	57	+0%

Peak memory is identical because memray tracks the high watermark, not total allocations. The savings are in transient allocation volume — the temporary np.ones() * 114 array (2.25 MiB for a 1024x768x3 uint8 buffer) is eliminated. The original memray stats --json benchmark measured per-call allocation totals: 4.53 MiB/call -> 2.28 MiB/call (-50%).

Original memray allocation benchmark (Apple M3 Max, NumPy 2.4)

np.full() vs np.ones() * scalar
Shape: (1024, 768, 3) (2.25 MiB)  |  10 iterations  |  Python 3.12.12  |  NumPy 2.4.2

  np.ones() * 114: 4.53 MiB / call
  np.full(114):    2.28 MiB / call

  Savings: 2.25 MiB (50%)

The savings equal exactly one array copy (2.25 MiB for a 1024x768x3 uint8 buffer). In the full hi_res pipeline this is a small but free win — no behavior change, no new dependencies.

---

Generated by codeflash optimization agent

Reproduce the benchmark locally

# Full comparison (timing + memory):
uv run codeflash compare 44da8d85541d 6a085c621f1c --memory \
  --inject benchmarks/test_benchmark_preprocess.py \
  --inject benchmarks/__init__.py \
  --inject pyproject.toml

# Or the original memray allocation benchmark:
pip install memray numpy
python benchmarks/bench_preprocess_padding.py --runs 10

Benchmark test source

"""Benchmark for YoloX preprocess() np.full optimization.

Directly benchmarks the preprocess() function which was changed from
np.ones(shape) * 114 to np.full(shape, 114) -- eliminating one temporary
array allocation per call.
"""

import numpy as np

from unstructured_inference.models.yolox import preprocess


def make_letter_200dpi_array() -> np.ndarray:
    """Create a letter-size image array at 200 DPI (1700x2200x3)."""
    return np.random.randint(0, 255, (2200, 1700, 3), dtype=np.uint8)


# Pre-allocate the input image once
_INPUT_IMAGE = make_letter_200dpi_array()
_INPUT_SHAPE = (1024, 768)


def run_preprocess():
    return preprocess(_INPUT_IMAGE, _INPUT_SHAPE)


def test_benchmark_preprocess(benchmark):
    benchmark(run_preprocess)

Test plan

• codeflash compare confirms 1.03-1.11x speedup, timing neutral
• Original memray stats confirm -50% per-call allocation (4.53 -> 2.28 MiB)
• No behavior change — np.full() produces identical output to np.ones() * scalar

[KRRT7]

@cragwolfe looks like it needs explicitly approval for auto merge