update eval scripts: add ONNX size tracking and output sanitization

[DingmaomaoBJTU]

Summary

• Add _compute_onnx_size() to measure combined ONNX + .data file sizes and include onnx_size_bytes in eval results
• Add _sanitize_output() to strip CLI chrome (Rich tables, device/IO banners) from eval_result.json, keeping only error-relevant content
• Minor formatting fixes in reporter.py

[DingmaomaoBJTU]

Nice additions - ONNX size tracking and output sanitization are useful for keeping eval_result.json lean. A few observations below.

[DingmaomaoBJTU]

Useful additions - ONNX size tracking and output sanitization will make eval results much cleaner. A few suggestions below.

[xieofxie]

a little strange.. any way we add a parameter in eval command to just drop them?

[timenick]

_BOX_CHARS only covers Unicode's LIGHT box-drawing style (─│┌┐└┘├┤┬┴┼), but winml perf uses Rich's default Table(), which renders with box.HEAVY_HEAD. I rendered a default Rich table locally and 3 of the 5 lines start with chars not in this set:

Row	First char	In _BOX_CHARS?
top border ┏━━━━━┳━━━━━┓	┏ (U+250F)	❌
header row ┃ Avg ┃ P90 ┃	┃ (U+2503)	❌
head separator ┡━━━━━╇━━━━━┩	┡ (U+2521)	❌
data row │ 12.5 │ 15.2 │	│	✅
bottom border └─────┴─────┘	└	✅

Net result: eval_result.json keeps the top border + header text + head separator while stripping data rows and the bottom border — arguably uglier than no sanitization at all (orphaned half-table chrome).

Cheap fix — use a Unicode block range check instead of a hand-curated set:

if 0x2500 <= ord(stripped[0]) <= 0x257F:  # Unicode "Box Drawing" block
    continue

That covers all four Rich styles (light, heavy, double, rounded) in one rule and won''t silently drift the next time someone changes the table style.

🤖 Generated with GitHub Copilot CLI

[timenick]

Appending stripped (post-.strip()) discards leading indentation, which destroys the structure of multi-line errors — the very content the docstring promises to preserve ("All classifier patterns are error-related and survive this filter"). For example, a traceback:

  File "foo.py", line 5, in bar
    raise RuntimeError("x")

becomes a visually-flat:

File "foo.py", line 5, in bar
raise RuntimeError("x")

which is noticeably harder to read.

Suggested fix — keep stripped only for the box/noise classifier checks, then append the original line (lightly trimmed):

if not stripped:
    continue
if 0x2500 <= ord(stripped[0]) <= 0x257F:
    continue
low = stripped.lower()
if any(low.startswith(pat) for pat in _NOISE_PATTERNS):
    continue
kept.append(line.rstrip())

🤖 Generated with GitHub Copilot CLI

[timenick]

A couple of patterns in _NOISE_PATTERNS can swallow legitimate diagnostic content with the current low.startswith(pat) matching:

• "inputs:" / "outputs:" — these silently strip lines like Inputs: expected (1, 224, 224, 3), got (1, 3, 224, 224) (a real shape-mismatch error), which is exactly the kind of "error-relevant" content a sanitizer is supposed to keep. Same concern for "device:" if a downstream error ever reads something like Device: <name> is not available, falling back to CPU.
• "samples/sec" — dead pattern under startswith semantics. Throughput: 80 samples/sec is already covered by "throughput:" above; no winml perf line literally starts with samples/sec.

Tightening options (cheapest first):

1. Just drop inputs: / outputs: / samples/sec. The remaining patterns are unambiguous CLI chrome.
2. Switch to exact-prefix-with-boundary: only strip when the line is pat followed by space or end, e.g. low == pat or low.startswith(pat + " "), so error lines that happen to start with Inputs: but carry non-trivial content survive.

🤖 Generated with GitHub Copilot CLI