Spaces:
Running on Zero
Running on Zero
test(02-04): add failing translation contract tests
Browse files- tests/test_translate.py (mt + slow markers): warm-up under 5 s and idempotent, six
contract sentences asserted on keywords with per-call ms printed, pieces end with </s>,
empty/whitespace raise ValueError, 200-char line bounded by max_decoding_length,
model_bytes matches disk, and no transformers/torch in sys.modules after warm-up
- Fails at collection: japanese_avatar.nlp.translate does not exist yet
- tests/test_translate.py +101 -0
tests/test_translate.py
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""JPN-04 contract for the CPU translator: six sentences, no transformers, warm in under 5 s.
|
| 2 |
+
|
| 3 |
+
Marked ``mt`` (and ``slow``) because it loads the CTranslate2 model (~0.3 s, +100 MB), which
|
| 4 |
+
keeps it out of the quick loop; plan 02-05 runs the whole loop and records its wall time. Every
|
| 5 |
+
assertion here is a keyword, never an exact string - the model is pinned, but a beam search is
|
| 6 |
+
not a contract on wording. Per-call milliseconds are printed for the record and NOT bounded:
|
| 7 |
+
research measured 16-129 ms locally at two threads, and the number that counts is the deployed
|
| 8 |
+
one, which plan 02-10 measures on the Space.
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
from __future__ import annotations
|
| 12 |
+
|
| 13 |
+
import sys
|
| 14 |
+
import time
|
| 15 |
+
|
| 16 |
+
import pytest
|
| 17 |
+
|
| 18 |
+
from japanese_avatar.nlp import translate as mt
|
| 19 |
+
|
| 20 |
+
pytestmark = [pytest.mark.mt, pytest.mark.slow]
|
| 21 |
+
|
| 22 |
+
# Japanese -> substrings of which at least one must appear (case-insensitive). Research Β§ Q4
|
| 23 |
+
# measured the full outputs; the keywords are the parts a paraphrase cannot drop.
|
| 24 |
+
CONTRACT = [
|
| 25 |
+
("γγγ«γ‘γ―γ", ("hello",)),
|
| 26 |
+
("ι§
γ―γ©γγ§γγγ", ("station",)),
|
| 27 |
+
("ζ₯ζ¬θͺγεεΌ·γγ¦γγΎγγ", ("japanese",)),
|
| 28 |
+
(
|
| 29 |
+
"δ»ζ₯γ―γγ倩ζ°γ§γγγγε
¬εγζ£ζ©γγ¦γγγθ²·γη©γ«θ‘γγΎγγγ",
|
| 30 |
+
("park", "shopping", "weather"),
|
| 31 |
+
),
|
| 32 |
+
("ζ₯ζ¬θͺγη·΄ηΏγγΎγγγγ", ("japanese", "practice")),
|
| 33 |
+
("ζ¨ζ₯δ½γγγΎγγγγ", ("yesterday",)),
|
| 34 |
+
]
|
| 35 |
+
|
| 36 |
+
WARMUP_FIRST_MAX_S = 5.0 # research: 0.25-0.58 s locally; the Space CPU is 3-6x slower
|
| 37 |
+
WARMUP_REPEAT_MAX_S = 0.01
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
@pytest.fixture(scope="module")
|
| 41 |
+
def warm() -> tuple[float, float]:
|
| 42 |
+
"""Load the translator once per module; returns (first, second) warm-up seconds."""
|
| 43 |
+
first = mt.warmup()
|
| 44 |
+
second = mt.warmup()
|
| 45 |
+
print(f"\ntranslator warm-up {first:.3f} s, repeat {second:.4f} s")
|
| 46 |
+
return first, second
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def test_warmup_is_fast_and_idempotent(warm):
|
| 50 |
+
first, second = warm
|
| 51 |
+
assert first < WARMUP_FIRST_MAX_S, f"first warm-up took {first:.2f} s"
|
| 52 |
+
assert second < WARMUP_REPEAT_MAX_S, f"repeat warm-up took {second:.4f} s; not cached"
|
| 53 |
+
assert mt.get_translator() is mt.get_translator()
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
@pytest.mark.parametrize(("text", "keywords"), CONTRACT, ids=[t for t, _ in CONTRACT])
|
| 57 |
+
def test_contract_sentences(warm, text, keywords):
|
| 58 |
+
started = time.perf_counter()
|
| 59 |
+
out = mt.translate(text)
|
| 60 |
+
ms = (time.perf_counter() - started) * 1000.0
|
| 61 |
+
print(f"\n{text} -> {out!r} ({ms:.0f} ms)")
|
| 62 |
+
assert isinstance(out, str) and out.strip(), f"empty translation for {text!r}"
|
| 63 |
+
assert any(k in out.lower() for k in keywords), f"{out!r} contains none of {keywords}"
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
def test_pieces_end_with_eos(warm):
|
| 67 |
+
pieces = mt.encode_pieces("γγγ«γ‘γ―")
|
| 68 |
+
assert pieces[-1] == "</s>"
|
| 69 |
+
assert len(pieces) > 1
|
| 70 |
+
assert all(isinstance(p, str) for p in pieces)
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
@pytest.mark.parametrize("text", ["", " ", "\n\t"])
|
| 74 |
+
def test_empty_and_whitespace_raise(warm, text):
|
| 75 |
+
with pytest.raises(ValueError):
|
| 76 |
+
mt.translate(text)
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def test_long_line_is_bounded(warm):
|
| 80 |
+
"""A MAX_TEXT_CHARS-long line decodes within max_decoding_length and stays non-empty."""
|
| 81 |
+
text = ("ζ₯ζ¬θͺγεεΌ·γγ¦γγΎγγ" * 20)[: mt.MAX_TEXT_CHARS]
|
| 82 |
+
assert len(text) == mt.MAX_TEXT_CHARS
|
| 83 |
+
started = time.perf_counter()
|
| 84 |
+
out = mt.translate(text)
|
| 85 |
+
ms = (time.perf_counter() - started) * 1000.0
|
| 86 |
+
print(f"\n{len(text)}-char line -> {len(out)} chars ({ms:.0f} ms)")
|
| 87 |
+
assert out.strip()
|
| 88 |
+
_, _, target = mt.get_translator()
|
| 89 |
+
assert len(target.encode(out, out_type=str)) <= mt.DEFAULT_MAX_DECODING_LENGTH
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def test_model_bytes_is_the_committed_model(warm):
|
| 93 |
+
assert mt.model_bytes() == (mt.MODEL_DIR / "model.bin").stat().st_size
|
| 94 |
+
assert mt.model_bytes() > 70_000_000
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def test_no_transformers_or_torch_imported(warm):
|
| 98 |
+
"""The Space's translation path is ctranslate2 + sentencepiece only (research Β§ Q4)."""
|
| 99 |
+
assert "transformers" not in sys.modules
|
| 100 |
+
assert "torch" not in sys.modules
|
| 101 |
+
assert "ctranslate2" in sys.modules and "sentencepiece" in sys.modules
|