@camstack/addon-pipeline 1.1.66 → 1.1.68

package/python/test_inference_pool_preprocess.py

@@ -11,7 +11,13 @@ np = pytest.importorskip("numpy")
 PIL = pytest.importorskip("PIL")
 from PIL import Image  # noqa: E402
 
-from inference_pool import _to_model_tensor, _preprocess_ctc_gray  # noqa: E402
+import inference_pool  # noqa: E402
+from inference_pool import (  # noqa: E402
+    _preprocess,
+    _preprocess_ctc_gray,
+    _preprocess_ov_uint8,
+    _to_model_tensor,
+)
 
 
 def _hwc(h: int, w: int) -> "np.ndarray":
@@ -63,3 +69,89 @@ def test_ctc_gray_caps_resized_width_at_input_w() -> None:
     img = Image.new("RGB", (800, 64), (50, 50, 50))
     out = _preprocess_ctc_gray(img, 384, 64)
     assert out.shape == (1, 1, 64, 384)
+
+
+# ---------------------------------------------------------------------------
+# Fix 2 — OpenVINO PrePostProcessor uint8 producer (`_preprocess_ov_uint8`)
+# ---------------------------------------------------------------------------
+
+
+def test_ov_uint8_letterbox_is_nhwc_uint8_at_model_size() -> None:
+    # A non-square frame is aspect-preserved + padded to a square uint8 canvas.
+    plan = {"inputW": 640, "inputH": 640, "size": 640, "letterbox": True, "normalization": "none"}
+    img = Image.new("RGB", (1280, 720), (10, 20, 30))
+    tensor, scale, pad = _preprocess_ov_uint8(img, plan)
+    assert tensor.shape == (1, 640, 640, 3)  # NHWC — OV transposes on-device
+    assert tensor.dtype == np.uint8  # NO float/255 in Python — folded into OV
+    assert 0 < scale < 1  # 640/1280
+    assert pad[1] > 0  # letterbox pads the shorter (vertical) axis
+
+
+def test_ov_uint8_letterbox_preserves_aspect_ratio() -> None:
+    # The real content occupies scale*orig; the rest is 114 grey padding — the
+    # aspect-ratio letterbox YOLO needs, preserved even though float/layout
+    # moved into OV.
+    plan = {"inputW": 640, "inputH": 640, "size": 640, "letterbox": True, "normalization": "none"}
+    img = Image.new("RGB", (1280, 640), (200, 100, 50))
+    tensor, scale, pad = _preprocess_ov_uint8(img, plan)
+    assert scale == pytest.approx(0.5, abs=1e-6)
+    # Padded rows (top/bottom) are the 114 letterbox fill.
+    grid = tensor[0]
+    assert np.all(grid[0] == 114)
+    assert np.all(grid[-1] == 114)
+
+
+def test_ov_uint8_plain_resize_matches_model_dims() -> None:
+    plan = {"inputW": 224, "inputH": 224, "size": 224, "letterbox": False, "normalization": "imagenet"}
+    img = Image.new("RGB", (500, 300), (128, 128, 128))
+    tensor, scale, pad = _preprocess_ov_uint8(img, plan)
+    assert tensor.shape == (1, 224, 224, 3)
+    assert tensor.dtype == np.uint8
+    assert (scale, pad) == (1.0, (0, 0))
+
+
+# ---------------------------------------------------------------------------
+# Fix 1 — bench preprocess cache is HIT on the (tagged) batch path
+# ---------------------------------------------------------------------------
+
+
+def test_bench_tag_caches_preprocess_and_counts_hit() -> None:
+    # A tagged frame (as the MSG_INFER_BATCH path now sets via `_bench_frame_id`)
+    # is preprocessed ONCE (miss) then served from cache (hit) — the fix that
+    # turns the sustained batch benchmark into a pure-inference measurement.
+    inference_pool._runtime = "onnxruntime"
+    inference_pool._bench_preprocess_cache.clear()
+    inference_pool._bench_cache_hits = 0
+    inference_pool._bench_cache_misses = 0
+
+    cfg = {"inputSize": 640, "_input_name": "images", "preprocessMode": "letterbox"}
+    img = Image.new("RGB", (640, 640), (12, 34, 56))
+    img._bench_frame_id = 4242
+
+    first = _preprocess(img, cfg)
+    assert inference_pool._bench_cache_misses == 1
+    assert inference_pool._bench_cache_hits == 0
+
+    # Five more inferences on the SAME pinned frame → all cache hits, and the
+    # returned tuple is the SAME object (no re-preprocess).
+    for _ in range(5):
+        again = _preprocess(img, cfg)
+        assert again is first
+    assert inference_pool._bench_cache_hits == 5
+    assert inference_pool._bench_cache_misses == 1
+
+
+def test_untagged_frame_never_caches() -> None:
+    # Live camera frames carry no tag → every call re-preprocesses (correct).
+    inference_pool._runtime = "onnxruntime"
+    inference_pool._bench_preprocess_cache.clear()
+    inference_pool._bench_cache_hits = 0
+    inference_pool._bench_cache_misses = 0
+
+    cfg = {"inputSize": 640, "_input_name": "images", "preprocessMode": "letterbox"}
+    img = Image.new("RGB", (640, 640), (9, 9, 9))  # no _bench_frame_id
+    _preprocess(img, cfg)
+    _preprocess(img, cfg)
+    assert inference_pool._bench_cache_hits == 0
+    assert inference_pool._bench_cache_misses == 0
+    assert len(inference_pool._bench_preprocess_cache) == 0