Declare an ONNX inference backend for basic-pitch

evaluation_function/backend_test.py

@@ -0,0 +1,64 @@
+"""
+backend_test.py
+===============
+Tests that the audio-to-MIDI transcription backend is actually available.
+
+Basic Pitch does not ship inference code of its own. It picks a backend at
+import time from whichever of CoreML, TensorFlow, TFLite or ONNX Runtime it
+finds installed, and if it finds none it fails while defining
+ICASSP_2022_MODEL_PATH:
+
+    NameError: name '_default_model_type' is not defined
+
+Basic Pitch declares those backends behind platform markers, so which one
+arrives depends on the machine doing the install:
+
+    - macOS       -> coremltools
+    - Linux/amd64 -> tensorflow
+    - Linux/arm64 -> the `tensorflow` wheel is an empty shim that requires
+                     `tensorflow-cpu-aws`, which does not get resolved into
+                     our lock, so no backend is installed at all
+
+The third case is not hypothetical: it is every container built on an Apple
+Silicon machine, where the worker dies at import and the container never
+starts. CI and the AWS build are amd64, so they never see it.
+
+The fix is to stop relying on those markers and declare a backend
+explicitly. ONNX Runtime is the one with wheels for every platform we
+build for, so these tests assert it is present.
+
+Run locally with:  python -m pytest evaluation_function/backend_test.py -v
+"""
+
+import importlib
+
+
+def test_onnx_runtime_is_installed():
+    """
+    A backend must be declared explicitly rather than inherited from a
+    platform marker, so that arm64 containers get one too.
+    """
+    assert importlib.util.find_spec("onnxruntime") is not None, (
+        "onnxruntime is not installed; basic-pitch needs an explicitly "
+        "declared inference backend so that every build platform gets one"
+    )
+
+
+def test_basic_pitch_resolves_a_model_path():
+    """
+    The symptom that takes the container down. Importing this name fails
+    with NameError when no backend is installed.
+    """
+    from basic_pitch import ICASSP_2022_MODEL_PATH
+
+    assert ICASSP_2022_MODEL_PATH is not None
+
+
+def test_model_loads_through_our_own_loader():
+    """
+    Guards the entry point the evaluation function actually calls, so a
+    missing or broken backend fails here rather than at container start.
+    """
+    from .audio_processing import load_basic_pitch_model
+
+    assert load_basic_pitch_model() is not None

pyproject.toml

@@ -21,7 +21,10 @@ matplotlib = "^3.11.0"
 pretty-midi = "^0.2.11"
 pandas = "^3.0.3"
 librosa = "^0.11.0"
-basic-pitch = "^0.4.0"
+# The "onnx" extra pins an inference backend explicitly. Without it,
+# basic-pitch relies on platform markers that leave linux/arm64 with no
+# backend at all, so the container dies at import. See backend_test.py.
+basic-pitch = { version = "^0.4.0", extras = ["onnx"] }
 setuptools = "<81"
 tqdm = "^4.68.4"
 huggingface-hub = "^1.23.0"