test: add unit test suite and CI workflow

.github/workflows/ci.yml

@@ -0,0 +1,53 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+          cache-dependency-path: pyproject.toml
+      # Installing the full package also acts as a clean-install smoke test
+      # for the dependency resolution fixed in #65.
+      - name: Install package with test extras
+        run: |
+          python -m pip install --upgrade pip
+          pip install -e ".[test]"
+      - name: Run tests
+        run: pytest -v tests/
+
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Install linters
+        run: pip install "black==24.*" "isort==5.*"
+      # Scoped to tests/ for now; widen to the package in a follow-up once the
+      # existing sources are confirmed clean against these versions.
+      # black and isort read their settings from pyproject.toml.
+      - name: black
+        run: black --check tests/
+      - name: isort
+        run: isort --check-only tests/

pyproject.toml

@@ -14,6 +14,9 @@ authors = [{name = "Meta Platforms, Inc."}]
 dependencies = [
   "neuralset==0.0.2",
   "neuraltrain==0.0.2",
+  # neuralset 0.0.2 references exca.steps.base.NoValue, which exca removed in
+  # 0.5.26 (moved to exca.steps.identity). Cap exca until neuralset is updated.
+  "exca>=0.5.20,<0.5.26",
   "torch>=2.5.1,<2.7",
   "numpy==2.2.6",
   "torchvision>=0.20,<0.22",
@@ -63,5 +66,8 @@ line-length = 88
 [tool.isort]
 profile = "black"
 
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+
 [tool.setuptools.packages.find]
 include = ["tribe*"]

tests/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

tests/test_demo_utils.py

@@ -0,0 +1,36 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Unit tests for the ``demo_utils`` input-suffix contract."""
+
+from tribev2.demo_utils import VALID_SUFFIXES
+
+
+def test_expected_input_modalities():
+    assert set(VALID_SUFFIXES) == {"text_path", "audio_path", "video_path"}
+
+
+def test_every_modality_has_at_least_one_suffix():
+    for name, suffixes in VALID_SUFFIXES.items():
+        assert suffixes, f"{name} has no accepted suffixes"
+
+
+def test_suffixes_are_lowercase_and_dot_prefixed():
+    for suffixes in VALID_SUFFIXES.values():
+        for suffix in suffixes:
+            assert suffix.startswith("."), f"{suffix!r} is missing a leading dot"
+            assert suffix == suffix.lower(), f"{suffix!r} is not lowercase"
+
+
+def test_known_suffixes_are_accepted():
+    assert ".txt" in VALID_SUFFIXES["text_path"]
+    assert ".wav" in VALID_SUFFIXES["audio_path"]
+    assert ".mp4" in VALID_SUFFIXES["video_path"]
+
+
+def test_suffixes_do_not_overlap_across_modalities():
+    all_suffixes = [s for suffixes in VALID_SUFFIXES.values() for s in suffixes]
+    assert len(all_suffixes) == len(set(all_suffixes)), "a suffix maps to >1 modality"

tests/test_eventstransforms.py

@@ -0,0 +1,28 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Unit tests for ``ExtractWordsFromAudio`` language validation.
+
+The language guard runs before any transcription subprocess, so these tests
+need no audio file, GPU, or the ``uvx``/``whisperx`` toolchain.
+"""
+
+from pathlib import Path
+
+import pytest
+
+from tribev2.eventstransforms import ExtractWordsFromAudio
+
+
+def test_unsupported_language_raises_value_error():
+    with pytest.raises(ValueError, match="not supported"):
+        ExtractWordsFromAudio._get_transcript_from_audio(
+            Path("nonexistent.wav"), "klingon"
+        )
+
+
+def test_default_language_is_english():
+    assert ExtractWordsFromAudio().language == "english"

tests/test_temporal_smoothing.py

@@ -0,0 +1,67 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Unit tests for ``TemporalSmoothing`` (CPU only, no data or weights needed)."""
+
+import torch
+from torch import nn
+
+from tribev2.model import TemporalSmoothing
+
+
+def test_build_returns_depthwise_conv1d():
+    conv = TemporalSmoothing(kernel_size=9, sigma=2.0).build(dim=4)
+
+    assert isinstance(conv, nn.Conv1d)
+    # Depthwise: groups == channels, so the weight is (dim, 1, kernel_size).
+    assert conv.groups == 4
+    assert conv.weight.shape == (4, 1, 9)
+    assert conv.bias is None
+
+
+def test_gaussian_kernel_is_normalized_per_channel():
+    dim, kernel_size = 4, 9
+    conv = TemporalSmoothing(kernel_size=kernel_size, sigma=2.0).build(dim=dim)
+
+    per_channel_sum = conv.weight.detach().sum(dim=-1).reshape(dim)
+    assert torch.allclose(per_channel_sum, torch.ones(dim), atol=1e-5)
+
+
+def test_gaussian_kernel_is_symmetric():
+    conv = TemporalSmoothing(kernel_size=9, sigma=2.0).build(dim=1)
+
+    kernel = conv.weight.detach()[0, 0]
+    assert torch.allclose(kernel, torch.flip(kernel, dims=[0]), atol=1e-6)
+
+
+def test_output_length_is_preserved():
+    conv = TemporalSmoothing(kernel_size=9, sigma=2.0).build(dim=4)
+
+    x = torch.randn(2, 4, 50)
+    y = conv(x)
+    assert y.shape == (2, 4, 50)
+
+
+def test_constant_signal_is_unchanged_in_interior():
+    # A normalized smoothing kernel must leave a constant signal unchanged,
+    # away from the zero-padded borders.
+    conv = TemporalSmoothing(kernel_size=9, sigma=2.0).build(dim=3)
+
+    x = torch.ones(1, 3, 50)
+    y = conv(x).detach()
+    # kernel_size=9, padding=4 means only indices 0-3 and 46-49 are
+    # contaminated by zero-padding; 10:40 is safely interior.
+    interior = y[:, :, 10:40]
+    assert torch.allclose(interior, torch.ones_like(interior), atol=1e-5)
+
+
+def test_default_sigma_is_none():
+    # With no sigma the conv keeps its randomly initialized, trainable weights
+    # (the Gaussian branch is skipped).
+    conv = TemporalSmoothing(kernel_size=5).build(dim=2)
+
+    assert isinstance(conv, nn.Conv1d)
+    assert conv.weight.shape == (2, 1, 5)