Remove PyTorch dependencies

pyproject.toml

@@ -35,16 +35,15 @@ dynamic = [
 ]
 dependencies = [
     "beartype>=0.22.9",
+    "brisque>=0.0.17",
     "flask>=3.0.3",
     "numpy>=1.26.4",
     "pillow>=11.0.0",
-    "piq>=0.8.0",
     "pydantic-settings>=2.6.1",
     "requests>=2.32.3",
     "responses>=0.25.3",
-    "torch>=2.5.1",
-    "torchmetrics>=1.5.1",
-    "torchvision>=0.20.1",
+    "scikit-image>=0.21.0",
+    "scipy>=1.12.0",
     "tzdata; sys_platform=='win32'",
     "vws-auth-tools>=2024.7.12",
     "werkzeug>=3.1.2",
@@ -139,11 +138,6 @@ fallback_version = "0.0.0"
 # Code to match this is in ``conf.py``.
 version_scheme = "post-release"
 
-[tool.uv]
-sources.torch = { index = "pytorch-cpu" }
-sources.torchvision = { index = "pytorch-cpu" }
-index = [ { name = "pytorch-cpu", url = "https://download.pytorch.org/whl/cpu", explicit = true } ]
-
 [tool.ruff]
 line-length = 79
 lint.select = [
@@ -315,12 +309,11 @@ pep621_dev_dependency_groups = [
     "release",
 ]
 per_rule_ignores.DEP002 = [
+    # scipy is a transitive dependency of brisque.
+    "scipy",
     # tzdata is needed on Windows for zoneinfo to work.
     # See https://docs.python.org/3/library/zoneinfo.html#data-sources.
     "tzdata",
-    # torchvision is used transitively via piq, but must be a direct dependency
-    # so that tool.uv.sources can route it to the CPU-only PyTorch index.
-    "torchvision",
 ]
 
 [tool.pyproject-fmt]

spelling_private_dict.txt

@@ -15,6 +15,7 @@ beartype
 binascii
 bool
 boolean
+brisque
 bytesio
 changelog
 chunked

src/mock_vws/_brisque.py

@@ -0,0 +1,23 @@
+"""Wrapper around the brisque package for image quality scoring."""
+
+from __future__ import annotations
+
+import io
+
+import numpy as np
+from brisque import (  # type: ignore[attr-defined]  # pyright: ignore[reportMissingTypeStubs]
+    BRISQUE,
+)
+from PIL import Image
+
+_brisque_scorer = BRISQUE(url=False)  # type: ignore[no-untyped-call]
+
+
+def brisque_score(image_content: bytes) -> float:
+    """Return a BRISQUE quality score for the given image bytes."""
+    image_file = io.BytesIO(initial_bytes=image_content)
+    image = Image.open(fp=image_file).convert(mode="RGB")
+    image_np: np.ndarray = np.array(object=image)
+    return float(
+        _brisque_scorer.score(img=image_np)  # type: ignore[no-untyped-call]  # pyright: ignore[reportUnknownMemberType]
+    )

src/mock_vws/_flask_server/Dockerfile

@@ -1,4 +1,6 @@
 FROM ghcr.io/astral-sh/uv:0.10.4-python3.13-trixie-slim AS base
+# Install build tools required by packages that compile C++ extensions (e.g. libsvm-official).
+RUN apt-get update && apt-get install -y --no-install-recommends g++=3.0.3 && rm -rf /var/lib/apt/lists/*
 # We set this pretend version as we do not have Git in our path, and we do
 # not care enough about having the version correct inside the Docker container
 # to install it.

src/mock_vws/image_matchers.py

@@ -1,14 +1,13 @@
 """Matchers for query and duplicate requests."""
 
 import io
-from typing import Protocol, runtime_checkable
+from typing import Protocol, cast, runtime_checkable
 
 import numpy as np
-import torch
 from beartype import beartype
 from PIL import Image
-from torchmetrics.image import (
-    StructuralSimilarityIndexMeasure,
+from skimage.metrics import (  # pylint: disable=no-name-in-module
+    structural_similarity,  # pyright: ignore[reportUnknownVariableType]
 )
 
 
@@ -78,42 +77,22 @@ def __call__(
         first_image_resized = first_image.resize(size=target_size)
         second_image_resized = second_image.resize(size=target_size)
 
-        first_image_np = np.array(object=first_image_resized, dtype=np.float32)
-        first_image_tensor = torch.tensor(data=first_image_np).float() / 255
-        first_image_tensor = first_image_tensor.view(
-            first_image_resized.size[1],
-            first_image_resized.size[0],
-            len(first_image_resized.getbands()),
+        first_image_np = (
+            np.array(object=first_image_resized, dtype=np.float32) / 255
         )
-
-        second_image_np = np.array(
-            object=second_image_resized,
-            dtype=np.float32,
-        )
-        second_image_tensor = torch.tensor(data=second_image_np).float() / 255
-        second_image_tensor = second_image_tensor.view(
-            second_image_resized.size[1],
-            second_image_resized.size[0],
-            len(second_image_resized.getbands()),
+        second_image_np = (
+            np.array(object=second_image_resized, dtype=np.float32) / 255
         )
 
-        first_image_tensor_batch_dimension = first_image_tensor.permute(
-            2,
-            0,
-            1,
-        ).unsqueeze(dim=0)
-        second_image_tensor_batch_dimension = second_image_tensor.permute(
-            2,
-            0,
-            1,
-        ).unsqueeze(dim=0)
-
-        ssim = StructuralSimilarityIndexMeasure(data_range=1.0)
-        ssim_value = ssim(
-            first_image_tensor_batch_dimension,
-            second_image_tensor_batch_dimension,
+        ssim_score: float = cast(
+            "float",
+            structural_similarity(  # type: ignore[no-untyped-call]
+                im1=first_image_np,
+                im2=second_image_np,
+                data_range=1.0,
+                channel_axis=2,
+            ),
         )
-        ssim_score = ssim_value.item()
 
         # Normalize SSIM score from -1 to 1 scale to 0 to 10 scale.
         # This maps -1 to 0 and 1 to 10.

src/mock_vws/target_raters.py

@@ -1,16 +1,13 @@
 """Raters for target quality."""
 
 import functools
-import io
 import math
 import secrets
 from typing import Protocol, runtime_checkable
 
-import numpy as np
-import torch
 from beartype import beartype
-from PIL import Image
-from piq.brisque import brisque  # pyright: ignore[reportMissingTypeStubs]
+
+from mock_vws._brisque import brisque_score
 
 
 @functools.cache
@@ -25,21 +22,11 @@ def _get_brisque_target_tracking_rating(*, image_content: bytes) -> int:
     Args:
         image_content: A target's image's content.
     """
-    image_file = io.BytesIO(initial_bytes=image_content)
-    image = Image.open(fp=image_file)
-    image_np = np.array(object=image, dtype=np.float32)
-    image_tensor = torch.tensor(data=image_np).float() / 255
-    image_tensor = image_tensor.view(
-        image.size[1],
-        image.size[0],
-        len(image.getbands()),
-    )
-    image_tensor = image_tensor.permute(2, 0, 1).unsqueeze(dim=0)
     try:
-        brisque_score = brisque(x=image_tensor, data_range=255)
+        score = brisque_score(image_content=image_content)
     except (AssertionError, IndexError):
         return 0
-    return math.ceil(int(brisque_score.item()) / 20)
+    return math.ceil(int(score) / 20)
 
 
 @runtime_checkable