Upgrade numpy to 2.x in OTX (#4970)

Co-authored-by: Copilot <[email protected]>

Author: leoll2

library/pyproject.toml

@@ -51,7 +51,8 @@ dependencies = [
     "onnxscript==0.5.3",
     "nncf==2.17.0",
     "anomalib[core]==1.1.3",
-    "numpy<2.0",
+    "imgaug==0.4.0",  # for anomalib
+    "numpy>2.0",
     "tensorboardx==2.6.4"
 ]
 
@@ -114,6 +115,14 @@ otx = "otx.cli:main"
 Documentation = "https://open-edge-platform.github.io/training_extensions/"
 Repository = "https://github.com/open-edge-platform/training_extensions/"
 
+[tool.uv]
+conflicts = [
+    [
+        { extra = "xpu" },
+        { extra = "cuda" },
+    ],
+]
+
 [tool.uv.sources]
 torch = [
     { index = "torch-xpu", extra = "xpu"},
@@ -126,6 +135,10 @@ torchvision = [
     { index = "torch-xpu", extra = "xpu"},
     { index = "torch-cuda", extra = "cuda"},
 ]
+# Anomalib depends on imgaug, which is abandoned and depends on an old version of numpy.
+# To avoid dependency resolution issues, we install imgaug from a maintained fork.
+# For details: https://github.com/aleju/imgaug/issues/859
+imgaug = { git = "https://github.com/imaug/imaug/", rev = "a7ca1b04d44700d075a23f14a97f86285ba3b33f" }
 
 [[tool.uv.index]]
 name = "torch-cuda"
@@ -285,12 +298,6 @@ ignore = [
 
     "E731",     # Do not assign a `lambda` expression, use a `def`
     "TD003",    # Missing issue link on the line following this TODO
-
-    # TODO (someone): Update legacy np.random.{method_name} call with np.random.Generator
-    # Legacy (MT19964) and new (PCG64) random generators use different pseudo-random number generators.
-    # (https://numpy.org/doc/stable/reference/random/legacy.html#legacy)
-    # For now, remaining legacy random methods is required until numpy version is explicitly updated in OTX.
-    "NPY002",   # Replace legacy np.random.{method_name} call with np.random.Generator
 ]
 
 # Allow autofix for all enabled rules (when `--fix`) is provided.

library/src/otx/backend/native/exporter/exportable_code/demo/demo_package/visualizers/visualizer.py

@@ -204,11 +204,12 @@ def __init__(
         self.color_map = self._create_color_map()
 
     def _create_color_map(self) -> np.ndarray:
+        rng = np.random.default_rng(42)
         classes = self.color_palette[:, ::-1]  # RGB to BGR
         color_map = np.zeros((256, 1, 3), dtype=np.uint8)
         classes_num = len(classes)
         color_map[:classes_num, 0, :] = classes
-        color_map[classes_num:, 0, :] = np.random.uniform(0, 255, size=(256 - classes_num, 3))
+        color_map[classes_num:, 0, :] = rng.uniform(0, 255, size=(256 - classes_num, 3))
         return color_map
 
     def _apply_color_map(self, input_2d_mask: np.ndarray) -> np.ndarray:

library/src/otx/backend/native/models/detection/detectors/detection_transformer.py

@@ -67,6 +67,7 @@ def __init__(
             )
         )
         self.optimizer_configuration = optimizer_configuration
+        self.rng = np.random.default_rng(42)
 
     def generate_scales(self, input_size: int, base_size_repeat: int = 3) -> list[int]:
         """Generates scales for multi-scale training."""
@@ -84,7 +85,7 @@ def _forward_features(self, images: Tensor, targets: dict[str, Any] | None = Non
     def forward(self, images: Tensor, targets: dict[str, Any] | None = None) -> dict[str, Tensor] | Tensor:
         """Forward pass of the model."""
         if self.multi_scale and self.training:
-            sz = int(np.random.choice(self.multi_scale))
+            sz = int(self.rng.choice(self.multi_scale))
             images = nn.functional.interpolate(images, size=[sz, sz])
 
         output = self._forward_features(images, targets)