fix(model): resolve NFS flock error when loading pretrained timm weights

.gitattributes

@@ -0,0 +1 @@
+anomaly_match/pretrained_cache/models--timm--*/blobs/* filter=lfs diff=lfs merge=lfs -text

.github/workflows/run_tests.yml

@@ -24,6 +24,8 @@ jobs:
       id-token: write
     steps:
       - uses: actions/checkout@v4
+        with:
+          lfs: true
       - name: Set up Python 3.11
         uses: actions/setup-python@v5
         with:

anomaly_match/models/FixMatch.py

@@ -54,8 +54,15 @@ def __init__(
         self.ema_m = ema_m
 
         # Create two versions of the model: one for training and one for evaluation with EMA
+        # eval_model skips pretrained weight download because its weights are immediately
+        # overwritten by copying from train_model below.
         self.train_model = net_builder(num_classes=num_classes, in_channels=in_channels)
-        self.eval_model = net_builder(num_classes=num_classes, in_channels=in_channels)
+        try:
+            self.eval_model = net_builder(
+                num_classes=num_classes, in_channels=in_channels, pretrained=False
+            )
+        except TypeError:
+            self.eval_model = net_builder(num_classes=num_classes, in_channels=in_channels)
         self.T = T
         self.p_cutoff = p_cutoff
         self.lambda_u = lambda_u

anomaly_match/pretrained_cache/models--timm--tf_efficientnet_lite0.in1k/refs/main

@@ -0,0 +1 @@
+e074755b39c3cd8ca77d917a7cfb62e70da6bf88

anomaly_match/pretrained_cache/models--timm--tf_efficientnet_lite0.in1k/snapshots/e074755b39c3cd8ca77d917a7cfb62e70da6bf88/model.safetensors

@@ -0,0 +1 @@
+../../blobs/1b2a452450b5cf3f8f42da5b6214e71d5fda5666bef4dfae0699407e5bef13e1

anomaly_match/utils/get_net_builder.py

@@ -4,10 +4,16 @@
 #   is part of this source code package. No part of the package, including
 #   this file, may be copied, modified, propagated, or distributed except according to
 #   the terms contained in the file 'LICENCE.txt'.
+from pathlib import Path
+
 import timm
 import torch.nn as nn
 from loguru import logger
 
+# Local cache directory for pretrained weights, avoids flock issues on NFS filesystems.
+_PACKAGE_DIR = Path(__file__).resolve().parent.parent
+_PRETRAINED_CACHE_DIR = _PACKAGE_DIR / "pretrained_cache"
+
 # Mapping from AnomalyMatch net names to timm model identifiers.
 # efficientnet-lite variants use the tf_ prefix for TF-style same-padding,
 # which is backward compatible with the previous efficientnet_lite_pytorch package.
@@ -113,7 +119,7 @@ def get_net_builder(net_name, pretrained=False, in_channels=3):
     if net_name == "test-cnn":
         logger.debug("Using test-cnn model (for testing only)")
 
-        def build_test_cnn(num_classes, in_channels):
+        def build_test_cnn(num_classes, in_channels, pretrained=None):
             return TestCNN(num_classes=num_classes, in_channels=in_channels)
 
         return build_test_cnn
@@ -124,10 +130,33 @@ def build_test_cnn(num_classes, in_channels):
         f"(timm: {timm_name})"
     )
 
-    def build_model(num_classes, in_channels, _timm_name=timm_name, _pretrained=use_pretrained):
+    def build_model(
+        num_classes, in_channels, _timm_name=timm_name, _pretrained=use_pretrained, pretrained=None
+    ):
+        effective_pretrained = pretrained if pretrained is not None else _pretrained
+        if effective_pretrained:
+            try:
+                return timm.create_model(
+                    _timm_name,
+                    pretrained=True,
+                    num_classes=num_classes,
+                    in_chans=in_channels,
+                    cache_dir=str(_PRETRAINED_CACHE_DIR),
+                )
+            except Exception:
+                logger.warning(
+                    f"Bundled pretrained weights not available (clone with git-lfs to avoid "
+                    f"re-downloading). Downloading {_timm_name} from HuggingFace."
+                )
+                return timm.create_model(
+                    _timm_name,
+                    pretrained=True,
+                    num_classes=num_classes,
+                    in_chans=in_channels,
+                )
         return timm.create_model(
             _timm_name,
-            pretrained=_pretrained,
+            pretrained=False,
             num_classes=num_classes,
             in_chans=in_channels,
         )