Upload 146 files

milk10k_effb2_metadata/MILK10K_EFFB2_METADATA_CLI.md

@@ -54,6 +54,7 @@ python train_milk10k_effb2_dual_metadata.py \
 
 Use the dedicated `DualConvNeXtMetadataClassifier` with two ImageNet-initialized
 ConvNeXt Base encoders. When `--image-size` is omitted, ConvNeXt uses 384x384.
+When no branch checkpoint or backend is specified, training uses timm.
 
 ```bash
 python train_milk10k_effb2_dual_metadata.py \

milk10k_effb2_metadata/cli.py

@@ -167,7 +167,7 @@ def parse_args() -> argparse.Namespace:
         "--backbone-backend",
         choices=["auto", "timm", "torchvision"],
         default="auto",
-        help="Backbone implementation used by checkpoints. auto detects timm vs torchvision from checkpoint keys.",
+        help="Backbone implementation. auto detects checkpoint backends and defaults to timm when no checkpoint is passed.",
     )
     parser.add_argument(
         "--imagenet-pretrained",

milk10k_effb2_metadata/milk10k_effb2_metadata/MILK10K_EFFB2_METADATA_CLI.md

@@ -54,6 +54,7 @@ python train_milk10k_effb2_dual_metadata.py \
 
 Use the dedicated `DualConvNeXtMetadataClassifier` with two ImageNet-initialized
 ConvNeXt Base encoders. When `--image-size` is omitted, ConvNeXt uses 384x384.
+When no branch checkpoint or backend is specified, training uses timm.
 
 ```bash
 python train_milk10k_effb2_dual_metadata.py \

milk10k_effb2_metadata/milk10k_effb2_metadata/cli.py

@@ -144,7 +144,7 @@ def parse_args() -> argparse.Namespace:
         "--backbone-backend",
         choices=["auto", "timm", "torchvision"],
         default="auto",
-        help="Backbone implementation used by checkpoints. auto detects timm vs torchvision from checkpoint keys.",
+        help="Backbone implementation. auto detects checkpoint backends and defaults to timm when no checkpoint is passed.",
     )
     parser.add_argument(
         "--imagenet-pretrained",

milk10k_effb2_metadata/milk10k_effb2_metadata/model_setup.py

@@ -50,8 +50,8 @@ def resolve_training_backbone_backends(args: argparse.Namespace, device: torch.d
         )
         return dermoscopic_backend, dermoscopic_backend
     if args.resume_checkpoint is None:
-        print("No branch checkpoints passed; using torchvision backbones initialized from ImageNet weights.")
-        return "torchvision", "torchvision"
+        print("No branch checkpoints passed; using timm backbones initialized from ImageNet weights.")
+        return "timm", "timm"
     checkpoint = torch.load(args.resume_checkpoint.expanduser().resolve(), map_location=device, weights_only=False)
     state = checkpoint["model_state"]
     clinical_backend = infer_branch_backend_from_state(state, "clinical_encoder.")

milk10k_effb2_metadata/model_setup.py

@@ -50,8 +50,8 @@ def resolve_training_backbone_backends(args: argparse.Namespace, device: torch.d
         )
         return dermoscopic_backend, dermoscopic_backend
     if args.resume_checkpoint is None:
-        print("No branch checkpoints passed; using torchvision backbones initialized from ImageNet weights.")
-        return "torchvision", "torchvision"
+        print("No branch checkpoints passed; using timm backbones initialized from ImageNet weights.")
+        return "timm", "timm"
     checkpoint = torch.load(args.resume_checkpoint.expanduser().resolve(), map_location=device, weights_only=False)
     state = checkpoint["model_state"]
     clinical_backend = infer_branch_backend_from_state(state, "clinical_encoder.")