Fix model-intrinsic params in the model (input channels / class count / network architecture) instead of exposing them in the config

body/Model.py

@@ -11,6 +11,10 @@ def _replace_unpicklable_identities(module: nn.Module) -> None:
             child.nonlin2 = nn.Identity()
     
 class ResEnc(network.Network):
+    # Intrinsic to the pretrained weights (5-channel 2.5D input, 12 body structures), not tunables — fixed here.
+    IN_CHANNELS = 5
+    NUM_CLASSES = 12
+
     def __init__(
         self,
         optimizer: network.OptimizerLoader = network.OptimizerLoader(),
@@ -18,25 +22,23 @@ class ResEnc(network.Network):
             "default:ReduceLROnPlateau": network.LRSchedulersLoader(0)
         },
         outputs_criterions: dict[str, network.TargetCriterionsLoader] = {"default": network.TargetCriterionsLoader()},
-        in_channels: int = 5,
-        nb_class: int = 132,
     ) -> None:
         super().__init__(
-            in_channels=in_channels,
+            in_channels=ResEnc.IN_CHANNELS,
             optimizer=optimizer,
             schedulers=schedulers,
             outputs_criterions=outputs_criterions,
             dim=2,
         )
         self.add_module("DecoderOutputs", ResidualEncoderUNet(
-                input_channels=in_channels,
+                input_channels=ResEnc.IN_CHANNELS,
                 n_stages=6,
                 features_per_stage=(24, 48, 96, 192, 256, 256),
                 conv_op=nn.Conv2d,
                 kernel_sizes=3,
                 strides=(1, 2, 2, 2, 2, 2),
                 n_blocks_per_stage=(1, 2, 2, 3, 3, 3),
-                num_classes=nb_class,
+                num_classes=ResEnc.NUM_CLASSES,
                 n_conv_per_stage_decoder=(1, 1, 1, 1, 1),
                 conv_bias=True,
                 norm_op=nn.InstanceNorm2d,

body/Prediction.yml

@@ -2,8 +2,6 @@ Predictor:
   Model:
     classpath: Model:ResEnc
     ResEnc:
-      in_channels: 5
-      nb_class: 12
       outputs_criterions: None
   Dataset:
     groups_src: