[`Detr`] Fix detr BatchNorm replacement issue (#25230)

* fix detr weird issue

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix copies

* fix copies

---------
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/models/conditional_detr/modeling_conditional_detr.py

@@ -310,19 +310,27 @@ class ConditionalDetrFrozenBatchNorm2d(nn.Module):
 
 
 # Copied from transformers.models.detr.modeling_detr.replace_batch_norm with Detr->ConditionalDetr
-def replace_batch_norm(m, name=""):
-    for attr_str in dir(m):
-        target_attr = getattr(m, attr_str)
-        if isinstance(target_attr, nn.BatchNorm2d):
-            frozen = ConditionalDetrFrozenBatchNorm2d(target_attr.num_features)
-            bn = getattr(m, attr_str)
-            frozen.weight.data.copy_(bn.weight)
-            frozen.bias.data.copy_(bn.bias)
-            frozen.running_mean.data.copy_(bn.running_mean)
-            frozen.running_var.data.copy_(bn.running_var)
-            setattr(m, attr_str, frozen)
-    for n, ch in m.named_children():
-        replace_batch_norm(ch, n)
+def replace_batch_norm(model):
+    r"""
+    Recursively replace all `torch.nn.BatchNorm2d` with `ConditionalDetrFrozenBatchNorm2d`.
+
+    Args:
+        model (torch.nn.Module):
+            input model
+    """
+    for name, module in model.named_children():
+        if isinstance(module, nn.BatchNorm2d):
+            new_module = ConditionalDetrFrozenBatchNorm2d(module.num_features)
+
+            new_module.weight.data.copy_(module.weight)
+            new_module.bias.data.copy_(module.bias)
+            new_module.running_mean.data.copy_(module.running_mean)
+            new_module.running_var.data.copy_(module.running_var)
+
+            model._modules[name] = new_module
+
+        if len(list(module.children())) > 0:
+            replace_batch_norm(module)
 
 
 # Copied from transformers.models.detr.modeling_detr.DetrConvEncoder

src/transformers/models/deformable_detr/modeling_deformable_detr.py

@@ -357,19 +357,27 @@ class DeformableDetrFrozenBatchNorm2d(nn.Module):
 
 
 # Copied from transformers.models.detr.modeling_detr.replace_batch_norm with Detr->DeformableDetr
-def replace_batch_norm(m, name=""):
-    for attr_str in dir(m):
-        target_attr = getattr(m, attr_str)
-        if isinstance(target_attr, nn.BatchNorm2d):
-            frozen = DeformableDetrFrozenBatchNorm2d(target_attr.num_features)
-            bn = getattr(m, attr_str)
-            frozen.weight.data.copy_(bn.weight)
-            frozen.bias.data.copy_(bn.bias)
-            frozen.running_mean.data.copy_(bn.running_mean)
-            frozen.running_var.data.copy_(bn.running_var)
-            setattr(m, attr_str, frozen)
-    for n, ch in m.named_children():
-        replace_batch_norm(ch, n)
+def replace_batch_norm(model):
+    r"""
+    Recursively replace all `torch.nn.BatchNorm2d` with `DeformableDetrFrozenBatchNorm2d`.
+
+    Args:
+        model (torch.nn.Module):
+            input model
+    """
+    for name, module in model.named_children():
+        if isinstance(module, nn.BatchNorm2d):
+            new_module = DeformableDetrFrozenBatchNorm2d(module.num_features)
+
+            new_module.weight.data.copy_(module.weight)
+            new_module.bias.data.copy_(module.bias)
+            new_module.running_mean.data.copy_(module.running_mean)
+            new_module.running_var.data.copy_(module.running_var)
+
+            model._modules[name] = new_module
+
+        if len(list(module.children())) > 0:
+            replace_batch_norm(module)
 
 
 class DeformableDetrConvEncoder(nn.Module):

src/transformers/models/deta/modeling_deta.py

@@ -295,19 +295,27 @@ class DetaFrozenBatchNorm2d(nn.Module):
 
 
 # Copied from transformers.models.detr.modeling_detr.replace_batch_norm with Detr->Deta
-def replace_batch_norm(m, name=""):
-    for attr_str in dir(m):
-        target_attr = getattr(m, attr_str)
-        if isinstance(target_attr, nn.BatchNorm2d):
-            frozen = DetaFrozenBatchNorm2d(target_attr.num_features)
-            bn = getattr(m, attr_str)
-            frozen.weight.data.copy_(bn.weight)
-            frozen.bias.data.copy_(bn.bias)
-            frozen.running_mean.data.copy_(bn.running_mean)
-            frozen.running_var.data.copy_(bn.running_var)
-            setattr(m, attr_str, frozen)
-    for n, ch in m.named_children():
-        replace_batch_norm(ch, n)
+def replace_batch_norm(model):
+    r"""
+    Recursively replace all `torch.nn.BatchNorm2d` with `DetaFrozenBatchNorm2d`.
+
+    Args:
+        model (torch.nn.Module):
+            input model
+    """
+    for name, module in model.named_children():
+        if isinstance(module, nn.BatchNorm2d):
+            new_module = DetaFrozenBatchNorm2d(module.num_features)
+
+            new_module.weight.data.copy_(module.weight)
+            new_module.bias.data.copy_(module.bias)
+            new_module.running_mean.data.copy_(module.running_mean)
+            new_module.running_var.data.copy_(module.running_var)
+
+            model._modules[name] = new_module
+
+        if len(list(module.children())) > 0:
+            replace_batch_norm(module)
 
 
 class DetaBackboneWithPositionalEncodings(nn.Module):

src/transformers/models/detr/modeling_detr.py

@@ -304,19 +304,27 @@ class DetrFrozenBatchNorm2d(nn.Module):
         return x * scale + bias
 
 
-def replace_batch_norm(m, name=""):
-    for attr_str in dir(m):
-        target_attr = getattr(m, attr_str)
-        if isinstance(target_attr, nn.BatchNorm2d):
-            frozen = DetrFrozenBatchNorm2d(target_attr.num_features)
-            bn = getattr(m, attr_str)
-            frozen.weight.data.copy_(bn.weight)
-            frozen.bias.data.copy_(bn.bias)
-            frozen.running_mean.data.copy_(bn.running_mean)
-            frozen.running_var.data.copy_(bn.running_var)
-            setattr(m, attr_str, frozen)
-    for n, ch in m.named_children():
-        replace_batch_norm(ch, n)
+def replace_batch_norm(model):
+    r"""
+    Recursively replace all `torch.nn.BatchNorm2d` with `DetrFrozenBatchNorm2d`.
+
+    Args:
+        model (torch.nn.Module):
+            input model
+    """
+    for name, module in model.named_children():
+        if isinstance(module, nn.BatchNorm2d):
+            new_module = DetrFrozenBatchNorm2d(module.num_features)
+
+            new_module.weight.data.copy_(module.weight)
+            new_module.bias.data.copy_(module.bias)
+            new_module.running_mean.data.copy_(module.running_mean)
+            new_module.running_var.data.copy_(module.running_var)
+
+            model._modules[name] = new_module
+
+        if len(list(module.children())) > 0:
+            replace_batch_norm(module)
 
 
 class DetrConvEncoder(nn.Module):

src/transformers/models/table_transformer/modeling_table_transformer.py

@@ -239,19 +239,27 @@ class TableTransformerFrozenBatchNorm2d(nn.Module):
 
 
 # Copied from transformers.models.detr.modeling_detr.replace_batch_norm with Detr->TableTransformer
-def replace_batch_norm(m, name=""):
-    for attr_str in dir(m):
-        target_attr = getattr(m, attr_str)
-        if isinstance(target_attr, nn.BatchNorm2d):
-            frozen = TableTransformerFrozenBatchNorm2d(target_attr.num_features)
-            bn = getattr(m, attr_str)
-            frozen.weight.data.copy_(bn.weight)
-            frozen.bias.data.copy_(bn.bias)
-            frozen.running_mean.data.copy_(bn.running_mean)
-            frozen.running_var.data.copy_(bn.running_var)
-            setattr(m, attr_str, frozen)
-    for n, ch in m.named_children():
-        replace_batch_norm(ch, n)
+def replace_batch_norm(model):
+    r"""
+    Recursively replace all `torch.nn.BatchNorm2d` with `TableTransformerFrozenBatchNorm2d`.
+
+    Args:
+        model (torch.nn.Module):
+            input model
+    """
+    for name, module in model.named_children():
+        if isinstance(module, nn.BatchNorm2d):
+            new_module = TableTransformerFrozenBatchNorm2d(module.num_features)
+
+            new_module.weight.data.copy_(module.weight)
+            new_module.bias.data.copy_(module.bias)
+            new_module.running_mean.data.copy_(module.running_mean)
+            new_module.running_var.data.copy_(module.running_var)
+
+            model._modules[name] = new_module
+
+        if len(list(module.children())) > 0:
+            replace_batch_norm(module)
 
 
 # Copied from transformers.models.detr.modeling_detr.DetrConvEncoder with Detr->TableTransformer