Fix _init_weights for ResNetPreTrainedModel (#31851)

* init

* test

---------

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/models/regnet/modeling_regnet.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 """PyTorch RegNet model."""
 
+import math
 from typing import Optional
 
 import torch
@@ -284,6 +285,13 @@ class RegNetPreTrainedModel(PreTrainedModel):
     def _init_weights(self, module):
         if isinstance(module, nn.Conv2d):
             nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
+        # copied from the `reset_parameters` method of `class Linear(Module)` in `torch`.
+        elif isinstance(module, nn.Linear):
+            nn.init.kaiming_uniform_(module.weight, a=math.sqrt(5))
+            if module.bias is not None:
+                fan_in, _ = nn.init._calculate_fan_in_and_fan_out(module.weight)
+                bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
+                nn.init.uniform_(module.bias, -bound, bound)
         elif isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
             nn.init.constant_(module.weight, 1)
             nn.init.constant_(module.bias, 0)

src/transformers/models/resnet/modeling_resnet.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 """PyTorch ResNet model."""
 
+import math
 from typing import Optional
 
 import torch
@@ -274,6 +275,13 @@ class ResNetPreTrainedModel(PreTrainedModel):
     def _init_weights(self, module):
         if isinstance(module, nn.Conv2d):
             nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
+        # copied from the `reset_parameters` method of `class Linear(Module)` in `torch`.
+        elif isinstance(module, nn.Linear):
+            nn.init.kaiming_uniform_(module.weight, a=math.sqrt(5))
+            if module.bias is not None:
+                fan_in, _ = nn.init._calculate_fan_in_and_fan_out(module.weight)
+                bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
+                nn.init.uniform_(module.bias, -bound, bound)
         elif isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
             nn.init.constant_(module.weight, 1)
             nn.init.constant_(module.bias, 0)

src/transformers/models/rt_detr/modeling_rt_detr_resnet.py

@@ -17,6 +17,7 @@ PyTorch RTDetr specific ResNet model. The main difference between hugginface Res
 See https://github.com/lyuwenyu/RT-DETR/blob/5b628eaa0a2fc25bdafec7e6148d5296b144af85/rtdetr_pytorch/src/nn/backbone/presnet.py#L126 for details.
 """
 
+import math
 from typing import Optional
 
 from torch import Tensor, nn
@@ -323,6 +324,13 @@ class RTDetrResNetPreTrainedModel(PreTrainedModel):
     def _init_weights(self, module):
         if isinstance(module, nn.Conv2d):
             nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
+        # copied from the `reset_parameters` method of `class Linear(Module)` in `torch`.
+        elif isinstance(module, nn.Linear):
+            nn.init.kaiming_uniform_(module.weight, a=math.sqrt(5))
+            if module.bias is not None:
+                fan_in, _ = nn.init._calculate_fan_in_and_fan_out(module.weight)
+                bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
+                nn.init.uniform_(module.bias, -bound, bound)
         elif isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
             nn.init.constant_(module.weight, 1)
             nn.init.constant_(module.bias, 0)

tests/test_modeling_common.py

@@ -3167,9 +3167,47 @@ class ModelTesterMixin:
         configs_no_init = _config_zero_init(config)
 
         for model_class in self.all_model_classes:
-            if model_class.__name__ not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES):
+            mappings = [
+                MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
+                MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
+                MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES,
+                MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES,
+            ]
+            is_classication_model = any(model_class.__name__ in get_values(mapping) for mapping in mappings)
+
+            if not is_classication_model:
                 continue
 
+            # TODO: ydshieh
+            is_special_classes = model_class.__name__ in [
+                "wav2vec2.masked_spec_embed",
+                "Wav2Vec2ForSequenceClassification",
+                "CLIPForImageClassification",
+                "RegNetForImageClassification",
+                "ResNetForImageClassification",
+            ]
+            special_param_names = [
+                r"wav2vec2\.masked_spec_embed",
+                r"wav2vec2\.feature_extractor\.conv_layers\..+\.conv\.weight",
+                r"wav2vec2\.feature_projection\.projection\.weight",
+                r"wav2vec2\.feature_projection\.projection\.bias",
+                r"wav2vec2\.encoder\.pos_conv_embed\.conv\.parametrizations\.weight\.original.",
+                r"classifier\.weight",
+                r"regnet\.embedder\.embedder\.convolution\.weight",
+                r"regnet\.encoder\.stages\..+\.layers\..+\.layer\..+\.convolution\.weight",
+                r"regnet\.encoder\.stages\..+\.layers\..+\.shortcut\.convolution\.weight",
+                r"regnet\.encoder\.stages\..+\.layers\..+\.layer\..+\.attention\..+\.weight",
+                r"regnet\.encoder\.stages\..+\.layers\..+\.layer\..+\.attention\..+\.bias",
+                r"classifier\..+\.weight",
+                r"classifier\..+\.bias",
+                r"resnet\.embedder\.embedder\.convolution\.weight",
+                r"resnet\.encoder\.stages\..+\.layers\..+\.shortcut\.convolution\.weight",
+                r"resnet\.encoder\.stages\..+\.layers\..+\.layer\..+\.convolution\.weight",
+                r"resnet\.encoder\.stages\..+\.layers\..+\.shortcut\.convolution\.weight",
+                r"resnet\.encoder\.stages\..+\.layers\..+\.layer\..+\.attention\..+\.weight",
+                r"resnet\.encoder\.stages\..+\.layers\..+\.layer\..+\.attention\..+\.bias",
+            ]
+
             with self.subTest(msg=f"Testing {model_class}"):
                 with tempfile.TemporaryDirectory() as tmp_dir:
                     model = model_class(configs_no_init)
@@ -3177,23 +3215,37 @@ class ModelTesterMixin:
 
                     # Fails when we don't set ignore_mismatched_sizes=True
                     with self.assertRaises(RuntimeError):
-                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
+                        new_model = model_class.from_pretrained(tmp_dir, num_labels=42)
 
                     logger = logging.get_logger("transformers.modeling_utils")
 
                     with CaptureLogger(logger) as cl:
-                        new_model = AutoModelForSequenceClassification.from_pretrained(
-                            tmp_dir, num_labels=42, ignore_mismatched_sizes=True
-                        )
+                        new_model = model_class.from_pretrained(tmp_dir, num_labels=42, ignore_mismatched_sizes=True)
                     self.assertIn("the shapes did not match", cl.out)
 
                     for name, param in new_model.named_parameters():
                         if param.requires_grad:
-                            self.assertIn(
-                                ((param.data.mean() * 1e9).round() / 1e9).item(),
-                                [0.0, 1.0],
-                                msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                            )
+                            param_mean = ((param.data.mean() * 1e9).round() / 1e9).item()
+                            if not (
+                                is_special_classes
+                                and any(len(re.findall(target, name)) > 0 for target in special_param_names)
+                            ):
+                                self.assertIn(
+                                    param_mean,
+                                    [0.0, 1.0],
+                                    msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                                )
+                            else:
+                                self.assertGreaterEqual(
+                                    param_mean,
+                                    -1.0,
+                                    msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                                )
+                                self.assertLessEqual(
+                                    param_mean,
+                                    1.0,
+                                    msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                                )
 
     def test_matched_shapes_have_loaded_weights_when_some_mismatched_shapes_exist(self):
         # 1. Create a dummy class. Should have buffers as well? To make sure we test __init__