Fix RT-DETR weights initialization (huggingface#31724)

* Fix init for rt-detr heads

* Fixup

* Add separate prior_prob value to config for initialization

* Add bbox init

* Change to 1 / num_labels init

* Adjust weights init test

* Fix style for test

src/transformers/models/rt_detr/configuration_rt_detr.py

@@ -37,6 +37,9 @@ class RTDetrConfig(PretrainedConfig):
     Args:
         initializer_range (`float`, *optional*, defaults to 0.01):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_bias_prior_prob (`float`, *optional*):
+            The prior probability used by the bias initializer to initialize biases for `enc_score_head` and `class_embed`.
+            If `None`, `prior_prob` computed as `prior_prob = 1 / (num_labels + 1)` while initializing model weights.
         layer_norm_eps (`float`, *optional*, defaults to 1e-05):
             The epsilon used by the layer normalization layers.
         batch_norm_eps (`float`, *optional*, defaults to 1e-05):
@@ -179,6 +182,7 @@ class RTDetrConfig(PretrainedConfig):
     def __init__(
         self,
         initializer_range=0.01,
+        initializer_bias_prior_prob=None,
         layer_norm_eps=1e-5,
         batch_norm_eps=1e-5,
         # backbone
@@ -239,6 +243,7 @@ def __init__(
         **kwargs,
     ):
         self.initializer_range = initializer_range
+        self.initializer_bias_prior_prob = initializer_bias_prior_prob
         self.layer_norm_eps = layer_norm_eps
         self.batch_norm_eps = batch_norm_eps
         # backbone

src/transformers/models/rt_detr/modeling_rt_detr.py

@@ -1148,14 +1148,27 @@ class RTDetrPreTrainedModel(PreTrainedModel):
     def _init_weights(self, module):
         """Initalize the weights"""
 
-        """initialize conv/fc bias value according to a given probability value."""
-        if isinstance(module, nn.Linear) and hasattr(module, "class_embed"):
-            prior_prob = self.config.initializer_range
+        """initialize linear layer bias value according to a given probability value."""
+        if isinstance(module, (RTDetrForObjectDetection, RTDetrDecoder)):
+            if module.class_embed is not None:
+                for layer in module.class_embed:
+                    prior_prob = self.config.initializer_bias_prior_prob or 1 / (self.config.num_labels + 1)
+                    bias = float(-math.log((1 - prior_prob) / prior_prob))
+                    nn.init.xavier_uniform_(layer.weight)
+                    nn.init.constant_(layer.bias, bias)
+
+            if module.bbox_embed is not None:
+                for layer in module.bbox_embed:
+                    nn.init.constant_(layer.layers[-1].weight, 0)
+                    nn.init.constant_(layer.layers[-1].bias, 0)
+
+        if isinstance(module, RTDetrModel):
+            prior_prob = self.config.initializer_bias_prior_prob or 1 / (self.config.num_labels + 1)
             bias = float(-math.log((1 - prior_prob) / prior_prob))
-            nn.init.xavier_uniform_(module.weight)
-            if module.bias is not None:
-                nn.init.constant_(module.bias, bias)
-        elif isinstance(module, (nn.Linear, nn.Conv2d, nn.BatchNorm2d)):
+            nn.init.xavier_uniform_(module.enc_score_head.weight)
+            nn.init.constant_(module.enc_score_head.bias, bias)
+
+        if isinstance(module, (nn.Linear, nn.Conv2d, nn.BatchNorm2d)):
             module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
             if module.bias is not None:
                 module.bias.data.zero_()

tests/models/rt_detr/test_modeling_rt_detr.py

@@ -584,6 +584,11 @@ def test_initialization(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
         configs_no_init = _config_zero_init(config)
+        configs_no_init.initializer_bias_prior_prob = 0.2
+        bias_value = -1.3863  # log_e ((1 - 0.2) / 0.2)
+
+        failed_cases = []
+
         for model_class in self.all_model_classes:
             model = model_class(config=configs_no_init)
             # Skip the check for the backbone
@@ -594,20 +599,36 @@ def test_initialization(self):
 
             for name, param in model.named_parameters():
                 if param.requires_grad:
-                    if (
+                    if ("class_embed" in name and "bias" in name) or "enc_score_head.bias" in name:
+                        bias_tensor = torch.full_like(param.data, bias_value)
+                        if not torch.allclose(param.data, bias_tensor, atol=1e-4):
+                            failed_cases.append(
+                                f"Parameter {name} of model {model_class} seems not properly initialized. "
+                                f"Biases should be initialized to {bias_value}, got {param.data}"
+                            )
+                    elif (
                         "level_embed" in name
                         or "sampling_offsets.bias" in name
                         or "value_proj" in name
                         or "output_proj" in name
                         or "reference_points" in name
+                        or "enc_score_head.weight" in name
+                        or ("class_embed" in name and "weight" in name)
                         or name in backbone_params
                     ):
                         continue
-                    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",
-                    )
+                    else:
+                        mean = param.data.mean()
+                        round_mean = (mean * 1e9).round() / 1e9
+                        round_mean = round_mean.item()
+                        if round_mean not in [0.0, 1.0]:
+                            failed_cases.append(
+                                f"Parameter {name} of model {model_class} seems not properly initialized. "
+                                f"Mean is {round_mean}, but should be in [0, 1]"
+                            )
+
+        message = "\n" + "\n".join(failed_cases)
+        self.assertTrue(not failed_cases, message)
 
     @parameterized.expand(["float32", "float16", "bfloat16"])
     @require_torch_gpu