ORTOptimizer for the model type Segformer (#1820)

* add segformer * black * make format * decoder_hidden_size not a list * tests pass now * use max * use zero --------- Co-authored-by: Zach Deane-Mayer <zach@ai-insight-solutions.com>
huggingface · Jun 5, 2024 · ac951ca · ac951ca
1 parent 7a0757a
commit ac951ca
Show file tree

Hide file tree

Showing 4 changed files with 25 additions and 4 deletions.
diff --git a/optimum/onnxruntime/modeling_ort.py b/optimum/onnxruntime/modeling_ort.py
@@ -1746,13 +1746,18 @@ class ORTModelForSemanticSegmentation(ORTModel):
             checkpoint="optimum/segformer-b0-finetuned-ade-512-512",
         )
     )
-    def forward(self, **kwargs):
-        use_torch = isinstance(next(iter(kwargs.values())), torch.Tensor)
+    def forward(
+        self,
+        pixel_values: Union[torch.Tensor, np.ndarray],
+        **kwargs,
+    ):
+        use_torch = isinstance(pixel_values, torch.Tensor)
         self.raise_on_numpy_input_io_binding(use_torch)
 
         if self.device.type == "cuda" and self.use_io_binding:
             io_binding = IOBindingHelper.prepare_io_binding(
                 self,
+                pixel_values,
                 **kwargs,
                 ordered_input_names=self._ordered_input_names,
             )
@@ -1769,7 +1774,7 @@ def forward(self, **kwargs):
             # converts output to namedtuple for pipelines post-processing
             return SemanticSegmenterOutput(logits=outputs["logits"])
         else:
-            onnx_inputs = self._prepare_onnx_inputs(use_torch=use_torch, **kwargs)
+            onnx_inputs = self._prepare_onnx_inputs(use_torch=use_torch, pixel_values=pixel_values, **kwargs)
 
             # run inference
             onnx_outputs = self.model.run(None, onnx_inputs)

diff --git a/optimum/onnxruntime/utils.py b/optimum/onnxruntime/utils.py
@@ -128,6 +128,7 @@ class ORTConfigManager:
         "nystromformer": "bert",
         "pegasus": "bert",
         "roberta": "bert",
+        "segformer": "vit",
         "t5": "bert",
         "vit": "vit",
         "whisper": "bart",

diff --git a/optimum/utils/normalized_config.py b/optimum/utils/normalized_config.py
@@ -102,6 +102,19 @@ class NormalizedVisionConfig(NormalizedConfig):
     INPUT_SIZE = "input_size"
 
 
+class NormalizedSegformerConfig(NormalizedVisionConfig):
+    NUM_ATTENTION_HEADS = "num_attention_heads"
+    HIDDEN_SIZE = "hidden_sizes"
+
+    # If the attribute is a list, return 0
+    # 0 means let the optimizer infer the correct value based on the model graph
+    def __getattr__(self, attr_name):
+        attr_value = super().__getattr__(attr_name)
+        if isinstance(attr_value, list):
+            attr_value = 0
+        return attr_value
+
+
 class NormalizedTextAndVisionConfig(NormalizedTextConfig, NormalizedVisionConfig):
     TEXT_CONFIG = None
     VISION_CONFIG = None
@@ -203,7 +216,6 @@ class NormalizedConfigManager:
         'owlvit',
         'perceiver',
         'roformer',
-        'segformer',
         'squeezebert',
         'table-transformer',
     """
@@ -258,6 +270,7 @@ class NormalizedConfigManager:
         "regnet": NormalizedVisionConfig,
         "resnet": NormalizedVisionConfig,
         "roberta": NormalizedTextConfig,
+        "segformer": NormalizedSegformerConfig,
         "speech-to-text": SpeechToTextLikeNormalizedTextConfig,
         "splinter": NormalizedTextConfig,
         "t5": T5LikeNormalizedTextConfig,

diff --git a/tests/onnxruntime/test_optimization.py b/tests/onnxruntime/test_optimization.py
@@ -36,6 +36,7 @@
     AutoOptimizationConfig,
     ORTConfig,
     ORTModelForImageClassification,
+    ORTModelForSemanticSegmentation,
     ORTModelForSequenceClassification,
     ORTOptimizer,
 )
@@ -171,6 +172,7 @@ def test_compare_original_seq2seq_model_with_optimized_model(self, model_cls, mo
 
     # Contribution note: Please add test models in alphabetical order. Find test models here: https://huggingface.co/hf-internal-testing.
     SUPPORTED_IMAGE_ARCHITECTURES_WITH_MODEL_ID = (
+        (ORTModelForSemanticSegmentation, "hf-internal-testing/tiny-random-segformer"),
         (ORTModelForImageClassification, "hf-internal-testing/tiny-random-vit"),
     )