fix: add VitPose config-driven dummy inputs

examples/recipes/nielsr_vitpose-base-simple/cpu/cpu/keypoint-detection_fp16_config.json

@@ -0,0 +1,67 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "pixel_values",
+        "dtype": "float32",
+        "shape": [
+          1,
+          3,
+          256,
+          192
+        ],
+        "value_range": [
+          -2.1179039478302,
+          2.640000343322754
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "heatmaps"
+      }
+    ],
+    "compatibility": {
+      "transformers_attention": "eager"
+    }
+  },
+  "optim": {},
+  "quant": {
+    "mode": "fp16",
+    "samples": 10,
+    "calibration_method": "minmax",
+    "weight_type": "uint8",
+    "activation_type": "uint8",
+    "per_channel": false,
+    "symmetric": false,
+    "weight_symmetric": null,
+    "activation_symmetric": null,
+    "save_calibration": false,
+    "distribution": "uniform",
+    "seed": null,
+    "calibration_load_path": null,
+    "calibration_save_path": null,
+    "op_types_to_quantize": null,
+    "nodes_to_exclude": null,
+    "task": "keypoint-detection",
+    "model_id": "nielsr/vitpose-base-simple",
+    "model_type": "vitpose",
+    "fp16_keep_io_types": true,
+    "fp16_op_block_list": null
+  },
+  "compile": null,
+  "loader": {
+    "task": "keypoint-detection",
+    "model_class": "VitPoseForPoseEstimation",
+    "model_type": "vitpose"
+  }
+}

examples/recipes/nielsr_vitpose-base-simple/cpu/cpu/keypoint-detection_fp32_config.json

@@ -0,0 +1,45 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "pixel_values",
+        "dtype": "float32",
+        "shape": [
+          1,
+          3,
+          256,
+          192
+        ],
+        "value_range": [
+          -2.1179039478302,
+          2.640000343322754
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "heatmaps"
+      }
+    ],
+    "compatibility": {
+      "transformers_attention": "eager"
+    }
+  },
+  "optim": {},
+  "quant": null,
+  "compile": null,
+  "loader": {
+    "task": "keypoint-detection",
+    "model_class": "VitPoseForPoseEstimation",
+    "model_type": "vitpose"
+  }
+}

src/winml/modelkit/models/hf/vitpose.py

@@ -33,8 +33,19 @@
 
 from __future__ import annotations
 
+from optimum.exporters.onnx.model_configs import VitPoseOnnxConfig
+from optimum.utils.input_generators import DummyVisionInputGenerator
 from transformers import VitPoseForPoseEstimation
 
+from ...export import register_onnx_overwrite
+
+
+@register_onnx_overwrite("vitpose", "keypoint-detection", library_name="transformers")
+class VitPoseIOConfig(VitPoseOnnxConfig):  # type: ignore[misc]  # optimum base is untyped
+    """Generate image inputs without reloading a processor from the model ID."""
+
+    DUMMY_INPUT_GENERATOR_CLASSES = (DummyVisionInputGenerator,)
+
 
 # (model_type, task) -> HuggingFace model class
 #

tests/unit/models/test_vitpose_mapping.py

@@ -15,9 +15,13 @@
 
 from unittest.mock import MagicMock
 
+from transformers import VitPoseConfig
+
+from winml.modelkit.export import resolve_io_specs
 from winml.modelkit.loader import resolve_task
 from winml.modelkit.models.hf import MODEL_CLASS_MAPPING
 from winml.modelkit.models.hf.vitpose import MODEL_CLASS_MAPPING as VITPOSE_MAPPING
+from winml.modelkit.models.hf.vitpose import VitPoseIOConfig
 
 
 class TestVitPoseMapping:
@@ -67,3 +71,30 @@ def test_sentinel_registered_in_mapping(self):
             is MODEL_CLASS_MAPPING[("vitpose", "keypoint-detection")]
         )
 
+
+class TestVitPoseIOConfig:
+    def test_resolves_dummy_input_without_loading_processor(self):
+        config = VitPoseConfig(
+            backbone_config={
+                "model_type": "vitpose_backbone",
+                "image_size": [256, 192],
+                "num_channels": 3,
+            },
+            num_labels=17,
+            scale_factor=4,
+            use_simple_decoder=True,
+        )
+
+        specs = resolve_io_specs(
+            "vitpose",
+            "keypoint-detection",
+            config,
+            height=256,
+            width=192,
+        )
+
+        assert specs["input_names"] == ["pixel_values"]
+        assert specs["output_names"] == ["heatmaps"]
+        assert specs["input_shapes"] == [(1, 3, 256, 192)]
+        assert specs["input_dtypes"] == ["float32"]
+        assert VitPoseIOConfig.__name__ == "VitPoseIOConfig"