Swin transformer conformance test

openvinotoolkit · Aug 8, 2024 · cf4aece · cf4aece
1 parent faeb768
commit cf4aece
Show file tree

Hide file tree

Showing 4 changed files with 26 additions and 7 deletions.
diff --git a/tests/post_training/data/ptq_reference_data.yaml b/tests/post_training/data/ptq_reference_data.yaml
@@ -34,6 +34,12 @@ torchvision/resnet18_backend_CUDA_TORCH:
  metric_value: 0.69152
 torchvision/resnet18_backend_FX_TORCH:
  metric_value: 0.6946
+torchvision/swin_v2_s_backend_FP32:
+ metric_value: 0.83712
+torchvision/swin_v2_s_backend_OV:
+ metric_value: 0.83638
+torchvision/swin_v2_s_backend_FX_TORCH:
+ metric_value: 0.82908
 timm/crossvit_9_240_backend_CUDA_TORCH:
  metric_value: 0.689
 timm/crossvit_9_240_backend_FP32:

diff --git a/tests/post_training/model_scope.py b/tests/post_training/model_scope.py
@@ -75,6 +75,17 @@
  "backends": [BackendType.FX_TORCH, BackendType.TORCH, BackendType.CUDA_TORCH, BackendType.OV, BackendType.ONNX],
  "batch_size": 128,
  },
+ {
+ "reported_name": "torchvision/swin_v2_s",
+ "model_id": "swin_v2_s",
+ "pipeline_cls": ImageClassificationTorchvision,
+ "compression_params": {
+ "model_type": ModelType.TRANSFORMER,
+ "advanced_parameters": AdvancedQuantizationParameters(smooth_quant_alpha=0.5),
+ },
+ "backends": [BackendType.FX_TORCH, BackendType.OV],
+ "batch_size": 1,
+ },
  # Timm models
  {
  "reported_name": "timm/crossvit_9_240",

diff --git a/tests/post_training/pipelines/image_classification_torchvision.py b/tests/post_training/pipelines/image_classification_torchvision.py
@@ -43,13 +43,14 @@ def prepare_model(self) -> None:
  model = model_cls(weights=self.model_weights)
  model.eval()
 
- self.static_input_size = [self.batch_size, 3, 224, 224]
+ default_input_size = [self.batch_size, 3, 224, 224]
+ self.dummy_tensor = self.model_weights.transforms()(torch.rand(default_input_size))
+ self.static_input_size = list(self.dummy_tensor.shape)
+
  self.input_size = self.static_input_size.copy()
  if self.batch_size > 1: # Dynamic batch_size shape export
  self.input_size[0] = -1
 
- self.dummy_tensor = torch.rand(self.static_input_size)
-
  if self.backend == BackendType.FX_TORCH:
  with torch.no_grad():
  with disable_patching():

diff --git a/tests/post_training/test_quantize_conformance.py b/tests/post_training/test_quantize_conformance.py
@@ -247,7 +247,8 @@ def test_ptq_quantization(
  err_msg = None
  test_model_param = None
  start_time = time.perf_counter()
- try:
+ if True:
+ # try:
  if test_case_name not in ptq_reference_data:
  raise nncf.ValidationError(f"{test_case_name} does not exist in 'reference_data.yaml'")
  test_model_param = PTQ_TEST_CASES[test_case_name]
@@ -271,9 +272,9 @@ def test_ptq_quantization(
  )
  pipeline: BaseTestPipeline = pipeline_cls(**pipeline_kwargs)
  pipeline.run()
- except Exception as e:
- err_msg = str(e)
- traceback.print_exc()
+ # except Exception as e:
+ # err_msg = str(e)
+ # traceback.print_exc()
 
  if pipeline is not None:
  pipeline.cleanup_cache()