Relax the check to enable Pad+Conv fusion with RELU6 to avoid numeric diffs in FP16.

PiperOrigin-RevId: 834438718

Author: junjiang-lab

ai_edge_torch/_convert/test/test_convert.py

@@ -100,6 +100,15 @@ def test_convert_resnet18(self):
         model_coverage.compare_tflite_torch(edge_model, torch_module, args)
     )
 
+  def test_convert_mobilenet_v2(self):
+    args = (torch.randn(4, 3, 224, 224),)
+    torch_module = torchvision.models.mobilenet_v2().eval()
+    edge_model = ai_edge_torch.convert(torch_module, args)
+
+    self.assertTrue(
+        model_coverage.compare_tflite_torch(edge_model, torch_module, args)
+    )
+
   def test_signature_args_ordering(self):
     """Tests conversion of a model with more than 10 arguments."""