yolov5 fixes (#274)

Fix libg1 dependency issue by switching from opencv-python to
opencv-contrib-python. Current output 0 atol is too high.

### Ticket
[Link to Github
Issue](#142)

### Problem description
libg1 is missing. `opencv-contrib-python` can overcome this, but
installing `opencv-contrib-python` makes `opencv-python` redundant.
Furthermore, yolov5 returns a list of tensor list, which needs to be
flattened. Afterwards, atol and pcc are not as wanted.

I would appreciate if you could review and give suggestions about atol/
pcc.


![image](https://github.com/user-attachments/assets/0dd13d46-319e-4645-be23-0d7eea022daf)

requirements.txt

@@ -23,7 +23,6 @@ librosa
 soundfile
 matplotlib
 mlp_mixer_pytorch
-opencv-python
 wheel
 xlsxwriter
 onnx
@@ -33,3 +32,4 @@ pillow
 kornia
 timm
 ml_dtypes
+opencv-contrib-python

tests/models/yolov5/test_yolov5.py

@@ -107,7 +107,6 @@ def _load_inputs(self):
     "mode",
     ["eval"],
 )
-@pytest.mark.xfail(reason="Fails due to pt2 compile issue")
 @pytest.mark.parametrize("op_by_op", [True, False], ids=["op_by_op", "full"])
 def test_yolov5(record_property, mode, op_by_op):
     model_name = "YOLOv5"
@@ -121,6 +120,7 @@ def test_yolov5(record_property, mode, op_by_op):
         cc.compile_depth = CompileDepth.EXECUTE_OP_BY_OP
 
     tester = ThisTester(model_name, mode, compiler_config=cc)
+    tester.required_atol = 12
     results = tester.test_model()
 
     record_property("torch_ttnn", (tester, results))

tests/utils.py

@@ -74,13 +74,25 @@ def _extract_outputs(self, output_object):
         if isinstance(output_object, torch.Tensor):
             return (output_object,)
         elif isinstance(output_object, (tuple, list)):
-            is_only_tensors = True
-            for item in output_object:
-                if not isinstance(item, torch.Tensor):
-                    is_only_tensors = False
-                    break
-            if is_only_tensors:
-                return tuple(output_object)
+
+            def flatten_tensor_lists(obj):
+                flattened = []
+                for item in obj:
+                    if isinstance(item, torch.Tensor):
+                        flattened.append(item)
+                    elif isinstance(item, (tuple, list)):
+                        flattened.extend(flatten_tensor_lists(item))
+                    else:
+                        raise NotImplementedError(
+                            f"Item type: ({type(item)}) is not a torch.Tensor or list/tuple of torch.Tensors"
+                        )
+                return flattened
+
+            try:
+                flattened_tensors = flatten_tensor_lists(output_object)
+                return tuple(flattened_tensors)
+            except NotImplementedError as e:
+                raise e
         elif hasattr(output_object, "to_tuple"):
 
             def flatten(t):