Merge branch 4.x

Author: asmorkalov

testdata/dnn/download_models.py

@@ -1034,6 +1034,11 @@ def handle_bad_download(self):
         url='https://github.com/opencv/opencv_zoo/raw/aab69020085e9b6390723b61f9789ec56b96b07e/models/text_recognition_crnn/text_recognition_CRNN_EN_2021sep.onnx',
         sha='dc8c70a52c6880f11859bf074bcd294a45860821',
         filename='text_recognition_CRNN_EN_2021sep.onnx'),
+    Model(
+        name='RAFT', # See https://github.com/opencv/opencv_zoo/tree/main/models/optical_flow_estimation_raft#raft for source
+        url='https://github.com/opencv/opencv_zoo/raw/281d232cd99cd920853106d853c440edd35eb442/models/optical_flow_estimation_raft/optical_flow_estimation_raft_2023aug.onnx',
+        sha='8165e43c7bd37cc651f66509532acdb3c292832b',
+        filename='onnx/models/optical_flow_estimation_raft_2023aug.onnx'),
 ]
 
 # Note: models will be downloaded to current working directory

testdata/dnn/onnx/data/input_cumsum_exclusive_inplace.npy

@@ -1457,6 +1457,15 @@ def forward(self, x):
 
 save_data_and_model_multy_inputs("einsum_2d", einsum, mat1, mat2, export_params=True)
 
+# 2d test case with ellipses
+mat1 = torch.randn(4, 5)
+mat2 = torch.randn(5, 8)
+equation  = "...ij, ...jk -> ...ik"
+einsum = Einsum(equation)
+output = einsum(mat1, mat2)
+
+save_data_and_model_multy_inputs("einsum_2d_ellipses", einsum, mat1, mat2, export_params=True)
+
 # 3d test case
 mat1 = torch.ones(2, 4, 5)
 mat2 = torch.ones(2, 5, 8)
@@ -2519,6 +2528,25 @@ def forward(self, x):
 x = torch.randn(2, 3, 4)
 save_data_and_model("cumsum_3d_dim_2", x, CumSum(dim=2), version=11)
 
+# test: CumSum exclusive layer should not be executed inplace
+dims = h.make_node("Constant", inputs=[], outputs=["dims1"], name="node-c1",
+                   value=h.make_tensor(name="c1v", data_type=onnx.TensorProto.INT64, dims=[], vals=np.asarray([1, ], dtype=np.int64)))
+one = h.make_node("Constant", inputs=[], outputs=["one1"], name="node-c2",
+                   value=h.make_tensor(name="c2v", data_type=onnx.TensorProto.FLOAT, dims=[], vals=np.asarray([1, ], dtype=np.float32)))
+
+mult = h.make_node("Mul", inputs=["input1", "one1"], outputs=["mul_output1"], name="node-m1")
+cumsum = h.make_node("CumSum", inputs=["mul_output1", "dims1"], outputs=["cumsum_output1"], name="node-r1", exclusive=1)
+
+graph = h.make_graph([dims, one, mult, cumsum], "graph123",
+                     [h.make_tensor_value_info("input1", onnx.TensorProto.FLOAT, [1, 3, 1, 1]),],
+                     [h.make_tensor_value_info("cumsum_output1", onnx.TensorProto.FLOAT, [1, 3, 1, 1])])
+cumsum_model = h.make_model(graph, producer_name="model_cumsum")
+onnx.checker.check_model(cumsum_model)
+
+input_np = np.array([1, 2, 3], dtype=np.float32).reshape(1, 3, 1, 1)
+output_np = np.array([0, 1, 3], dtype=np.float32).reshape(1, 3, 1, 1)
+save_data_and_onnx_model("cumsum_exclusive_inplace", input_np, output_np, cumsum_model)
+
 # where layer
 class Where(nn.Module):
     def __init__(self, *args, **kwargs):

testdata/dnn/onnx/data/input_einsum_2d_ellipses_0.npy

@@ -0,0 +1,17 @@
+pytorch2.0.0:�

+D
+one
+two
2/Einsum"Einsum*$
+equation"...ij, ...jk -> ...ik�
	torch_jitZ
+one
+

+
+Z
+two
+

+
+b-
+
2(
+&
"
+Einsum2_dim_0
+Einsum2_dim_1B