5.x merge 4.x

testdata/dnn/download_models.py

@@ -963,13 +963,12 @@ def is_archive(self):
         sha='2b181b55d1d7af718eaca6cabdeb741217b64c73',
         filename='wechat_2021-01/sr.caffemodel'),
     Model(
-        name='yolov7_not_simplified',
+        name='yolov7',
         url=[
-            'https://docs.google.com/uc?export=download&id=1ljSh81ydO5ylsnDoV_mt3zj5RX_cgLu8',
-            'https://dl.opencv.org/models/yolov7/yolov7_not_simplified.onnx'
+            'https://dl.opencv.org/models/yolov7/yolov7.onnx'
         ],
-        sha='fcd0fa401c83bf2b29e18239a9c2c989c9b8669d',
-        filename='onnx/models/yolov7_not_simplified.onnx'),
+        sha='9f5199c266418462771a26a7b8ea25a90412ce2e',
+        filename='onnx/models/yolov7.onnx'),
     Model(
         name='yolox_s_inf_decoder',
         url=[
@@ -988,6 +987,13 @@ def is_archive(self):
         url='https://dl.opencv.org/models/yolov8/yolov8n.onnx',
         sha='136807b88d0b02bc226bdeb9741141d857752e10',
         filename='onnx/models/yolov8n.onnx'),
+
+    Model(
+        name='yolo_nas_s',
+        url='https://dl.opencv.org/models/yolo-nas/yolo_nas_s.onnx',
+        sha='089942fbdf8591875a7a6ff10ac50fb6864e7aa4',
+        filename='onnx/models/yolo_nas_s.onnx'),
+
     Model(
         name='NanoTrackV1 (ONNX, backbone)',
         url='https://raw.githubusercontent.com/zihaomu/opencv_extra_data_backup/main/NanoTrack/models/nanotrack_backbone_sim.onnx',

testdata/dnn/onnx/conformance/node/test_group_normalization_epsilon/test_data_set_0/input_0.pb

@@ -0,0 +1 @@
+
B
xJ�
x��?h��>��z?�j@$�?�.z��8s?b��hdӽ�9�>(�>�%�?^�B?�0�=B�>]ת>�=�?R�iJ�>�Z�/d#��S'?�K]?��=��C@�(��Hm;= �?�2�?��?��>���>�Ec������!��� >*z�?��?�Oƾmǚ��6��&õ�gڿ��?�x�FKྙ[���	G?

testdata/dnn/onnx/conformance/node/test_group_normalization_epsilon/test_data_set_0/input_1.pb

@@ -0,0 +1 @@
+
BscaleJ4��Y�L=e��>

testdata/dnn/onnx/conformance/node/test_group_normalization_epsilon/test_data_set_0/input_2.pb

@@ -0,0 +1 @@
+
BbiasJ �����k��QN�>

testdata/dnn/onnx/conformance/node/test_group_normalization_example/test_data_set_0/input_0.pb

@@ -0,0 +1 @@
+
B
xJ�
x��?h��>��z?�j@$�?�.z��8s?b��hdӽ�9�>(�>�%�?^�B?�0�=B�>]ת>�=�?R�iJ�>�Z�/d#��S'?�K]?��=��C@�(��Hm;= �?�2�?��?��>���>�Ec������!��� >*z�?��?�Oƾmǚ��6��&õ�gڿ��?�x�FKྙ[���	G?

testdata/dnn/onnx/conformance/node/test_group_normalization_example/test_data_set_0/input_1.pb

@@ -0,0 +1 @@
+
BscaleJ4��Y�L=e��>

testdata/dnn/onnx/conformance/node/test_group_normalization_example/test_data_set_0/input_2.pb

@@ -0,0 +1 @@
+
BbiasJ �����k��QN�>

testdata/dnn/onnx/conformance/node/test_group_normalization_example/test_data_set_0/output_0.pb

@@ -0,0 +1,2 @@
+
B
yJ�
�!�,�>>�(���)��N���K�����ew���?4U*=�?��
�:�2?��>�z�>@�>�D.��!о�u����>J�>�8������b��������?��>��?��G?86B?���>|��>�+?��@� Y�A\��}�����������U���z�>�
+7?ڇo?Q�<#�>�W�>��&>T�V?

testdata/dnn/onnx/generate_onnx_models_with_onnxscript.py

@@ -3,6 +3,8 @@
 import onnx
 import onnxscript as ost
 from onnxscript import opset19 as op # opset19 is the lastest by 202309
+from onnxscript import opset11
+from onnxscript import opset13
 
 np.random.seed(0)
 
@@ -69,8 +71,6 @@ def greater_input_dtype_int64(x: ost.FLOAT[27, 9]) ->ost.BOOL[27, 9]:
     return y
 make_model_and_data(greater_input_dtype_int64, np.random.randint(0, 100, size=[27, 9], dtype=np.int64), force_saving_input_as_dtype_float32=True, force_saving_output_as_dtype_float32=True)
 
-from onnxscript import opset11
-
 @ost.script()
 def two_resizes_with_shared_subgraphs(x: ost.FLOAT["batch", 1, "height", "width"], y: ost.FLOAT[1, 1, 3, 2], z: ost.FLOAT[1, 1, 2, 1]) ->ost.FLOAT["batch", 1, "height", "width"]:
     shape_src_1 = opset11.Shape(x)
@@ -310,3 +310,32 @@ def einsum_const_inputs(input_0: ost.FLOAT[3, 2, 2, 4]) -> ost.FLOAT[3, 2, 2, 2]
     return op.Einsum(input_0, input_1, equation="bhwc, hkc -> bhwk")
 
 make_model_and_data(einsum_const_inputs, input_0_data)
+
+''' This subgraph looks the same as LayerNorm expanded, but it has
+    axes=1 in ReduceMean which does not meet the requirement of LayerNorm:
+        - axes[-1] = -1 or the axis of last dimension
+        - adjacent axes, e.g. [1, 2, 3] or [-3, -2, -1]
+'''
+
+n = 1
+c = 4
+h = w = 8
+mul_weight = np.random.rand(c, 1, 1).astype(np.float32)
+add_weight = np.random.rand(c, 1, 1).astype(np.float32)
+
+@ost.script()
+def layer_norm_no_fusion(x: ost.FLOAT[n, c, h, w]) -> ost.FLOAT[n, c, h, w]:
+    reduce_mean = opset13.ReduceMean(x, axes=[1], keepdims=1)
+    sub = opset13.Sub(x, reduce_mean)
+
+    pow = opset13.Pow(sub, opset13.Constant(value=onnx.helper.make_tensor("", onnx.TensorProto.FLOAT, [], np.array([2], dtype=np.float32))))
+    reduce_mean_1 = opset13.ReduceMean(pow, axes=[1], keepdims=1)
+    add = opset13.Add(reduce_mean_1, opset13.Constant(value=onnx.helper.make_tensor("", onnx.TensorProto.FLOAT, [], np.array([9.999999974752427e-7], dtype=np.float32))))
+    sqrt = opset13.Sqrt(add)
+
+    div = opset13.Div(sub, sqrt)
+    mul = opset13.Mul(opset13.Constant(value=onnx.helper.make_tensor("", onnx.TensorProto.FLOAT, [c, 1, 1], mul_weight)), div)
+    add = opset13.Add(mul, opset13.Constant(value=onnx.helper.make_tensor("", onnx.TensorProto.FLOAT, [c, 1, 1], add_weight)))
+
+    return add
+make_model_and_data(layer_norm_no_fusion, np.random.rand(n, c, h, w).astype(np.float32))