Update onnx-metadata.js

source/onnx-metadata.json

@@ -14164,9 +14164,91 @@
       }
     ],
     "examples": [
+      {
+        "summary": "epsilon",
+        "code": "c = 4\nnum_groups = 2\nx = np.random.randn(3, c, 2, 2).astype(np.float32)\nscale = np.random.randn(c).astype(np.float32)\nbias = np.random.randn(c).astype(np.float32)\nepsilon = 1e-2\ny = _group_normalization(x, num_groups, scale, bias, epsilon).astype(np.float32)\n\nnode = onnx.helper.make_node(\n    \"GroupNormalization\",\n    inputs=[\"x\", \"scale\", \"bias\"],\n    outputs=[\"y\"],\n    epsilon=epsilon,\n    num_groups=num_groups,\n)\n\nexpect(\n    node,\n    inputs=[x, scale, bias],\n    outputs=[y],\n    name=\"test_group_normalization_epsilon\",\n)"
+      },
+      {
+        "summary": "groupnormalization",
+        "code": "c = 4\nnum_groups = 2\nx = np.random.randn(3, c, 2, 2).astype(np.float32)\nscale = np.random.randn(c).astype(np.float32)\nbias = np.random.randn(c).astype(np.float32)\ny = _group_normalization(x, num_groups, scale, bias).astype(np.float32)\n\nnode = onnx.helper.make_node(\n    \"GroupNormalization\",\n    inputs=[\"x\", \"scale\", \"bias\"],\n    outputs=[\"y\"],\n    num_groups=num_groups,\n)\n\nexpect(\n    node,\n    inputs=[x, scale, bias],\n    outputs=[y],\n    name=\"test_group_normalization_example\",\n)"
+      }
+    ]
+  },
+  {
+    "name": "GroupNormalization",
+    "module": "ai.onnx",
+    "version": 21,
+    "description": "A GroupNormalization function. Carries out group normalization as described in\nthe paper https://arxiv.org/abs/1803.08494\n\nThis operator transforms input according to\n```\ny = scale * (x - mean) / sqrt(variance + epsilon) + bias,\n```\nwhere the mean and variance are computed per instance per group of channels, and\n`scale` and `bias` should be specified for each group of channels. The number of\ngroups `num_groups` should be divisible by the number of channels so that there are\nan equal number of channels per group.\n\nThe overall computation has two stages: the first stage normalizes the elements to\nhave zero mean and unit variance for each instance in each group, and the second\nstage scales and shifts the results of the first stage. The floating-point precision\nused in the first stage is determined by the `stash_type` attribute. For example,\nif `stash_type` is 1, the operator casts all input variables to 32-bit float,\nperforms the computation, and finally casts the normalized results back to the\noriginal type of `X`. The second stage does not depend on `stash_type`.\n\nWhen the number of groups is the same as the number of channels, this operator is\nequivalent to InstanceNormalization. When there is only one group, this operator\nis equivalent to LayerNormalization.\n",
+    "attributes": [
+      {
+        "name": "epsilon",
+        "type": "float32",
+        "required": false,
+        "default": 9.999999747378752e-06,
+        "description": "The epsilon value to use to avoid division by zero."
+      },
+      {
+        "name": "num_groups",
+        "type": "int64",
+        "required": true,
+        "description": "The number of groups of channels. It should be a divisor of the number of channels `C`."
+      },
+      {
+        "name": "stash_type",
+        "type": "int64",
+        "required": false,
+        "default": 1,
+        "description": "The floating-point precision used in stage one of the computation."
+      }
+    ],
+    "inputs": [
+      {
+        "name": "X",
+        "type": "T",
+        "description": "Input data tensor. Dimensions for image cases are `(N x C x H x W)`, where `N` is the batch size, `C` is the number of channels, and `H` and `W` are the height and width of the data. Statistics are computed for every group of channels over `C`, `H`, and `W`. For non-image cases, the dimensions are in the form of `(N x C x D1 x D2 ... Dn)`."
+      },
+      {
+        "name": "scale",
+        "type": "T",
+        "description": "Scale tensor of shape `(C)`."
+      },
+      {
+        "name": "bias",
+        "type": "T",
+        "description": "Bias tensor of shape `(C)`."
+      }
+    ],
+    "min_input": 3,
+    "max_input": 3,
+    "outputs": [
+      {
+        "name": "Y",
+        "type": "T",
+        "description": "The output tensor of the same shape as `X`."
+      }
+    ],
+    "min_output": 1,
+    "max_output": 1,
+    "type_constraints": [
+      {
+        "description": "Constrain input and output types to float tensors.",
+        "type_param_str": "T",
+        "allowed_type_strs": [
+          "tensor(bfloat16)",
+          "tensor(float16)",
+          "tensor(float)",
+          "tensor(double)"
+        ]
+      }
+    ],
+    "examples": [
+      {
+        "summary": "epsilon",
+        "code": "c = 4\nnum_groups = 2\nx = np.random.randn(3, c, 2, 2).astype(np.float32)\nscale = np.random.randn(c).astype(np.float32)\nbias = np.random.randn(c).astype(np.float32)\nepsilon = 1e-2\ny = _group_normalization(x, num_groups, scale, bias, epsilon).astype(np.float32)\n\nnode = onnx.helper.make_node(\n    \"GroupNormalization\",\n    inputs=[\"x\", \"scale\", \"bias\"],\n    outputs=[\"y\"],\n    epsilon=epsilon,\n    num_groups=num_groups,\n)\n\nexpect(\n    node,\n    inputs=[x, scale, bias],\n    outputs=[y],\n    name=\"test_group_normalization_epsilon\",\n)"
+      },
       {
         "summary": "groupnormalization",
-        "code": "x = np.random.randn(3, 4, 2, 2).astype(np.float32)\nnum_groups = 2\nscale = np.random.randn(num_groups).astype(np.float32)\nbias = np.random.randn(num_groups).astype(np.float32)\ny = _group_normalization(x, num_groups, scale, bias).astype(np.float32)\n\nnode = onnx.helper.make_node(\n    \"GroupNormalization\",\n    inputs=[\"x\", \"scale\", \"bias\"],\n    outputs=[\"y\"],\n    num_groups=num_groups,\n)\n\nexpect(\n    node,\n    inputs=[x, scale, bias],\n    outputs=[y],\n    name=\"test_group_normalization_example\",\n)\n\nx = np.random.randn(3, 4, 2, 2).astype(np.float32)\nnum_groups = 2\nscale = np.random.randn(num_groups).astype(np.float32)\nbias = np.random.randn(num_groups).astype(np.float32)\nepsilon = 1e-2\ny = _group_normalization(x, num_groups, scale, bias, epsilon).astype(np.float32)\n\nnode = onnx.helper.make_node(\n    \"GroupNormalization\",\n    inputs=[\"x\", \"scale\", \"bias\"],\n    outputs=[\"y\"],\n    epsilon=epsilon,\n    num_groups=num_groups,\n)\n\nexpect(\n    node,\n    inputs=[x, scale, bias],\n    outputs=[y],\n    name=\"test_group_normalization_epsilon\",\n)"
+        "code": "c = 4\nnum_groups = 2\nx = np.random.randn(3, c, 2, 2).astype(np.float32)\nscale = np.random.randn(c).astype(np.float32)\nbias = np.random.randn(c).astype(np.float32)\ny = _group_normalization(x, num_groups, scale, bias).astype(np.float32)\n\nnode = onnx.helper.make_node(\n    \"GroupNormalization\",\n    inputs=[\"x\", \"scale\", \"bias\"],\n    outputs=[\"y\"],\n    num_groups=num_groups,\n)\n\nexpect(\n    node,\n    inputs=[x, scale, bias],\n    outputs=[y],\n    name=\"test_group_normalization_example\",\n)"
       }
     ]
   },