onnx 1.13.1__cp39-cp39-win32.whl → 1.14.1__cp39-cp39-win32.whl

onnx/defs/nn/defs.cc

@@ -220,11 +220,7 @@ std::function<void(OpSchema&)> PoolOpSchemaGenerator(
  ```
  output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - {kernelSpatialShape}) / strides_spatial_shape[i] + 1)
  ```
- if ceil_mode is enabled
-
- ```
- * pad_shape[i] is sum of pads along axis i
- ```
+ if ceil_mode is enabled `pad_shape[i]` is the sum of pads along axis `i`.
 
  `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:
  ```
@@ -311,14 +307,19 @@ std::function<void(OpSchema&)> PoolOpSchemaGenerator(
 
 ONNX_OPERATOR_SET_SCHEMA(
     AveragePool,
-    11,
+    19,
     OpSchema()
         .FillUsing(PoolOpSchemaGenerator(
             "AveragePool",
             "average",
             "The output of each pooling window is divided by the number of elements (exclude pad when attribute count_include_pad is zero).",
-            false,
-            false))
+            true, /* use_dilation: dilations attribute has been added in opset 19. */
+            false /* supports8bit: does not support 8bit. */))
+        .Attr(
+            "dilations",
+            "Dilation value along each spatial axis of filter. If not present, the dilation defaults to 1 along each spatial axis.",
+            AttributeProto::INTS,
+            OPTIONAL_VALUE)
         .Attr(
             "count_include_pad",
             "Whether include pad pixels when calculating values for the edges. Default is 0, doesn't count include pad.",
@@ -557,11 +558,7 @@ std::function<void(OpSchema&)> LpPoolOpSchemaGenerator(const char* name) {
  ```
  output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - {kernelSpatialShape}) / strides_spatial_shape[i] + 1)
  ```
- if ceil_mode is enabled
-
- ```
- * pad_shape[i] is sum of pads along axis i
- ```
+ if ceil_mode is enabled `pad_shape[i]` is the sum of pads along axis `i`.
 
  `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:
  ```
@@ -1371,6 +1368,104 @@ output_shape can also be explicitly specified in which case pads values are auto
 
 ONNX_OPERATOR_SET_SCHEMA(ConvTranspose, 11, OpSchema().FillUsing(ConvTransposeOpSchemaGenerator("a filter")));
 
+static const char* DeformConv_ver19_doc = R"DOC(
+Performs deformable convolution as described in https://arxiv.org/abs/1703.06211 and https://arxiv.org/abs/1811.11168.
+This operator specification supports the general N-D case. Note that most common use cases have 2D or 3D data.
+)DOC";
+
+ONNX_OPERATOR_SET_SCHEMA(
+    DeformConv,
+    19,
+    OpSchema()
+        .SetDoc(DeformConv_ver19_doc)
+        .Input(
+            0,
+            "X",
+            "Input data tensor. For 2D image data, it has shape (N, C, H, W) where N is the batch size, "
+            "C is the number of input channels, and H and W are the height and width. "
+            "In general, the shape is (N, C, D1, D2, ... , Dn) for n-dimensional data, where "
+            "D1 to Dn are the spatial dimension sizes. Most common use cases have n = 2 or 3.",
+            "T")
+        .Input(
+            1,
+            "W",
+            "Weight tensor that will be used in the convolutions. It has shape (oC, C/group, kH, kW), "
+            "where oC is the number of output channels and kH and kW are the kernel height and width. "
+            "For more than 2 dimensions, it has shape (oC, C/group, k1, k2, ... , kn).",
+            "T")
+        .Input(
+            2,
+            "offset",
+            "Offset tensor denoting the offset for the sampling locations in the convolution kernel. "
+            "It has shape (N, offset_group * kH * kW * 2, oH, oW) for 2D data or "
+            "(N, offset_group * k1 * k2 * ... * kn * n, o1, o2, ... , on) for nD data. Use linear interpolation"
+            "for fractional offset values. Sampling locations outside of the padded input tensor gives zero.",
+            "T")
+        .Input(
+            3,
+            "B",
+            "Optional 1D bias of length oC to be added to the convolution. Default is a tensor of zeros.",
+            "T",
+            OpSchema::Optional)
+        .Input(
+            4,
+            "mask",
+            "The mask tensor to be applied to each position in the convolution kernel. "
+            "It has shape (N, offset_group * kH * kW, oH, oW) for 2D data or "
+            "(N, offset_group * k1 * k2 * ... * kn * n, o1, o2, ... , on) for nD data. Default is a "
+            "tensor of ones.",
+            "T",
+            OpSchema::Optional)
+        .Output(
+            0,
+            "Y",
+            "Output data tensor that contains the result of convolution. It has shape (N, oC, oH, oW) "
+            "for 2D data or (N, oC, o1, o2, ..., on) for nD data",
+            "T")
+        .TypeConstraint(
+            "T",
+            {"tensor(float16)", "tensor(float)", "tensor(double)"},
+            "Constrain input and output types to float tensors.")
+        .Attr(
+            "dilations",
+            "Dilation value along each spatial axis of the kernel. Default is 1 along each axis.",
+            AttributeProto::INTS,
+            OPTIONAL_VALUE)
+        .Attr(
+            "group",
+            "Number of groups the input and output channels, C and oC, are divided into. C and oC must both "
+            "be divisible by group. Default is 1.",
+            AttributeProto::INT,
+            static_cast<int64_t>(1))
+        .Attr(
+            "kernel_shape",
+            "Shape of the convolution kernel. If not present, it is inferred from the shape of input W.",
+            AttributeProto::INTS,
+            OPTIONAL_VALUE)
+        .Attr(
+            "offset_group",
+            "Number of groups of offset. C must be divisible by offset_group. Default is 1.",
+            AttributeProto::INT,
+            static_cast<int64_t>(1))
+        .Attr(
+            "pads",
+            "Padding for the beginning and end along each spatial axis. The values represent the number of pixels "
+            "added to the beginning and end of the corresponding axis and can take any nonnegative value. "
+            "The format should be as follows: [x1_begin, x2_begin, ..., x1_end, x2_end, ...], where xi_begin "
+            "is the number of pixels added at the beginning of axis `i` and xi_end is the number of pixels "
+            "added at the end of axis `i`. Default is 0 along each axis.",
+            AttributeProto::INTS,
+            OPTIONAL_VALUE)
+        .Attr(
+            "strides",
+            "Stride along each spatial axis. Default is 1 along each axis.",
+            AttributeProto::INTS,
+            OPTIONAL_VALUE)
+        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
+          propagateElemTypeFromInputToOutput(ctx, 0, 0);
+          convPoolShapeInference(ctx, true, false, 0, 1);
+        }));
+
 // For GlobalPool operations.
 void globalPoolTypeShapeInference(InferenceContext& ctx) {
   propagateElemTypeFromInputToOutput(ctx, 0, 0);
@@ -1507,8 +1602,8 @@ statistics in inference mode (training_mode=False, default),
 and the running statistics in training mode (training_mode=True).
 There are multiple cases for the number of outputs, which we list below:
 
-Output case #1: Y, running_mean, running_var (training_mode=True)
-Output case #2: Y (training_mode=False)
+* Output case #1: Y, running_mean, running_var (training_mode=True)
+* Output case #2: Y (training_mode=False)
 
 When training_mode=False, extra outputs are invalid.
 The outputs are updated as follows when training_mode=True:
@@ -1517,17 +1612,15 @@ running_mean = input_mean * momentum + current_mean * (1 - momentum)
 running_var = input_var * momentum + current_var * (1 - momentum)
 
 Y = (X - current_mean) / sqrt(current_var + epsilon) * scale + B
-
+```
 where:
-
+```
 current_mean = ReduceMean(X, axis=all_except_channel_index)
 current_var =  ReduceVar(X, axis=all_except_channel_index)
-
-Notice that ReduceVar refers to the population variance, and it equals to
-sum(sqrd(x_i - x_avg)) / N
-where N is the population size (this formula does not use sample size N - 1).
-
 ```
+Notice that `ReduceVar` refers to the population variance, and it equals to
+`sum(sqrd(x_i - x_avg)) / N`
+where `N` is the population size (this formula does not use sample size `N - 1`).
 
 The computation of ReduceMean and ReduceVar uses float to avoid overflow for float16 inputs.
 
@@ -1919,10 +2012,7 @@ ONNX_OPERATOR_SET_SCHEMA(
             true,
             1,
             OpSchema::Differentiable)
-        .TypeConstraint(
-            "T",
-            OpSchema::all_tensor_types_with_bfloat(),
-            "Constrain input and output to all tensor types.")
+        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output to all tensor types.")
         .Attr(
             "axis",
             "Indicate up to which input dimensions "
@@ -1953,14 +2043,14 @@ ONNX_OPERATOR_SET_SCHEMA(
 static const char* LRN_ver13_doc = R"DOC(
 Local Response Normalization proposed in the [AlexNet paper](https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf).
 It normalizes over local input regions.
-The local region is defined across the channels. For an element X[n, c, d1, ..., dk] in a tensor
-of shape (N x C x D1 x D2, ..., Dk), its region is
-{X[n, i, d1, ..., dk] | max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2))}.
+The local region is defined across the channels. For an element `X[n, c, d1, ..., dk]` in a tensor
+of shape `(N x C x D1 x D2, ..., Dk)`, its region is
+`{X[n, i, d1, ..., dk] | max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2))}`.
 
-square_sum[n, c, d1, ..., dk] = sum(X[n, i, d1, ..., dk] ^ 2),
-where max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2)).
+`square_sum[n, c, d1, ..., dk] = sum(X[n, i, d1, ..., dk] ^ 2)`,
+where `max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2))`.
 
-Y[n, c, d1, ..., dk] = X[n, c, d1, ..., dk] / (bias + alpha / size * square_sum[n, c, d1, ..., dk] ) ^ beta
+`Y[n, c, d1, ..., dk] = X[n, c, d1, ..., dk] / (bias + alpha / size * square_sum[n, c, d1, ..., dk] ) ^ beta`
 )DOC";
 
 ONNX_OPERATOR_SET_SCHEMA(
@@ -2204,7 +2294,7 @@ ONNX_OPERATOR_SET_SCHEMA(
 
 static const char* mvn_ver13_doc = R"DOC(
       A MeanVarianceNormalization Function: Perform mean variance normalization
-      on the input tensor X using formula: <br/> ``` (X-EX)/sqrt(E(X-EX)^2) ```
+      on the input tensor X using formula: `(X-EX)/sqrt(E(X-EX)^2)`
 )DOC";
 
 static const std::vector<int64_t> mvn_default_axes = {0, 2, 3};
@@ -2361,10 +2451,10 @@ Col2Im behaves similarly to PyTorch's fold https://pytorch.org/docs/stable/gener
 but it only supports *batched* multi-dimensional image tensors.
 Another implementation in Python with N-dimension support can be found at https://github.com/f-dangel/unfoldNd/.
 
-NOTE: Although specifying image_shape looks redundant because it could be calculated from
-      convolution formulas, it is required as input for more advanced scenarios as explained
-      at PyTorch's implementation (https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/Col2Im.cpp#L10)
-
+NOTE:
+  Although specifying image_shape looks redundant because it could be calculated from
+  convolution formulas, it is required as input for more advanced scenarios as explained
+  at PyTorch's implementation (https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/Col2Im.cpp#L10)
 )DOC";
 
 ONNX_OPERATOR_SET_SCHEMA(
@@ -2443,7 +2533,7 @@ ONNX_OPERATOR_SET_SCHEMA(
             OpSchema::Differentiable)
         .TypeConstraint(
             "T",
-            OpSchema::all_tensor_types_with_bfloat(),
+            OpSchema::all_tensor_types_ir4(),
             "Constrain input and output types to all numeric tensor types.")
         .TypeAndShapeInferenceFunction([](InferenceContext& ctx) { col2imShapeInference(ctx); }));
 
@@ -2676,20 +2766,20 @@ ONNX_OPERATOR_SET_SCHEMA(
         }));
 
 static const char* GroupNormalization_ver18_doc = R"DOC(
-A GroupNormalization function. Carries out group normalization as described in 
-the paper https://arxiv.org/abs/1803.08494 
+A GroupNormalization function. Carries out group normalization as described in
+the paper https://arxiv.org/abs/1803.08494
 
 This operator transforms input according to
 ```
 y = scale * (x - mean) / sqrt(variance + epsilon) + bias,
 ```
-where the mean and variance are computed per instance per group of channels, and 
-`scale` and `bias` should be specified for each group of channels. The number of 
-groups `num_groups` should be divisible by the number of channels so that there are 
+where the mean and variance are computed per instance per group of channels, and
+`scale` and `bias` should be specified for each group of channels. The number of
+groups `num_groups` should be divisible by the number of channels so that there are
 an equal number of channels per group.
 
-When the number of groups is the same as the number of channels, this operator is 
-equivalent to InstanceNormalization. When there is only one group, this operator 
+When the number of groups is the same as the number of channels, this operator is
+equivalent to InstanceNormalization. When there is only one group, this operator
 is equivalent to LayerNormalization.
 )DOC";
 

onnx/defs/nn/old.cc

@@ -236,6 +236,14 @@ const char* auto_pad_doc2 =
     "SAME_UPPER or SAME_LOWER mean pad the input so that the output spatial size match the input."
     "In case of odd number add the extra padding at the end for SAME_UPPER and at the "
     "beginning for SAME_LOWER. VALID mean no padding.";
+const char* auto_pad_doc3 =
+    "auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where "
+    "default value is NOTSET, which means explicit padding is used. "
+    "SAME_UPPER or SAME_LOWER mean pad the input so that "
+    "`output_shape[i] = ceil(input_shape[i] / strides[i])` for each axis `i`. "
+    "The padding is split between the two sides equally or almost equally (depending "
+    "on whether it is even or odd). In case the padding is an odd number, the extra "
+    "padding is added at the end for SAME_UPPER and at the beginning for SAME_LOWER.";
 
 void convPoolShapeInference1(
     InferenceContext& ctx,
@@ -575,6 +583,124 @@ std::function<void(OpSchema&)> PoolOpSchemaGenerator_10(
   };
 }
 
+std::vector<std::string> GetSupportedDataTypesForPoolingOps_1(bool supports8bit) {
+  if (supports8bit) {
+    return {"tensor(float16)", "tensor(float)", "tensor(double)", "tensor(int8)", "tensor(uint8)"};
+  }
+  return {"tensor(float16)", "tensor(float)", "tensor(double)"};
+}
+
+std::function<void(OpSchema&)> PoolOpSchemaGenerator_11(
+    const char* name,
+    const char* opName,
+    const char* additionalDescription,
+    bool use_dilation,
+    bool supports8bit = false) {
+  return [=](OpSchema& schema) {
+    std::string doc;
+    POPULATE_OP_DOC_STR(
+        doc = R"DOC(
+ {name} consumes an input tensor X and applies {opName} pooling across
+ the tensor according to kernel sizes, stride sizes, and pad lengths.
+ {opName} pooling consisting of computing the {opName} on all values of a
+ subset of the input tensor according to the kernel size and downsampling the
+ data into the output tensor Y for further processing. The output spatial shape will be following:
+ ```
+ output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - {kernelSpatialShape}) / strides_spatial_shape[i] + 1)
+ ```
+ or
+ ```
+ output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - {kernelSpatialShape}) / strides_spatial_shape[i] + 1)
+ ```
+ if ceil_mode is enabled
+
+ ```
+ * pad_shape[i] is sum of pads along axis i
+ ```
+
+ `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:
+ ```
+ VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - {kernelSpatialShape} + 1) / strides_spatial_shape[i])
+ SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])
+ ```
+ And pad shape will be following if `SAME_UPPER` or `SAME_LOWER`:
+ ```
+ pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + {kernelSpatialShape} - input_spatial_shape[i]
+ ```
+ {additionalDescription}
+ )DOC";
+        ReplaceAll(doc, "{name}", name);
+        ReplaceAll(doc, "{opName}", opName);
+        ReplaceAll(doc, "{additionalDescription}", additionalDescription);
+        ReplaceAll(
+            doc,
+            "{kernelSpatialShape}",
+            use_dilation ? "((kernel_spatial_shape[i] - 1) * dilations[i] + 1)" : "kernel_spatial_shape[i]"););
+    schema.SetDoc(doc);
+    schema.Attr("kernel_shape", "The size of the kernel along each axis.", AttributeProto::INTS);
+    schema.Attr(
+        "strides",
+        "Stride along each spatial axis. If not present, the stride defaults to 1 along each spatial axis.",
+        AttributeProto::INTS,
+        OPTIONAL_VALUE);
+    schema.Attr("auto_pad", auto_pad_doc3, AttributeProto::STRING, std::string("NOTSET"));
+    schema.Attr("pads", pads_doc2, AttributeProto::INTS, OPTIONAL_VALUE);
+    schema.Attr(
+        "ceil_mode",
+        "Whether to use ceil or floor (default) to compute the output shape.",
+        AttributeProto::INT,
+        static_cast<int64_t>(0));
+    schema.Input(
+        0,
+        "X",
+        "Input data tensor from the previous operator; "
+        "dimensions for image case 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 the "
+        "width of the data. For non image case, the "
+        "dimensions are in the form of "
+        "(N x C x D1 x D2 ... Dn), where N is the batch "
+        "size. Optionally, if dimension denotation is "
+        "in effect, the operation expects the input "
+        "data tensor to arrive with the dimension denotation "
+        "of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].",
+        "T",
+        OpSchema::Single,
+        true,
+        1,
+        OpSchema::Differentiable);
+    schema.Output(
+        0,
+        "Y",
+        "Output data tensor from average or max pooling across "
+        "the input tensor. Dimensions will vary based "
+        "on various kernel, stride, and pad sizes. Floor value of "
+        "the dimension is used",
+        "T",
+        OpSchema::Single,
+        true,
+        1,
+        OpSchema::Differentiable);
+    schema.TypeConstraint(
+        "T",
+        GetSupportedDataTypesForPoolingOps_1(supports8bit),
+        supports8bit ? "Constrain input and output types to float and 8 bit tensors."
+                     : "Constrain input and output types to float tensors.");
+    schema.TypeAndShapeInferenceFunction([use_dilation](InferenceContext& ctx) {
+      propagateElemTypeFromInputToOutput(ctx, 0, 0);
+      if (ctx.getNumOutputs() > 1) {
+        // MaxPool with two outputs case.
+        auto output_type = ctx.getOutputType(1);
+        if (output_type->value_case() == TypeProto::kTensorType ||
+            output_type->value_case() == TypeProto::VALUE_NOT_SET) {
+          output_type->mutable_tensor_type()->set_elem_type(TensorProto::INT64);
+        }
+      }
+      convPoolShapeInference1(ctx, use_dilation, true, 0, 1);
+    });
+  };
+}
+
 ONNX_OPERATOR_SET_SCHEMA(
     AveragePool,
     1,
@@ -613,6 +739,22 @@ ONNX_OPERATOR_SET_SCHEMA(
             AttributeProto::INT,
             static_cast<int64_t>(0)));
 
+ONNX_OPERATOR_SET_SCHEMA(
+    AveragePool,
+    11,
+    OpSchema()
+        .FillUsing(PoolOpSchemaGenerator_11(
+            "AveragePool",
+            "average",
+            "The output of each pooling window is divided by the number of elements (exclude pad when attribute count_include_pad is zero).",
+            true,
+            false))
+        .Attr(
+            "count_include_pad",
+            "Whether include pad pixels when calculating values for the edges. Default is 0, doesn't count include pad.",
+            AttributeProto::INT,
+            static_cast<int64_t>(0)));
+
 ONNX_OPERATOR_SET_SCHEMA(
     MaxPool,
     1,
@@ -980,15 +1122,6 @@ static const char* GlobalLpPool_ver1_doc = R"DOC(
  the values in the same channel. This is equivalent to LpPool with kernel size
  equal to the spatial dimension of input tensor.)DOC";
 
-const char* auto_pad_doc3 =
-    "auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where "
-    "default value is NOTSET, which means explicit padding is used. "
-    "SAME_UPPER or SAME_LOWER mean pad the input so that "
-    "`output_shape[i] = ceil(input_shape[i] / strides[i])` for each axis `i`. "
-    "The padding is split between the two sides equally or almost equally (depending "
-    "on whether it is even or odd). In case the padding is an odd number, the extra "
-    "padding is added at the end for SAME_UPPER and at the beginning for SAME_LOWER.";
-
 std::function<void(OpSchema&)> LpPoolOpSchemaGenerator_11(const char* name) {
   return [=](OpSchema& schema) {
     std::string doc;

onnx/defs/operator_sets.h

@@ -1058,6 +1058,49 @@ class OpSet_Onnx_ver18 {
   }
 };
 
+// Forward declarations for ai.onnx version 19
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Equal);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, AveragePool);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Cast);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, CastLike);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Constant);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, DeformConv);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, DequantizeLinear);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Identity);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, If);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Loop);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Pad);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, QuantizeLinear);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Resize);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Reshape);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Scan);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Shape);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Size);
+
+// Iterate over schema from ai.onnx version 19
+class OpSet_Onnx_ver19 {
+ public:
+  static void ForEachSchema(std::function<void(OpSchema&&)> fn) {
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Equal)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, AveragePool)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Cast)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, CastLike)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Constant)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, DeformConv)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, DequantizeLinear)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Identity)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, If)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Loop)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Pad)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, QuantizeLinear)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Reshape)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Resize)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Scan)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Shape)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 19, Size)>());
+  }
+};
+
 inline void RegisterOnnxOperatorSetSchema() {
   RegisterOpSetSchema<OpSet_Onnx_ver1>();
   RegisterOpSetSchema<OpSet_Onnx_ver2>();
@@ -1077,6 +1120,7 @@ inline void RegisterOnnxOperatorSetSchema() {
   RegisterOpSetSchema<OpSet_Onnx_ver16>();
   RegisterOpSetSchema<OpSet_Onnx_ver17>();
   RegisterOpSetSchema<OpSet_Onnx_ver18>();
+  RegisterOpSetSchema<OpSet_Onnx_ver19>();
   // 0 means all versions of ONNX schema have been loaded
   OpSchemaRegistry::Instance()->SetLoadedSchemaVersion(0);
 }
@@ -1084,6 +1128,7 @@ inline void RegisterOnnxOperatorSetSchema() {
 inline void RegisterOnnxOperatorSetSchema(int target_version) {
   // Update here if opset_version bumps
   // These calls for schema registration here are required to be in descending order for this to work correctly
+  RegisterOpSetSchema<OpSet_Onnx_ver19>(target_version);
   RegisterOpSetSchema<OpSet_Onnx_ver18>(target_version);
   RegisterOpSetSchema<OpSet_Onnx_ver17>(target_version);
   RegisterOpSetSchema<OpSet_Onnx_ver16>(target_version);

onnx/defs/optional/defs.cc

@@ -141,10 +141,14 @@ ONNX_OPERATOR_SET_SCHEMA(
           if (input_type == nullptr) {
             fail_type_inference("Input type is null. Input must have Type information.");
           }
-          if (input_type->has_optional_type() && !input_type->optional_type().has_elem_type()) {
-            fail_type_inference("Optional-type input must contain an element with type information.");
+          if (input_type->has_optional_type()) {
+            if (!input_type->optional_type().has_elem_type()) {
+              fail_type_inference("Optional-type input must contain an element with type information.");
+            }
+            ctx.getOutputType(0)->CopyFrom(input_type->optional_type().elem_type());
+          } else {
+            propagateShapeAndTypeFromFirstInput(ctx);
           }
-          ctx.getOutputType(0)->CopyFrom(input_type->optional_type().elem_type());
         }));
 
-} // namespace ONNX_NAMESPACE
+} // namespace ONNX_NAMESPACE

onnx/defs/parser.cc

@@ -106,6 +106,32 @@ Status OnnxParser::Parse(char open, IdList& idlist, char close) {
   return Status::OK();
 }
 
+Status OnnxParser::Parse(IdList& idlist, AttrList& attrlist) {
+  idlist.Clear();
+  attrlist.Clear();
+  do {
+    std::string id;
+    ParseIdentifier(id);
+    auto next = NextChar();
+    if (next == ':' || next == '=')
+      Parse(*attrlist.Add(), id);
+    else
+      *idlist.Add() = id;
+  } while (Matches(','));
+  return Status::OK();
+}
+
+Status OnnxParser::Parse(char open, IdList& idlist, AttrList& attrlist, char close) {
+  if (Matches(open)) {
+    PARSE(idlist, attrlist);
+    MATCH(close);
+  } else {
+    idlist.Clear();
+    attrlist.Clear();
+  }
+  return Status::OK();
+}
+
 Status OnnxParser::Parse(TensorShapeProto& shape) {
   shape.clear_dim();
   do {
@@ -366,6 +392,12 @@ Status OnnxParser::Parse(TensorProto& tensorProto, const TypeProto& tensorTypePr
   return Status::OK();
 }
 
+bool OnnxParser::NextIsIdentifier() {
+  std::string id("");
+  (void)PeekIdentifier(id);
+  return !(id.empty());
+}
+
 bool OnnxParser::NextIsType() {
   std::string id("");
   (void)PeekIdentifier(id);
@@ -387,8 +419,19 @@ Status OnnxParser::ParseSingleAttributeValue(AttributeProto& attr) {
   auto next = NextChar();
   if (isalpha(next) || next == '_') {
     if (NextIsType()) {
-      attr.set_type(AttributeProto_AttributeType_TENSOR);
-      Parse(*attr.mutable_t());
+      TypeProto typeProto;
+      Parse(typeProto);
+      next = NextChar();
+      if ((next == '{') || (next == '=') || (NextIsIdentifier())) {
+        attr.set_type(AttributeProto_AttributeType_TENSOR);
+        auto& tensorProto = *attr.mutable_t();
+        ParseOptionalIdentifier(*tensorProto.mutable_name());
+        (void)Matches('='); // Optional, to unify handling of initializers
+        Parse(tensorProto, typeProto);
+      } else {
+        attr.set_type(AttributeProto_AttributeType_TYPE_PROTO);
+        attr.mutable_tp()->CopyFrom(typeProto);
+      }
     } else {
       attr.set_type(AttributeProto_AttributeType_GRAPH);
       Parse(*attr.mutable_g());
@@ -424,6 +467,10 @@ Status OnnxParser::Parse(AttributeProto& attr) {
   attr.Clear();
   std::string name;
   CHECK_PARSER_STATUS(ParseIdentifier(name));
+  return Parse(attr, name);
+}
+
+Status OnnxParser::Parse(AttributeProto& attr, std::string& name) {
   attr.set_name(name);
   if (Matches(':')) {
     CHECK_PARSER_STATUS(ParseIdentifier(name));
@@ -559,7 +606,7 @@ Status OnnxParser::Parse(FunctionProto& fn) {
   ParseIdentifier(id);
   fn.set_name(id);
 
-  PARSE('<', *fn.mutable_attribute(), '>');
+  PARSE('<', *fn.mutable_attribute(), *fn.mutable_attribute_proto(), '>');
   PARSE('(', *fn.mutable_input(), ')');
   MATCH('=');
   MATCH('>', false);