onnx 1.16.1__cp38-cp38-win_amd64.whl → 1.17.0__cp38-cp38-win_amd64.whl

onnx/defs/schema.h

@@ -15,6 +15,7 @@
 #include <ostream>
 #include <set>
 #include <string>
+#include <string_view>
 #include <tuple>
 #include <unordered_map>
 #include <unordered_set>
@@ -763,12 +764,36 @@ class OpSchema final {
     return all_tensor_types_ir4;
   }
 
+  static const std::vector<std::string>& all_non_complex_numeric_types_plus_bool_ir4() {
+    static const std::vector<std::string> all_non_complex_numeric_types_plus_bool_ir4 = {
+        "tensor(uint8)",
+        "tensor(uint16)",
+        "tensor(uint32)",
+        "tensor(uint64)",
+        "tensor(int8)",
+        "tensor(int16)",
+        "tensor(int32)",
+        "tensor(int64)",
+        "tensor(bfloat16)",
+        "tensor(float16)",
+        "tensor(float)",
+        "tensor(double)",
+        "tensor(bool)"};
+    return all_non_complex_numeric_types_plus_bool_ir4;
+  }
+
   static const std::vector<std::string>& all_float_types_ir4() {
     static const std::vector<std::string> all_float_types_ir4 = {
         "tensor(bfloat16)", "tensor(float16)", "tensor(float)", "tensor(double)"};
     return all_float_types_ir4;
   }
 
+  static const std::vector<std::string>& all_float_types_plus_Xint8_ir4() {
+    static const std::vector<std::string> all_float_types_ir4 = {
+        "tensor(bfloat16)", "tensor(float16)", "tensor(float)", "tensor(double)", "tensor(int8)", "tensor(uint8)"};
+    return all_float_types_ir4;
+  }
+
   static const std::vector<std::string>& all_float_types_ir9() {
     static const std::vector<std::string> all_float_types_ir9 = {
         "tensor(bfloat16)",
@@ -809,6 +834,16 @@ class OpSchema final {
     return all_tensor_types_ir10;
   }
 
+  static const std::vector<std::string>& all_non_complex_tensor_types_ir10() {
+    static const std::vector<std::string> all_non_complex_tensor_types_ir10 = {
+        "tensor(uint8)",      "tensor(uint16)",         "tensor(uint32)",       "tensor(uint64)",
+        "tensor(int8)",       "tensor(int16)",          "tensor(int32)",        "tensor(int64)",
+        "tensor(bfloat16)",   "tensor(float16)",        "tensor(float)",        "tensor(double)",
+        "tensor(string)",     "tensor(bool)",           "tensor(float8e4m3fn)", "tensor(float8e4m3fnuz)",
+        "tensor(float8e5m2)", "tensor(float8e5m2fnuz)", "tensor(uint4)",        "tensor(int4)"};
+    return all_non_complex_tensor_types_ir10;
+  }
+
   static const std::vector<std::string>& all_tensor_sequence_types() {
     static const std::vector<std::string> all_tensor_sequence_types = {
         "seq(tensor(uint8))",
@@ -1098,6 +1133,27 @@ class OpSchema final {
       std::set<std::string>* updated_ops = nullptr) const;
   void UpdateFunctionProtoOpsetImportVersion(FunctionProto& function_proto, int opset_version) const;
 
+  /**
+   * @brief A common function to generate a prefix string for use in fail_check during the verify function.
+   * @param  node_name If empty, the returned string will not include the node name.
+   * @return std::string The prefix string.
+   */
+  std::string VerifyFailPrefix(std::string_view node_name) const;
+
+  /**
+   * @brief Verifies if the input number matches the pattern specified in the schema.
+   * @param input_num The number of inputs to be verified against the schema.
+   * @param node_info The prefix string used if the check fails.
+   */
+  void VerifyInputNum(int input_num, std::string_view node_name = "") const;
+
+  /**
+   * @brief Verifies if the output number matches the pattern specified in the schema.
+   * @param output_num The number of outputs to be verified against the schema.
+   * @param node_info The prefix string used if the check fails.
+   */
+  void VerifyOutputNum(int output_num, std::string_view node_name = "") const;
+
   std::string name_;
   std::string file_;
   std::string doc_;
@@ -1153,7 +1209,7 @@ class OpSchemaRegistry final : public ISchemaRegistry {
       // Increase the highest version when you make BC-breaking changes to the
       // operator schema on specific domain. Update the lowest version when it's
       // determined to remove too old version history.
-      map_[ONNX_DOMAIN] = std::make_pair(1, 21);
+      map_[ONNX_DOMAIN] = std::make_pair(1, 22);
       map_[AI_ONNX_ML_DOMAIN] = std::make_pair(1, 5);
       map_[AI_ONNX_TRAINING_DOMAIN] = std::make_pair(1, 1);
       // ONNX's preview domain contains operators subject to change, so
@@ -1163,7 +1219,7 @@ class OpSchemaRegistry final : public ISchemaRegistry {
       // Version corresponding last release of ONNX. Update this to match with
       // the max version above in a *release* version of ONNX. But in other
       // versions, the max version may be ahead of the last-release-version.
-      last_release_version_map_[ONNX_DOMAIN] = 21;
+      last_release_version_map_[ONNX_DOMAIN] = 22;
       last_release_version_map_[AI_ONNX_ML_DOMAIN] = 5;
       last_release_version_map_[AI_ONNX_TRAINING_DOMAIN] = 1;
       last_release_version_map_[AI_ONNX_PREVIEW_TRAINING_DOMAIN] = 1;

onnx/defs/shape_inference.h

@@ -105,6 +105,10 @@ struct InferenceContext {
   virtual const SparseTensorProto* getInputSparseData(size_t index) const = 0;
   // Gets the shape inputs computed by partial data propagation.
   virtual const TensorShapeProto* getSymbolicInput(size_t index) const = 0;
+  // To display a name the user can use to narrow its search.
+  virtual std::string getDisplayName() const {
+    return "";
+  }
 };
 
 // We use data propagation to perform partial evaluation of the model, to compute statically
@@ -263,7 +267,15 @@ inline void propagateElemTypeFromDtypeToOutput(
   } else {
     // This is not expected to happen
     fail_type_inference(
-        "Output ", outputIndex, " expected to have: ", expected_value_case, " or UNDEFINED. Got: ", output_value_case);
+        "Output ",
+        outputIndex,
+        " expected to have: ",
+        expected_value_case,
+        " or UNDEFINED. Got: ",
+        output_value_case,
+        " in ",
+        ctx.getDisplayName(),
+        ".");
   }
 }
 
@@ -277,18 +289,18 @@ inline void propagateElemTypeFromDtypeToOutput(InferenceContext& ctx, const Attr
   const auto attr_type = attr->type();
   if (attr_type == AttributeProto::TENSOR) {
     if (attr->t().dims().size() != 1) {
-      fail_type_inference("Attribute expected to have a one-dim tensor");
+      fail_type_inference("Attribute expected to have a one-dim tensor in ", ctx.getDisplayName(), ".");
     }
     data_type = attr->t().data_type();
     expected_value_case = TypeProto::kTensorType;
   } else if (attr_type == AttributeProto::SPARSE_TENSOR) {
     if (attr->sparse_tensor().dims().size() != 1) {
-      fail_type_inference("Attribute expected to have a one-dim sparse tensor");
+      fail_type_inference("Attribute expected to have a one-dim sparse tensor in ", ctx.getDisplayName(), ".");
     }
     data_type = attr->sparse_tensor().values().data_type();
     expected_value_case = TypeProto::kSparseTensorType;
   } else {
-    fail_type_inference("Attribute expected to have tensor or sparse tensor type");
+    fail_type_inference("Attribute expected to have tensor or sparse tensor type in ", ctx.getDisplayName(), ".");
   }
 
   propagateElemTypeFromDtypeToOutput(ctx, data_type, outputIndex, expected_value_case);
@@ -326,7 +338,10 @@ inline const TensorShapeProto& getInputShape(const InferenceContext& ctx, size_t
   const auto* input_type = ctx.getInputType(n);
   const auto value_case = input_type->value_case();
   if (value_case != TypeProto::kTensorType && value_case != TypeProto::kSparseTensorType) {
-    fail_type_inference("Attribute expected to have tensor or sparse tensor type");
+    fail_type_inference("Input ", n, "expected to be a tensor or a sparse tensor type in ", ctx.getDisplayName(), ".");
+  }
+  if (!hasShape(*input_type)) {
+    fail_shape_inference("Input ", n, " must have a non null shape in ", ctx.getDisplayName(), ".");
   }
   if (value_case == TypeProto::kTensorType) {
     return input_type->tensor_type().shape();
@@ -344,7 +359,7 @@ inline const TensorShapeProto* getOptionalInputShape(InferenceContext& ctx, size
 
   const auto value_case = input_type->value_case();
   if (value_case != TypeProto::kTensorType && value_case != TypeProto::kSparseTensorType) {
-    fail_type_inference("Attribute expected to have tensor or sparse tensor type");
+    fail_type_inference("Input ", n, "expected to be a tensor or a sparse tensor type in ", ctx.getDisplayName(), ".");
   }
   if (value_case == TypeProto::kTensorType) {
     return &input_type->tensor_type().shape();
@@ -372,7 +387,10 @@ inline void appendSingleDimCopiedFromInputTypeToOutputType(
         " does not match type of output: ",
         outputIndex,
         "type: ",
-        output_value_case);
+        output_value_case,
+        " in ",
+        ctx.getDisplayName(),
+        ".");
   }
   if (TypeProto::kTensorType == input_value_case) {
     auto* dim = output_type->mutable_tensor_type()->mutable_shape()->add_dim();
@@ -382,7 +400,13 @@ inline void appendSingleDimCopiedFromInputTypeToOutputType(
     *dim = input_type->sparse_tensor_type().shape().dim(static_cast<int>(fromDimIndex));
   } else {
     fail_type_inference(
-        "Input ", inputIndex, " and Output ", outputIndex, " expected to have tensor or sparse tensor type");
+        "Input ",
+        inputIndex,
+        " and Output ",
+        outputIndex,
+        " expected to have tensor or sparse tensor type in ",
+        ctx.getDisplayName(),
+        ".");
   }
 }
 
@@ -440,7 +464,14 @@ updateOutputElemType(InferenceContext& ctx, size_t outputIndex, int32_t elemType
     setTensorElementType(elemType, expected_type, *output_type);
   } else {
     // This is not expected to happen
-    fail_type_inference("Output ", outputIndex, " expected to have tensor or sparse tensor type: ", expected_type);
+    fail_type_inference(
+        "Output ",
+        outputIndex,
+        " expected to have tensor or sparse tensor type: ",
+        expected_type,
+        " in ",
+        ctx.getDisplayName(),
+        ".");
   }
 }
 
@@ -462,16 +493,17 @@ inline void propagateElemTypeFromAttributeToOutput(
       updateOutputElemType(ctx, outputIndex, default_value, expected_type);
       return;
     } else {
-      fail_type_inference("Value of attribute ", attributeName, " not specified");
+      fail_type_inference("Value of attribute ", attributeName, " not specified in ", ctx.getDisplayName(), ".");
     }
   }
   if (!attr_proto->has_i()) {
-    fail_type_inference("Attribute ", attributeName, " should be of integer type and specify a type.");
+    fail_type_inference(
+        "Attribute ", attributeName, " should be of integer type and specify a type in ", ctx.getDisplayName(), ".");
   }
   auto attr_value = attr_proto->i();
   auto elem_type = static_cast<TensorProto_DataType>(attr_value);
   if (!TensorProto_DataType_IsValid(elem_type)) {
-    fail_type_inference("Attribute ", attributeName, " does not specify a valid type.");
+    fail_type_inference("Attribute ", attributeName, " does not specify a valid type in ", ctx.getDisplayName(), ".");
   }
   updateOutputElemType(ctx, outputIndex, elem_type, expected_type);
 }
@@ -497,7 +529,7 @@ inline TensorShapeProto*
 getOutputShape(InferenceContext& ctx, size_t n, TypeProto::ValueCase default_type = TypeProto::kTensorType) {
   auto output_type = ctx.getOutputType(n);
   if (output_type == nullptr) {
-    fail_type_inference("Output ", n, " expected to have tensor or sparse type");
+    fail_type_inference("Output ", n, " expected to have tensor or sparse type in ", ctx.getDisplayName(), ".");
   }
   const auto output_value_case = output_type->value_case();
   if (output_value_case == TypeProto::kTensorType || output_value_case == TypeProto::kSparseTensorType) {
@@ -505,7 +537,7 @@ getOutputShape(InferenceContext& ctx, size_t n, TypeProto::ValueCase default_typ
   } else if (output_value_case == TypeProto::VALUE_NOT_SET) {
     return getTensorMutableShape(default_type, *output_type);
   } else {
-    fail_type_inference("Output ", n, " expected to have tensor type");
+    fail_type_inference("Output ", n, " expected to have tensor type in ", ctx.getDisplayName(), ".");
   }
 }
 
@@ -562,13 +594,13 @@ inline void propagateShapeFromAttributeToOutput(
   auto attr_proto = ctx.getAttribute(attributeName);
   if ((nullptr == attr_proto) || (!attr_proto->has_type()) ||
       (attr_proto->type() != AttributeProto_AttributeType_INTS)) {
-    fail_shape_inference("Attribute ", attributeName, " should specify a shape");
+    fail_shape_inference("Attribute ", attributeName, " should specify a shape in ", ctx.getDisplayName(), ".");
   }
   auto& int_list = attr_proto->ints();
   TensorShapeProto shape;
   for (auto dim_size : int_list) {
     if (dim_size < 0) {
-      fail_shape_inference("Negative values are not allowed in a shape specification");
+      fail_shape_inference("Negative values are not allowed in a shape specification in ", ctx.getDisplayName(), ".");
     }
     shape.add_dim()->set_dim_value(dim_size);
   }
@@ -745,7 +777,16 @@ inline void checkInputRank(InferenceContext& ctx, size_t input_index, int expect
   if (hasInputShape(ctx, input_index)) {
     auto rank = getInputShape(ctx, input_index).dim_size();
     if (rank != expected_rank) {
-      fail_shape_inference("Input ", input_index, " expected to have rank ", expected_rank, " but has rank ", rank);
+      fail_shape_inference(
+          "Input ",
+          input_index,
+          " expected to have rank ",
+          expected_rank,
+          " but has rank ",
+          rank,
+          " in ",
+          ctx.getDisplayName(),
+          ".");
     }
   }
 }
@@ -798,7 +839,15 @@ inline void unifyInputDim(InferenceContext& ctx, size_t input_index, int dim_ind
     // This shape is expected to have rank > dim_index:
     if (input_shape.dim_size() <= dim_index) {
       fail_shape_inference(
-          "Input ", input_index, " expected to have rank >", dim_index, " but has rank ", input_shape.dim_size());
+          "Input ",
+          input_index,
+          " expected to have rank >",
+          dim_index,
+          " but has rank ",
+          input_shape.dim_size(),
+          " in ",
+          ctx.getDisplayName(),
+          ".");
     }
     const Dim& input_dim = input_shape.dim(dim_index);
     // Now, unify dim and input_dim:

onnx/defs/tensor/defs.cc

@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <cmath>
 #include <numeric>
+#include <optional>
 
 #include "onnx/defs/data_propagators.h"
 #include "onnx/defs/function.h"
@@ -135,7 +136,7 @@ ONNX_OPERATOR_SET_SCHEMA(
           PropagateShapeDataFromInputToOutput(ctx, 0);
         }));
 
-static const char* CastLike_ver19_doc = R"DOC(
+static const char* CastLike_ver21_doc = R"DOC(
 The operator casts the elements of a given input tensor (the first input) to
 the same data type as the elements of the second input tensor.
 See documentation of the Cast operator for further details.
@@ -145,7 +146,7 @@ ONNX_OPERATOR_SET_SCHEMA(
     CastLike,
     21,
     OpSchema()
-        .SetDoc(CastLike_ver19_doc)
+        .SetDoc(CastLike_ver21_doc)
         .Attr(
             "saturate",
             "The parameter defines how the conversion behaves if an input value is out of "
@@ -175,19 +176,11 @@ ONNX_OPERATOR_SET_SCHEMA(
             OpSchema::Differentiable)
         .TypeConstraint(
             "T1",
-            {"tensor(float16)",    "tensor(float)",          "tensor(double)",       "tensor(int8)",
-             "tensor(int16)",      "tensor(int32)",          "tensor(int64)",        "tensor(uint8)",
-             "tensor(uint16)",     "tensor(uint32)",         "tensor(uint64)",       "tensor(bool)",
-             "tensor(string)",     "tensor(bfloat16)",       "tensor(float8e4m3fn)", "tensor(float8e4m3fnuz)",
-             "tensor(float8e5m2)", "tensor(float8e5m2fnuz)", "tensor(uint4)",        "tensor(int4)"},
+            OpSchema::all_non_complex_tensor_types_ir10(),
             "Constrain input types. Casting from complex is not supported.")
         .TypeConstraint(
             "T2",
-            {"tensor(float16)",    "tensor(float)",          "tensor(double)",       "tensor(int8)",
-             "tensor(int16)",      "tensor(int32)",          "tensor(int64)",        "tensor(uint8)",
-             "tensor(uint16)",     "tensor(uint32)",         "tensor(uint64)",       "tensor(bool)",
-             "tensor(string)",     "tensor(bfloat16)",       "tensor(float8e4m3fn)", "tensor(float8e4m3fnuz)",
-             "tensor(float8e5m2)", "tensor(float8e5m2fnuz)", "tensor(uint4)",        "tensor(int4)"},
+            OpSchema::all_non_complex_tensor_types_ir10(),
             "Constrain output types. Casting to complex is not supported.")
         .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
           propagateElemTypeFromInputToOutput(ctx, 1, 0);
@@ -2323,7 +2316,7 @@ ONNX_OPERATOR_SET_SCHEMA(
         .SetDoc(Resize_ver19_doc)
         .TypeAndShapeInferenceFunction([](InferenceContext& ctx) { resizeShapeInference_opset18_to_19(ctx); }));
 
-static const char* GridSample_ver20_doc = R"DOC(
+static const char* GridSample_ver22_doc = R"DOC(
 Given an input `X` and a flow-field `grid`, computes the output `Y` using `X` values and pixel locations from the `grid`.
 For spatial input `X` with shape (N, C, H, W), the `grid` will have shape (N, H_out, W_out, 2),
 the output `Y` will have shape (N, C, H_out, W_out). For volumetric input `X` with shape (N, C, D, H, W),
@@ -2346,7 +2339,7 @@ See also in [torch.nn.functional.grid_sample](https://pytorch.org/docs/stable/ge
 
 ONNX_OPERATOR_SET_SCHEMA(
     GridSample,
-    20,
+    22,
     OpSchema()
         .Attr(
             "mode",
@@ -2412,13 +2405,10 @@ ONNX_OPERATOR_SET_SCHEMA(
             OpSchema::Differentiable)
         .TypeConstraint(
             "T1",
-            OpSchema::all_tensor_types(),
+            OpSchema::all_tensor_types_ir4(),
             "Constrain input `X` and output `Y` types to all tensor types.")
-        .TypeConstraint(
-            "T2",
-            {"tensor(float16)", "tensor(float)", "tensor(double)"},
-            "Constrain grid types to float tensors.")
-        .SetDoc(GridSample_ver20_doc)
+        .TypeConstraint("T2", OpSchema::all_float_types_ir4(), "Constrain grid types to float tensors.")
+        .SetDoc(GridSample_ver22_doc)
         .TypeAndShapeInferenceFunction([](InferenceContext& ctx) { gridSampleShapeInference(ctx); }));
 
 static const char* AffineGrid_ver20_doc = R"DOC(
@@ -2895,8 +2885,18 @@ ONNX_OPERATOR_SET_SCHEMA(
           // and 1 element vector for now. In future when version update for
           // this op is done we should only allow scalar or change the spec to
           // allow both.
+          std::optional<int64_t> depth_value;
           if (hasInputShape(ctx, 1)) {
             auto& depth_shape = getInputShape(ctx, 1);
+            if (const TensorProto* depth_data = ctx.getInputData(1)) {
+              if (depth_data->data_type() == TensorProto::INT64) {
+                depth_value = ParseData<int64_t>(depth_data)[0];
+              } else if (depth_data->data_type() == TensorProto::INT32) {
+                depth_value = ParseData<int32_t>(depth_data)[0];
+              } else if (depth_data->data_type() == TensorProto::FLOAT) {
+                depth_value = static_cast<int64_t>(ParseData<float>(depth_data)[0]);
+              }
+            }
             if (depth_shape.dim_size() != 0 && depth_shape.dim_size() != 1) {
               fail_type_inference("Input 'depth' must be a scalar or rank 1 tensor.");
             }
@@ -2947,6 +2947,8 @@ ONNX_OPERATOR_SET_SCHEMA(
                 } else if (indices_shape.dim(i - 1).has_dim_param()) {
                   dim->set_dim_param(indices_shape.dim(i - 1).dim_param());
                 }
+              } else if (depth_value) {
+                dim->set_dim_value(*depth_value);
               }
             }
           }

onnx/defs/tensor/old.cc

@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <cmath>
 #include <numeric>
+#include <optional>
 
 #include "onnx/defs/data_propagators.h"
 #include "onnx/defs/function.h"
@@ -12,6 +13,104 @@
 
 namespace ONNX_NAMESPACE {
 
+static const char* GridSample_ver20_doc = R"DOC(
+Given an input `X` and a flow-field `grid`, computes the output `Y` using `X` values and pixel locations from the `grid`.
+For spatial input `X` with shape (N, C, H, W), the `grid` will have shape (N, H_out, W_out, 2),
+the output `Y` will have shape (N, C, H_out, W_out). For volumetric input `X` with shape (N, C, D, H, W),
+the `grid` will have shape (N, D_out, H_out, W_out, 3), the output `Y` will have shape (N, C, D_out, H_out, W_out).
+More generally, for an input `X` of rank r+2 with shape (N, C, d1, d2, ..., dr),
+the `grid` will have shape (N, D1_out, D2_out, ..., Dr_out, r), the output `Y` will have shape (N, C, D1_out, D2_out, ..., Dr_out).
+
+The tensor `X` contains values at centers of square pixels (voxels, etc) locations such as (n, c, d1_in, d2_in, ..., dr_in).
+The (n, d1_out, d2_out, ..., dr_out, :) values from the tensor `grid` are the normalized positions for interpolating the values
+at the (n, c, d1_out, d2_out, ..., dr_out) locations from the output tensor `Y` using a specified interpolation method (the mode)
+and a padding mode (for `grid` positions falling outside the 2-dimensional image).
+
+For example, the values in `grid[n, h_out, w_out, :]` are size-2 vectors specifying normalized positions in the 2-dimensional space of `X`.
+They are used to interpolate output values of `Y[n, c, h_out, w_out]`.
+
+The GridSample operator is often used in doing grid generator and sampler in the
+[Spatial Transformer Networks](https://arxiv.org/abs/1506.02025).
+See also in [torch.nn.functional.grid_sample](https://pytorch.org/docs/stable/generated/torch.nn.functional.grid_sample.html).
+)DOC";
+
+ONNX_OPERATOR_SET_SCHEMA(
+    GridSample,
+    20,
+    OpSchema()
+        .Attr(
+            "mode",
+            "Three interpolation modes: linear (default), nearest and cubic. "
+            "The \"linear\" mode includes linear and N-linear interpolation modes depending on the number of spatial dimensions "
+            "of the input tensor (i.e. linear for 1 spatial dimension, bilinear for 2 spatial dimensions, etc.). "
+            "The \"cubic\" mode also includes N-cubic interpolation modes following the same rules. The \"nearest\" mode rounds "
+            "to the nearest even index when the sampling point falls halfway between two indices.",
+            AttributeProto::STRING,
+            std::string("linear"))
+        .Attr(
+            "padding_mode",
+            "Support padding modes for outside grid values: `zeros`(default), `border`, `reflection`. "
+            "zeros: use 0 for out-of-bound grid locations, "
+            "border: use border values for out-of-bound grid locations, "
+            "reflection: use values at locations reflected by the border for out-of-bound grid locations. "
+            "If index 0 represents the margin pixel, the reflected value at index -1 will be the same as the value at index 1. "
+            "For location far away from the border, it will keep being reflected until becoming in bound. "
+            "If pixel location x = -3.5 reflects by border -1 and becomes x' = 1.5, then reflects by border 1 and becomes x'' = 0.5.",
+            AttributeProto::STRING,
+            std::string("zeros"))
+        .Attr(
+            "align_corners",
+            "If align_corners=1, the extrema (-1 and 1) are considered as referring to the center points of the input's corner pixels (voxels, etc.). "
+            "If align_corners=0, they are instead considered as referring to the corner points of the input's corner pixels (voxels, etc.), "
+            "making the sampling more resolution agnostic.",
+            AttributeProto::INT,
+            static_cast<int64_t>(0))
+        .Input(
+            0,
+            "X",
+            "Input tensor of rank r+2 that has shape (N, C, D1, D2, ..., Dr), where N is the batch size, "
+            "C is the number of channels, D1, D2, ..., Dr are the spatial dimensions.",
+            "T1",
+            OpSchema::Single,
+            true,
+            1,
+            OpSchema::Differentiable)
+        .Input(
+            1,
+            "grid",
+            "Input offset of shape (N, D1_out, D2_out, ..., Dr_out, r), where D1_out, D2_out, ..., "
+            "Dr_out are the spatial dimensions of the grid and output, and r is the number of spatial dimensions. "
+            "Grid specifies the sampling locations normalized by the input spatial dimensions. "
+            "Therefore, it should have most values in the range of [-1, 1]. If the grid has values outside the range of [-1, 1], "
+            "the corresponding outputs will be handled as defined by padding_mode. Following computer vision convention, "
+            "the coordinates in the length-r location vector are listed from the innermost tensor dimension to the outermost, "
+            "the opposite of regular tensor indexing.",
+            "T2",
+            OpSchema::Single,
+            true,
+            1,
+            OpSchema::NonDifferentiable)
+        .Output(
+            0,
+            "Y",
+            "Output tensor of rank r+2 that has shape (N, C, D1_out, D2_out, ..., Dr_out) of the sampled values. "
+            "For integer input types, intermediate values are computed as floating point and cast to integer at the end.",
+            "T1",
+            OpSchema::Single,
+            true,
+            1,
+            OpSchema::Differentiable)
+        .TypeConstraint(
+            "T1",
+            OpSchema::all_tensor_types(),
+            "Constrain input `X` and output `Y` types to all tensor types.")
+        .TypeConstraint(
+            "T2",
+            {"tensor(float16)", "tensor(float)", "tensor(double)"},
+            "Constrain grid types to float tensors.")
+        .SetDoc(GridSample_ver20_doc)
+        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) { gridSampleShapeInference(ctx); }));
+
 static const char* Cast_ver19_doc = R"DOC(
 The operator casts the elements of a given input tensor to a data type
 specified by the 'to' argument and returns an output tensor of the same size in
@@ -1380,7 +1479,7 @@ ONNX_OPERATOR_SET_SCHEMA(
 
 static const char* Slice_ver11_doc = R"DOC(
 Produces a slice of the input tensor along multiple axes. Similar to numpy:
-https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html
+https://numpy.org/doc/stable/reference/routines.indexing.html
 Slices uses `starts`, `ends`, `axes` and `steps` inputs to specify the start and end
 dimension and step for each axis in the list of axes, it uses this information to
 slice the input `data` tensor. If a negative value is passed for any of the
@@ -4443,7 +4542,7 @@ ONNX_OPERATOR_SET_SCHEMA(
 
 static const char* Slice_ver1_doc = R"DOC(
 Produces a slice of the input tensor along multiple axes. Similar to numpy:
-https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html
+https://numpy.org/doc/stable/reference/routines.indexing.html
 Slices uses `axes`, `starts` and `ends` attributes to specify the start and end
 dimension for each axis in the list of axes, it uses this information to
 slice the input `data` tensor. If a negative value is passed for any of the
@@ -4559,7 +4658,7 @@ ONNX_OPERATOR_SET_SCHEMA(
 
 static const char* Slice_ver10_doc = R"DOC(
 Produces a slice of the input tensor along multiple axes. Similar to numpy:
-https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html
+https://numpy.org/doc/stable/reference/routines.indexing.html
 Slices uses `starts`, `ends`, `axes` and `steps` inputs to specify the start and end
 dimension and step for each axis in the list of axes, it uses this information to
 slice the input `data` tensor. If a negative value is passed for any of the
@@ -5152,8 +5251,18 @@ ONNX_OPERATOR_SET_SCHEMA(
           // and 1 element vector for now. In future when version update for
           // this op is done we should only allow scalar or change the spec to
           // allow both.
+          std::optional<int64_t> depth_value;
           if (hasInputShape(ctx, 1)) {
             auto& depth_shape = getInputShape(ctx, 1);
+            if (const TensorProto* depth_data = ctx.getInputData(1)) {
+              if (depth_data->data_type() == TensorProto::INT64) {
+                depth_value = ParseData<int64_t>(depth_data)[0];
+              } else if (depth_data->data_type() == TensorProto::INT32) {
+                depth_value = ParseData<int32_t>(depth_data)[0];
+              } else if (depth_data->data_type() == TensorProto::FLOAT) {
+                depth_value = static_cast<int64_t>(ParseData<float>(depth_data)[0]);
+              }
+            }
             if (depth_shape.dim_size() != 0 && depth_shape.dim_size() != 1) {
               fail_type_inference("Input 'depth' must be a scalar or rank 1 tensor.");
             }
@@ -5204,6 +5313,8 @@ ONNX_OPERATOR_SET_SCHEMA(
                 } else if (indices_shape.dim(i - 1).has_dim_param()) {
                   dim->set_dim_param(indices_shape.dim(i - 1).dim_param());
                 }
+              } else if (depth_value) {
+                dim->set_dim_value(*depth_value);
               }
             }
           }