Revert D10098106: [pytorch][PR] [WIP] New version of PT1 model format

Differential Revision:
D10098106

Original commit changeset: 94ec7fc57c84

fbshipit-source-id: 38f729b0970618f38359797b806cbbcd865f4715

9 files changed, 1058 insertions, 139 deletions

diff --git a/caffe2/core/blob_serialization.cc b/caffe2/core/blob_serialization.cc
index f27d16adf3..8126b3d594 100644
--- a/caffe2/core/blob_serialization.cc
+++ b/caffe2/core/blob_serialization.cc
@@ -309,12 +309,6 @@ void TensorSerializer::Serialize(
         proto.add_string_data(SerializeBlob(temp_blob, ""));
       }
     } break;
-    case TensorProto_DataType_SPECIAL: {
-      CAFFE_THROW("SPECIAL Tensor is not handled yet.");
-    } break;
-    case TensorProto_DataType_NO_CONTENT: {
-      CAFFE_THROW("NO_CONTENT Tensor should not be serialized.");
-    } break;
       // Note: we intentially do not provide "default:" so if any new data types
       // are added, the compiler should warn the user to add the case here.
   }
@@ -526,14 +520,7 @@ void TensorDeserializer::Deserialize(const TensorProto& proto, Tensor* tensor) {
                 (i + chunkBegin) * temp_blob.meta().itemsize(),
             1);
       }
-    } break;
-    case TensorProto_DataType_SPECIAL: {
-      CAFFE_THROW("SPECIAL Tensor is not handled yet.");
-    } break;
-    case TensorProto_DataType_NO_CONTENT: {
-      CAFFE_THROW("NO_CONTENT Tensor should not be deserialized.");
-    } break;
-      // Note: we intentially do not provide "default:" so if any new data types
+    }
   }
   context->FinishDeviceComputation();
 }
diff --git a/caffe2/proto/caffe2.proto b/caffe2/proto/caffe2.proto
index 9dc745edbd..7187001029 100644
--- a/caffe2/proto/caffe2.proto
+++ b/caffe2/proto/caffe2.proto
@@ -15,46 +15,23 @@ package caffe2;
 message TensorProto {
   // The dimensions in the tensor.
   repeated int64 dims = 1;
-  // The strides of the tensor.
-  repeated int64 strides = 12;
-
-  // Data type
   enum DataType {
     UNDEFINED = 0;
-
-    // Basic types
-    FLOAT = 1;     // float
-    INT32 = 2;     // int
-    BYTE = 3;      // byte, when deserialized, is going to be restored as uint8
-    STRING = 4;    // string
-
-    // Less-commonly used data types
-    BOOL = 5;      // bool
-    UINT8 = 6;     // uint8_t
-    INT8 = 7;      // int8_t
-    UINT16 = 8;    // uint16_t
-    INT16 = 9;     // int16_t
-    INT64 = 10;    // int64_t
+    FLOAT = 1;  // float
+    INT32 = 2;  // int
+    BYTE = 3;  // BYTE, when deserialized, is going to be restored as uint8.
+    STRING = 4;  // string
+    // Less-commonly used data types.
+    BOOL = 5;  // bool
+    UINT8 = 6;  // uint8_t
+    INT8 = 7;  // int8_t
+    UINT16 = 8;  // uint16_t
+    INT16 = 9;  // int16_t
+    INT64 = 10;  // int64_t
     FLOAT16 = 12;  // at::Half
-    DOUBLE = 13;   // double
-
-    // Special data type, type information is stored in the special type field
-    SPECIAL = 51;
-    // Use TensorProto to specify the shape and type
-    NO_CONTENT = 52;
+    DOUBLE = 13;  // double
   }
   optional DataType data_type = 2 [default = FLOAT];
-  // if data_type is SPECIAL, use this field to express the type info
-  optional SpecialType special_type = 13;
-
-  // Data storage
-  enum StorageType {
-    TYPED = 1;
-    RAW = 2;
-    EXTERNAL = 3;
-    ALIAS = 4;
-  }
-  optional StorageType storage_type = 14 [default = TYPED];
   // For float
   repeated float float_data = 3 [packed = true];
   // For int32, uint8, int8, uint16, int16, bool, and float16
@@ -69,13 +46,6 @@ message TensorProto {
   repeated double double_data = 9 [packed = true];
   // For int64
   repeated int64 int64_data = 10 [packed = true];
-  // For raw data
-  optional bytes raw_data = 15;
-  // External data by file name
-  optional string external_data = 16;
-  // For argument, to share the content
-  optional string alias = 17;
-
   // Optionally, a name for the tensor.
   optional string name = 7;
 
@@ -83,23 +53,13 @@ message TensorProto {
   // it was serialized from. This is useful in cases like snapshotting a whole
   // workspace in a multi-GPU environment.
   optional DeviceOption device_detail = 8;
-
   // When loading from chunks this is going to indicate where to put data in the
   // full array. When not used full data have to be present
   message Segment {
     required int64 begin = 1;
     required int64 end = 2;
-    optional int64 chunk_num = 51;
-    optional int64 chunk_id = 52;
   }
   optional Segment segment = 11;
-  optional string debug_info = 18;
-
-  // For PyTorch serialized tensor.
-  optional bool require_gradient = 19;
-  optional bool is_buffer = 20;
-
-  repeated Argument annotations = 21;
 }
 
 message QTensorProto {
@@ -126,11 +86,7 @@ message TensorShape {
   repeated int32 unknown_dims = 3;
   optional bool unknown_shape = 4 [default = false];
   optional string name = 5;
-}
 
-// This is prepared for non-tensor types.
-message SpecialType {
-  optional string name = 1;
 }
 
 message TensorShapes {
@@ -141,17 +97,13 @@ message TensorShapes {
 // values, or repeated float, int and string arrays.
 message Argument {
   optional string name = 1;
-
   optional float f = 2;
   optional int64 i = 3;
   optional bytes s = 4;
-  optional TensorProto t = 10;
   optional NetDef n = 8;
-
   repeated float floats = 5;
   repeated int64 ints = 6;
   repeated bytes strings = 7;
-  repeated TensorProto tensors = 11;
   repeated NetDef nets = 9;
 }
 
@@ -200,11 +152,7 @@ message DeviceOption {
 // Operator Definition.
 message OperatorDef {
   repeated string input = 1; // the name of the input blobs
-  // the input name in the schema, for named inputs
-  repeated string mapped_inputs = 11;
   repeated string output = 2; // the name of output top blobs
-  // the outputname in the schema, for named outputs
-  repeated string mapped_outputs = 12;
   optional string name = 3; // the operator name. This is optional.
   // the operator type. This is needed to create the object from the operator
   // registry.
@@ -238,16 +186,6 @@ message OperatorDef {
   // This is an optional string with no assumed characteristics as
   // operators can be constructed in any language.
   optional string debug_info = 10;
-
-  // additional annotations
-  repeated Argument annotations = 13;
-
-  // for jit ir exporting
-  optional string aten_function = 14;
-
-  // for operator versioning
-  optional string domain = 15;
-  optional string op_version = 16;
 }
 
 // Network definition.
diff --git a/caffe2/proto/torch.proto b/caffe2/proto/torch.proto
index f31c3b65ec..43dfd02b14 100644
--- a/caffe2/proto/torch.proto
+++ b/caffe2/proto/torch.proto
@@ -4,77 +4,547 @@ import "caffe2/proto/caffe2.proto";
 
 package torch;
 
-enum ProtoVersion {
+// Overview
+//
+// ONNX is an open specification that is comprised of the following components:
+//
+// 1)  A definition of an extensible computation graph model.
+// 2)  Definitions of standard data types.
+// 3)  Definitions of built-in operators.
+//
+// This document describes the syntax of models and their computation graphs,
+// as well as the standard data types. Together, they are referred to as the ONNX
+// Intermediate Representation, or 'IR' for short.
+//
+// The normative semantic specification of the ONNX IR is found in docs/IR.md.
+// Definitions of the built-in neural network operators may be found in docs/Operators.md.
+
+// Notes
+//
+// Release
+//
+// We are still in the very early stage of defining ONNX. The current
+// version of ONNX is a starting point. While we are actively working
+// towards a complete spec, we would like to get the community involved
+// by sharing our working version of ONNX.
+//
+// Protobuf compatibility
+//
+// To simplify framework compatibility, ONNX is defined using the subset of
+// protobuf that is compatible with both protobuf v2 and v3. This means that we
+// do not use any protobuf features that are only available in one of the two
+// versions.
+//
+// Here are the most notable contortions we have to carry out to work around
+// these limitations:
+//
+//   - No 'map' (added protobuf 3.0). We instead represent mappings as lists
+//     of key-value pairs, where order does not matter and duplicates
+//     are not allowed.
+
+// Versioning
+//
+// ONNX versioning is specified in docs/IR.md and elaborated on in docs/Versioning.md
+//
+// To be compatible with both proto2 and proto3, we will use a version number
+// that is not defined by the default value but an explicit enum number.
+enum Version {
+  // proto3 requires the first enum value to be zero.
+  // We add this just to appease the compiler.
   _START_VERSION = 0;
-  IR_VERSION_NEWEST = 0x0000000000000101;
+  // The version field is always serialized and we will use it to store the
+  // version that the  graph is generated from. This helps us set up version
+  // control.
+  // For the IR, we are using simple numbers starting with with 0x00000001,
+  // which was the version we published on Oct 10, 2017.
+  IR_VERSION_2017_10_10 = 0x0000000000000001;
+
+  // IR_VERSION 2 published on Oct 30, 2017
+  // - Added type discriminator to AttributeProto to support proto3 users
+  IR_VERSION_2017_10_30 = 0x0000000000000002;
+
+  // IR VERSION 3 published on Nov 3, 2017
+  // - For operator versioning:
+  //    - Added new message OperatorSetIdProto
+  //    - Added opset_import in ModelProto
+  // - For vendor extensions, added domain in NodeProto
+  IR_VERSION_NEWEST_ONNX = 0x0000000000000003;
+
+  // PYTORCH IR VERSION
+  IR_VERSION_NEWEST = 0x0000000000000103;
 }
 
-message MethodDef {
-  // method name
-  optional string name = 1; // method name
+// Attributes
+//
+// A named attribute containing either singular float, integer, string, graph,
+// and tensor values, or repeated float, integer, string, graph, and tensor values.
+// An AttributeProto MUST contain the name field, and *only one* of the
+// following content fields, effectively enforcing a C/C++ union equivalent.
+message AttributeProto {
+
+  // Note: this enum is structurally identical to the OpSchema::AttrType
+  // enum defined in schema.h.  If you rev one, you likely need to rev the other.
+  enum AttributeType {
+    UNDEFINED = 0;
+    FLOAT = 1;
+    INT = 2;
+    STRING = 3;
+    TENSOR = 4;
+    GRAPH = 5;
+
+    FLOATS = 6;
+    INTS = 7;
+    STRINGS = 8;
+    TENSORS = 9;
+    GRAPHS = 10;
+  }
+
+  // The name field MUST be present for this version of the IR.
+  optional string name = 1;           // namespace Attribute
 
-  // static graph
-  optional caffe2.NetDef graph = 2;
-  // method is represented as torch script
-  optional string torch_script = 3;
+  // if ref_attr_name is not empty, ref_attr_name is the attribute name in parent function.
+  // In this case, this AttributeProto does not contain data, and it's a reference of attribute
+  // in parent scope.
+  // NOTE: This should ONLY be used in function (sub-graph). It's invalid to be used in main graph.
+  optional string ref_attr_name = 21;
 
-  // the names of inputs and outputs
-  repeated string inputs = 4;
-  repeated string outputs = 5;
+  // A human-readable documentation for this attribute. Markdown is allowed.
+  optional string doc_string = 13;
 
-  // whether this method is main or not.
-  // by default, `forward` should the main method.
-  optional bool is_main = 6;
+  // The type field MUST be present for this version of the IR.
+  // For 0.0.1 versions of the IR, this field was not defined, and
+  // implementations needed to use has_field hueristics to determine
+  // which value field was in use.  For IR_VERSION 0.0.2 or later, this
+  // field MUST be set and match the f|i|s|t|... field in use.  This
+  // change was made to accomodate proto3 implementations.
+  optional AttributeType type = 20;   // discriminator that indicates which field below is in use
 
-  optional string debug_info = 7;
+  // Exactly ONE of the following fields must be present for this version of the IR
+  optional float f = 2;               // float
+  optional int64 i = 3;               // int
+  optional bytes s = 4;               // UTF-8 string
+  optional TensorProto t = 5;         // tensor value
+  optional GraphProto g = 6;          // graph
+  // Do not use field below, it's deprecated.
+  // optional ValueProto v = 12;         // value - subsumes everything but graph
 
-  repeated caffe2.Argument annotations = 8;
+  repeated float floats = 7;          // list of floats
+  repeated int64 ints = 8;            // list of ints
+  repeated bytes strings = 9;         // list of UTF-8 strings
+  repeated TensorProto tensors = 10;  // list of tensors
+  repeated GraphProto graphs = 11;    // list of graph
 }
 
+// Defines information on value, including the name, the type, and
+// the shape of the value.
+message ValueInfoProto {
+  // This field MUST be present in this version of the IR.
+  optional string name = 1;     // namespace Value
+  // This field MUST be present in this version of the IR.
+  optional TypeProto type = 2;
+  // A human-readable documentation for this value. Markdown is allowed.
+  optional string doc_string = 3;
+}
+
+// Nodes
+//
+// Computation graphs are made up of a DAG of nodes, which represent what is
+// commonly called a "layer" or "pipeline stage" in machine learning frameworks.
+//
+// For example, it can be a node of type "Conv" that takes in an image, a filter
+// tensor and a bias tensor, and produces the convolved output.
+message NodeProto {
+  repeated string input = 1;    // namespace Value
+  repeated string output = 2;   // namespace Value
 
-message ModuleDef {
-  repeated ModuleDef submodules = 1;
+  // An optional identifier for this node in a graph.
+  // This field MAY be absent in ths version of the IR.
+  optional string name = 3;     // namespace Node
 
-  // We suppose to store the modules in one of the following format:
-  //   - methods (static graph or torch script)
-  //   - pickle
-  //   - cpp_arena
-  repeated MethodDef methods = 2;
-  // because the old pickle modules may not be supported by torch_script,
-  // have to stored as pickle_arena at this moment.
-  optional bytes pickle_arena = 3;
-  // should be exposed by the Class Archive, so user can save
-  // module specific data which cannot be store in the graph or torch_script
-  optional bytes cpp_arena = 4;
+  // The symbolic identifier of the Operator to execute.
+  optional string op_type = 4;  // namespace Operator
+  // The domain of the OperatorSet that specifies the operator named by op_type.
+  optional string domain = 7;   // namespace Domain
 
-  // the names of inputs and outputs of the module are inferred
-  // from the main method.
+  // Additional named attributes.
+  repeated AttributeProto attribute = 5;
 
-  optional string debug_info = 5;
+  // A human-readable documentation for this node. Markdown is allowed.
+  // Equivalent to string debug_info
+  optional string doc_string = 6;
 
-  repeated caffe2.Argument annotations = 6;
+  // Additional annotations, attributes are defined in Schema
+  // To be added as annotations:
+  //    string engine
+  //    string list control_input
+  //    int64 is_gradient_op
+  repeated AttributeProto annotations = 8;
+
+  // Besides the node type, PyTorhc also serialize ATen function signature
+  optional caffe2.DeviceOption device_option = 51;
+  optional string aten_function = 52;
 }
 
-message ModelDef {
+// Models
+//
+// ModelProto is a top-level file/container format for bundling a ML model and
+// associating its computation graph with metadata.
+//
+// The semantics of the model are described by the associated GraphProto.
+//
+// Model ==> Caffe2 MetaNetDef
+//       ==> PyTorch Module
+message ModelProto {
+  // The version of the IR this model targets. See Version enum above.
+  // This field MUST be present.
   optional int64 ir_version = 1;
 
-  // main module of the model
-  optional ModuleDef main_module = 2;
+  // The OperatorSets this model relies on.
+  // All ModelProtos MUST have at least one entry that
+  // specifies which version of the ONNX OperatorSet is
+  // being imported.
+  //
+  // All nodes in the ModelProto's graph will bind against the operator
+  // with the same-domain/same-op_type operator with the HIGHEST version
+  // in the referenced operator sets.
+  repeated OperatorSetIdProto opset_import = 8;
+
+  // The name of the framework or tool used to generate this model.
+  // This field SHOULD be present to indicate which implementation/tool/framework
+  // emitted the model.
+  optional string producer_name = 2;
+
+  // The version of the framework or tool used to generate this model.
+  // This field SHOULD be present to indicate which implementation/tool/framework
+  // emitted the model.
+  optional string producer_version = 3;
+
+  // Domain name of the model.
+  // We use reverse domain names as name space indicators. For example:
+  // `com.facebook.fair` or `com.microsoft.cognitiveservices`
+  //
+  // Together with `model_version` and GraphProto.name, this forms the unique identity of
+  // the graph.
+  optional string domain = 4;
+
+  // The version of the graph encoded. See Version enum below.
+  optional int64 model_version = 5;
+
+  // A human-readable documentation for this model. Markdown is allowed.
+  optional string doc_string = 6;
 
-  repeated caffe2.TensorProto parameters = 3;
-  repeated caffe2.TensorProto value_infos = 4;
+  // The parameterized graph that is evaluated to execute the model.
+  // The main graph, in single graph case, it is ONNX compatible.
+  optional GraphProto graph = 7;
 
-  // to distinguish whether exported from c2 or torch
-  optional string producer_name = 5;
+  // The remaining nets in MetaNetDef.
+  // Submodules and methods in PyTorch.
+  repeated GraphProto methods = 15;
+
+  // Named metadata values; keys should be distinct.
+  // Many meta data in MetaNetDef and preditor are piggy backed here.
+  // 1) project
+  // 2) model_class
+  // 3) internal_version
+  // 4) predictor_type
+  // 5) predictor_id
+  // 6) execute_plan
+  // 7) applicationSpecificInfo (another string map, need to verify it has no duplicate.)
+  // 8) engine
+  // 9) publish time
+  repeated StringStringEntryProto metadata_props = 14;
+
+  // Model name
+  optional string name = 16;
+
+  // Model name
+  repeated AttributeProto annotations = 17;
+
+  // Mapping from list name to blob name list, must be string list type.
+  // Equivalent to blobs in MetaNetDef.
+  repeated AttributeProto blob_lists = 51;
+
+  // Mapping from plan name to serialized plan, must be string list type.
+  // Equivalent to plans in MetaNetDef.
+  repeated AttributeProto plans = 52;
+};
+
+// StringStringEntryProto follows the pattern for cross-proto-version maps.
+// See https://developers.google.com/protocol-buffers/docs/proto3#maps
+message StringStringEntryProto {
+  optional string key = 1;
+  optional string value= 2;
+};
+
+// Graphs
+//
+// A graph defines the computational logic of a model and is comprised of a parameterized
+// list of nodes that form a directed acyclic graph based on their inputs and outputs.
+// This is the equivalent of the "network" or "graph" in many deep learning
+// frameworks.
+// Graph ==> NetDef in Caffe2
+//       ==> Submodule/Method in PyTorch
+message GraphProto {
+  // The nodes in the graph, sorted topologically.
+  repeated NodeProto node = 1;
+
+  // The name of the graph.
+  optional string name = 2;   // namespace Graph
+
+  // A list of named tensor values, used to specify constant inputs of the graph.
+  // Each TensorProto entry must have a distinct name (within the list) that
+  // also appears in the input list.
+  repeated TensorProto initializer = 5;
+
+  // A human-readable documentation for this graph. Markdown is allowed.
+  optional string doc_string = 10;
+
+  // The inputs and outputs of the graph.
+  repeated ValueInfoProto input = 11;
+  repeated ValueInfoProto output = 12;
+
+  // Information for the values in the graph. The ValueInfoProto.name's
+  // must be distinct. It is optional for a value to appear in value_info list.
+  repeated ValueInfoProto value_info = 13;
+
+  // Additional annotations.
+  repeated AttributeProto annotations = 14;
+
+  // DO NOT USE the following fields, they were deprecated from earlier versions.
+  // repeated string input = 3;
+  // repeated string output = 4;
+  // optional int64 ir_version = 6;
+  // optional int64 producer_version = 7;
+  // optional string producer_tag = 8;
+  // optional string domain = 9;
+}
 
-  // put build version here
-  optional string producer_version = 6;
+// Tensors
+//
+// A serialized tensor value.
+message TensorProto {
+  enum DataType {
+    UNDEFINED = 0;
+    // Basic types.
+    FLOAT = 1;   // float
+    UINT8 = 2;   // uint8_t
+    INT8 = 3;    // int8_t
+    UINT16 = 4;  // uint16_t
+    INT16 = 5;   // int16_t
+    INT32 = 6;   // int32_t
+    INT64 = 7;   // int64_t
+    STRING = 8;  // string
+    BOOL = 9;    // bool
 
-  optional string name = 7;
+    // Advanced types
+    FLOAT16 = 10;
+    DOUBLE = 11;
+    UINT32 = 12;
+    UINT64 = 13;
+    COMPLEX64 = 14;     // complex with float32 real and imaginary components
+    COMPLEX128 = 15;    // complex with float64 real and imaginary components
+    // Future extensions go here.
 
-  optional string debug_info = 8;
+    // Special data type, real type information is stored in ValueInfoProto.
+    // If data_type is SPECIAL, raw_data should be used.
+    SPECIAL = 51;
+  }
+
+  // The shape of the tensor.
+  repeated int64 dims = 1;
+  repeated int64 strides = 14;
+
+  // The data type of the tensor.
+  optional DataType data_type = 2;
+
+  // For very large tensors, we may want to store them in chunks, in which
+  // case the following fields will specify the segment that is stored in
+  // the current TensorProto.
+  message Segment {
+    optional int64 begin = 1;
+    optional int64 end = 2;
+    optional int64 chuck_num = 51;
+    optional int64 chuck_id = 52;
+  }
+  // Used as offset in the external shared data.
+  optional Segment segment = 3;
+
+  // Tensor content must be organized in row-major order.
+  //
+  // Depending on the data_type field, exactly one of the fields below with
+  // name ending in _data is used to store the elements of the tensor.
+
+  // For float and complex64 values
+  // Complex64 tensors are encoded as a single array of floats,
+  // with the real components appearing in odd numbered positions,
+  // and the corresponding imaginary component apparing in the
+  // subsequent even numbered position. (e.g., [1.0 + 2.0i, 3.0 + 4.0i]
+  // is encoded as [1.0, 2.0 ,3.0 ,4.0]
+  // When this field is present, the data_type field MUST be FLOAT or COMPLEX64.
+  repeated float float_data = 4 [packed = true];
+
+  // For int32, uint8, int8, uint16, int16, bool, and Half values
+  // float16 values must be bit-wise converted to an uint16_t prior
+  // to writing to the buffer.
+  // When this field is present, the data_type field MUST be
+  // INT32, INT16, INT8, UINT16, INT8, BOOL, or FLOAT16
+  repeated int32 int32_data = 5 [packed = true];
+
+  // For strings.
+  // Each element of string_data is a UTF-8 encoded Unicode
+  // string. No trailing null, no leading BOM. The protobuf "string"
+  // scalar type is not used to match ML community conventions.
+  // When this field is present, the data_type field MUST be STRING
+  repeated bytes string_data = 6;
+
+  // For int64.
+  // When this field is present, the data_type field MUST be INT64
+  repeated int64 int64_data = 7 [packed = true];
+
+  // Optionally, a name for the tensor.
+  optional string name = 8; // namespace Value
+
+  // A human-readable documentation for this tensor. Markdown is allowed.
+  optional string doc_string = 12;
+
+  // Serializations can either use one of the fields above, or use this
+  // raw bytes field. The only exception is the string case, where one is
+  // required to store the content in the repeated bytes string_data field.
+  //
+  // When this raw_data field is used to store tensor value, elements MUST
+  // be stored in as fixed-width, little-endian order.
+  // Floating-point data types MUST be stored in IEEE 754 format.
+  // Complex64 elements must be written as two consecutive FLOAT values, real component first.
+  // Complex128 elements must be written as two consecutive DOUBLE values, real component first.
+  // Boolean type MUST be written one byte per tensor element (00000001 for true, 00000000 for false).
+  //
+  // Note: the advantage of specific field rather than the raw_data field is
+  // that in some cases (e.g. int data), protobuf does a better packing via
+  // variable length storage, and may lead to smaller binary footprint.
+  // When this field is present, the data_type field MUST NOT be STRING or UNDEFINED
+  optional bytes raw_data = 9;
+
+  // For double
+  // Complex64 tensors are encoded as a single array of doubles,
+  // with the real components appearing in odd numbered positions,
+  // and the corresponding imaginary component apparing in the
+  // subsequent even numbered position. (e.g., [1.0 + 2.0i, 3.0 + 4.0i]
+  // is encoded as [1.0, 2.0 ,3.0 ,4.0]
+  // When this field is present, the data_type field MUST be DOUBLE or COMPLEX128
+  repeated double double_data = 10 [packed = true];
+
+  // For uint64 and uint32 values
+  // When this field is present, the data_type field MUST be
+  // UINT32 or UINT64
+  repeated uint64 uint64_data = 11 [packed = true];
+
+  // External data by file name
+  optional string external_data = 13;
+
+  // If two tensors represent the same weights/content, use alias.
+  // Must exist a TensorProto named alias in the initializer list.
+  // To avoid the duplicate tensor in attribute, such as value in Constant node.
+  // This is useful, if everything is stored just in the proto.
+  optional string alias = 16;
+
+  // Additional annotations.
+  repeated AttributeProto annotations = 17;
+
+  // Device info
+  optional caffe2.DeviceOption device_option = 51;
+
+  // For PyTorch serialized tensor.
+  optional int64 require_gradient = 52;
+  optional int64 is_buffer = 53;
+}
+
+// Defines a tensor shape. A dimension can be either an integer value
+// or a symbolic variable. A symbolic variable represents an unknown
+// dimension.
+message TensorShapeProto {
+  message Dimension {
+    oneof value {
+      int64 dim_value = 1;
+      string dim_param = 2;   // namespace Shape
+    };
+    // Standard denotation can optionally be used to denote tensor
+    // dimensions with standard semantic descriptions to ensure
+    // that operations are applied to the correct axis of a tensor.
+    // Refer to https://github.com/onnx/onnx/blob/master/docs/DimensionDenotation.md#denotation-definition
+    // for pre-defined dimension denotations.
+    optional string denotation = 3;
+  };
+  // To represent a scalar, using no dim to represent 0-d tensor.
+  repeated Dimension dim = 1;
+
+  repeated Dimension stride = 51;
+}
+
+// Types
+//
+// The standard ONNX data types.
+message TypeProto {
+
+  message Tensor {
+    // This field MUST NOT have the value of UNDEFINED
+    // This field MUST be present for this version of the IR.
+    optional TensorProto.DataType elem_type = 1;
+    optional TensorShapeProto shape = 2;
+  }
+
+  // Sequence type: List, Tuple
+  message Sequence {
+    // elem_type and elem_type_list cannot appear together.
+    // If all the element types are the same, we use elem_type,
+    // otherwise, we specify the type of each element in elem_type_list.
+    optional TypeProto elem_type = 1;
+    repeated TypeProto elem_type_list = 51;
+    enum SequenceType {
+      UNDEFINED = 0;
+      LIST = 1;
+      TUPLE = 2;
+    }
+    optional SequenceType sequence_type = 52;
+  }
+
+  // Map<K, V>, (not necessary at this moment)
+  message Map {
+    optional TensorProto.DataType key_type = 1;
+    optional TypeProto value_type = 2;
+  }
+
+  // Special type of blobs, based on the type_name, we can choose the right
+  // serializer and deserialzier.
+  message SpecialBlob {
+    optional string type_name = 1;
+  }
+
+  oneof value {
+    // The type of a tensor.
+    Tensor tensor_type = 1;
+    Sequence sequence_type = 4;
+    Map map_type = 5;
+    SpecialBlob special_type = 51;
+  }
+
+  // An optional denotation can be used to denote the whole
+  // type with a standard semantic description as to what is
+  // stored inside. Refer to https://github.com/onnx/onnx/blob/master/docs/TypeDenotation.md#type-denotation-definition
+  // for pre-defined type denotations.
+  optional string denotation = 6;
+}
 
-  // annotations, it is used for MetaNetDef's metadata
-  repeated caffe2.Argument annotations = 9;
+// Operator Sets
+//
+// OperatorSets are uniquely identified by a (domain, opset_version) pair.
+message OperatorSetIdProto {
+  // The domain of the operator set being identified.
+  // The empty string ("") or absence of this field implies the operator
+  // set that is defined as part of the ONNX specification.
+  // This field MUST be present in this version of the IR when referring to any other operator set.
+  optional string domain = 1;
 
+  // The version of the operator set being identified.
+  // This field MUST be present in this version of the IR.
+  optional int64 version = 2;
 }
diff --git a/caffe2/python/convert.py b/caffe2/python/convert.py
index 44f81d6e2d..50eaf220c7 100644
--- a/caffe2/python/convert.py
+++ b/caffe2/python/convert.py
@@ -8,3 +8,59 @@ from __future__ import unicode_literals
 from caffe2.proto import caffe2_pb2, torch_pb2
 
 import caffe2.python._import_c_extension as C
+
+
+def ArgumentToAttributeProto(arg):
+    serialized_arg = None
+    if hasattr(arg, 'SerializeToString') and callable(arg.SerializeToString):
+        serialized_arg = arg.SerializeToString()
+    elif isinstance(arg, bytes):
+        serialized_arg = arg
+    else:
+        raise ValueError('No SerializeToString method is detected. '
+                         'neither arg is bytes.\ntype is {}'.format(type(arg)))
+    attr = torch_pb2.AttributeProto()
+    attr.ParseFromString(C.argument_to_attribute_proto(serialized_arg))
+    return attr
+
+
+def AttributeProtoToArgument(attr):
+    serialized_attr = None
+    if hasattr(attr, 'SerializeToString') and callable(attr.SerializeToString):
+        serialized_attr = attr.SerializeToString()
+    elif isinstance(attr, bytes):
+        serialized_attr = attr
+    else:
+        raise ValueError('No SerializeToString method is detected. '
+                         'neither attr is bytes.\ntype is {}'.format(type(attr)))
+    arg = caffe2_pb2.Argument()
+    arg.ParseFromString(C.attribute_proto_to_argument(serialized_attr))
+    return arg
+
+
+def OperatorDefToNodeProto(op_def):
+    serialized_op_def = None
+    if hasattr(op_def, 'SerializeToString') and callable(op_def.SerializeToString):
+        serialized_op_def = op_def.SerializeToString()
+    elif isinstance(op_def, bytes):
+        serialized_op_def = op_def
+    else:
+        raise ValueError('No SerializeToString method is detected. '
+                         'neither op_def is bytes.\ntype is {}'.format(type(op_def)))
+    node = torch_pb2.NodeProto()
+    node.ParseFromString(C.operator_def_to_node_proto(serialized_op_def))
+    return node
+
+
+def NodeProtoToOperatorDef(node_proto):
+    serialized_node_proto = None
+    if hasattr(node_proto, 'SerializeToString') and callable(node_proto.SerializeToString):
+        serialized_node_proto = node_proto.SerializeToString()
+    elif isinstance(node_proto, bytes):
+        serialized_node_proto = node_proto
+    else:
+        raise ValueError('No SerializeToString method is detected. '
+                         'neither node_proto is bytes.\ntype is {}'.format(type(node_proto)))
+    op_def = caffe2_pb2.OperatorDef()
+    op_def.ParseFromString(C.node_proto_to_operator_def(serialized_node_proto))
+    return op_def
diff --git a/caffe2/python/convert_test.py b/caffe2/python/convert_test.py
index 82c969c901..c8de7e9750 100644
--- a/caffe2/python/convert_test.py
+++ b/caffe2/python/convert_test.py
@@ -12,5 +12,239 @@ class TestOperator(unittest.TestCase):
     def setUp(self):
         workspace.ResetWorkspace()
 
+    def testArgument2AttributeProto(self):
+        arg_f = caffe2_pb2.Argument()
+        arg_f.name = "TestArgF"
+        arg_f.f = 10.0
+        attr_f = convert.ArgumentToAttributeProto(arg_f)
+        self.assertEqual(attr_f.name, arg_f.name)
+        self.assertEqual(attr_f.f, arg_f.f)
+
+        arg_i = caffe2_pb2.Argument()
+        arg_i.name = "TestArgI"
+        arg_i.i = 100
+        attr_i = convert.ArgumentToAttributeProto(arg_i)
+        self.assertEqual(attr_i.name, arg_i.name)
+        self.assertEqual(attr_i.i, arg_i.i)
+
+        arg_s = caffe2_pb2.Argument()
+        arg_s.name = "TestArgS"
+        arg_s.s = "TestS".encode("utf-8")
+        attr_s = convert.ArgumentToAttributeProto(arg_s)
+        self.assertEqual(attr_s.name, arg_s.name)
+        self.assertEqual(attr_s.s, arg_s.s)
+
+        # TODO: test net arg
+
+        arg_floats = caffe2_pb2.Argument()
+        arg_floats.name = "TestArgFloats"
+        arg_floats.floats.extend([10.0, 11.0, 12.0])
+        attr_floats = convert.ArgumentToAttributeProto(arg_floats)
+        self.assertEqual(attr_floats.name, arg_floats.name)
+        self.assertEqual(attr_floats.floats, arg_floats.floats)
+
+        arg_ints = caffe2_pb2.Argument()
+        arg_ints.name = "TestArgInts"
+        arg_ints.ints.extend([100, 101, 102])
+        attr_ints = convert.ArgumentToAttributeProto(arg_ints)
+        self.assertEqual(attr_ints.name, arg_ints.name)
+        self.assertEqual(attr_ints.ints, arg_ints.ints)
+
+        arg_strings = caffe2_pb2.Argument()
+        arg_strings.name = "TestArgStrings"
+        arg_strings.strings.extend([
+            "TestStrings1".encode("utf-8"),
+            "TestStrings2".encode("utf-8"),
+        ])
+        attr_strings = convert.ArgumentToAttributeProto(arg_strings)
+        self.assertEqual(attr_strings.name, arg_strings.name)
+        self.assertEqual(attr_strings.strings, arg_strings.strings)
+
+        # TODO: test nets arg
+
+    def testAttributeProto2Argument(self):
+        attr_f = torch_pb2.AttributeProto()
+        attr_f.type = torch_pb2.AttributeProto.FLOAT
+        attr_f.name = "TestAttrF"
+        attr_f.f = 10.0
+        arg_f = convert.AttributeProtoToArgument(attr_f)
+        self.assertEqual(arg_f.name, attr_f.name)
+        self.assertEqual(arg_f.f, attr_f.f)
+
+        attr_i = torch_pb2.AttributeProto()
+        attr_i.type = torch_pb2.AttributeProto.INT
+        attr_i.name = "TestArgI"
+        attr_i.i = 100
+        arg_i = convert.AttributeProtoToArgument(attr_i)
+        self.assertEqual(arg_i.name, attr_i.name)
+        self.assertEqual(arg_i.i, attr_i.i)
+
+        attr_s = torch_pb2.AttributeProto()
+        attr_s.type = torch_pb2.AttributeProto.STRING
+        attr_s.name = "TestArgS"
+        attr_s.s = "TestS".encode("utf-8")
+        arg_s = convert.AttributeProtoToArgument(attr_s)
+        self.assertEqual(arg_s.name, attr_s.name)
+        self.assertEqual(arg_s.s, attr_s.s)
+
+        # TODO: test graph attribute
+
+        attr_floats = torch_pb2.AttributeProto()
+        attr_floats.type = torch_pb2.AttributeProto.FLOATS
+        attr_floats.name = "TestAttrFloats"
+        attr_floats.floats.extend([10.0, 11.0, 12.0])
+        arg_floats = convert.AttributeProtoToArgument(attr_floats)
+        self.assertEqual(arg_floats.name, attr_floats.name)
+        self.assertEqual(arg_floats.floats, attr_floats.floats)
+
+        attr_ints = torch_pb2.AttributeProto()
+        attr_ints.type = torch_pb2.AttributeProto.INTS
+        attr_ints.name = "TestArgInts"
+        attr_ints.ints.extend([100, 101, 102])
+        arg_ints = convert.AttributeProtoToArgument(attr_ints)
+        self.assertEqual(arg_ints.name, attr_ints.name)
+        self.assertEqual(arg_ints.ints, attr_ints.ints)
+
+        attr_strings = torch_pb2.AttributeProto()
+        attr_strings.type = torch_pb2.AttributeProto.STRINGS
+        attr_strings.name = "TestArgStrings"
+        attr_strings.strings.extend([
+            "TestStrings1".encode("utf-8"),
+            "TestStrings2".encode("utf-8"),
+        ])
+        arg_strings = convert.AttributeProtoToArgument(attr_strings)
+        self.assertEqual(arg_strings.name, attr_strings.name)
+        self.assertEqual(arg_strings.strings, attr_strings.strings)
+
+        # TODO: test graphs attribute
+
+
+    def testOperatorDef2NodeProto(self):
+        op_def = caffe2_pb2.OperatorDef()
+        op_def.input.extend(["A", "B", "C"])
+        op_def.output.extend(["X", "Y"])
+        op_def.name = "TestOpName"
+        op_def.type = "TestOp"
+        arg1 = caffe2_pb2.Argument()
+        arg1.name = "TestArg1"
+        arg1.i = 1
+        arg2 = caffe2_pb2.Argument()
+        arg2.name = "TestArg2"
+        arg1.s = "TestInfo".encode("utf-8")
+        op_def.arg.extend([arg1, arg2])
+        op_def.device_option.CopyFrom(caffe2_pb2.DeviceOption())
+        op_def.engine = "TestEngine".encode("utf-8")
+        op_def.control_input.extend(["input1", "input2"])
+        op_def.is_gradient_op = True
+        op_def.debug_info = "TestDebugInfo"
+
+        node = convert.OperatorDefToNodeProto(op_def)
+
+        self.assertEqual(node.input, op_def.input)
+        self.assertEqual(node.output, op_def.output)
+        self.assertEqual(node.name, op_def.name)
+        self.assertEqual(node.op_type, op_def.type)
+        self.assertEqual(node.attribute[0].name, op_def.arg[0].name)
+        self.assertEqual(node.attribute[1].name, op_def.arg[1].name)
+        self.assertEqual(node.device_option, op_def.device_option)
+        node_engine = [a.s.decode("utf-8") for a in node.annotations if a.name == "engine"][0]
+        self.assertEqual(node_engine, op_def.engine)
+        node_control_input = [a.strings for a in node.annotations if a.name == "control_input"][0]
+        self.assertEqual(len(node_control_input), len(op_def.control_input))
+        for x, y in zip(node_control_input, op_def.control_input):
+            self.assertEqual(x.decode("utf-8"), y)
+        self.assertEqual(node.doc_string, op_def.debug_info)
+        node_is_gradient_op = [a.i for a in node.annotations if a.name == "is_gradient_op"][0]
+        self.assertEqual(node_is_gradient_op, int(op_def.is_gradient_op))
+
+    def testNodeProto2OperatorDef(self):
+        node = torch_pb2.NodeProto()
+        node.input.extend(["A", "B", "C"])
+        node.output.extend(["X", "Y"])
+        node.name = "TestOpName"
+        node.op_type = "TestOp"
+        attr1 = torch_pb2.AttributeProto()
+        attr1.name = "TestAttr1"
+        attr1.type = torch_pb2.AttributeProto.STRING
+        attr1.s = "TestInfo".encode("utf-8")
+        attr2 = torch_pb2.AttributeProto()
+        attr2.name = "TestAttr2"
+        attr2.type = torch_pb2.AttributeProto.INT
+        attr2.i = 10
+        node.attribute.extend([attr1, attr2])
+        node.device_option.CopyFrom(caffe2_pb2.DeviceOption())
+        anno1 = torch_pb2.AttributeProto()
+        anno1.name = "engine"
+        anno1.type = torch_pb2.AttributeProto.STRING
+        anno1.s = "TestEngine".encode("utf-8")
+        anno2 = torch_pb2.AttributeProto()
+        anno2.name = "control_input"
+        anno2.type = torch_pb2.AttributeProto.STRINGS
+        anno2.strings.extend(["input1".encode("utf-8"), "input2".encode("utf-8")])
+        anno3 = torch_pb2.AttributeProto()
+        anno3.name = "is_gradient_op"
+        anno3.type = torch_pb2.AttributeProto.INT
+        anno3.i = 1
+        node.annotations.extend([anno1, anno2, anno3])
+        node.doc_string = "TestDocString".encode("utf-8")
+
+        op_def = convert.NodeProtoToOperatorDef(node)
+
+        self.assertEqual(op_def.input, node.input)
+        self.assertEqual(op_def.output, node.output)
+        self.assertEqual(op_def.name, node.name)
+        self.assertEqual(op_def.type, node.op_type)
+        self.assertEqual(op_def.arg[0].name, node.attribute[0].name)
+        self.assertEqual(op_def.arg[1].name, node.attribute[1].name)
+        self.assertEqual(op_def.device_option, node.device_option)
+        node_engine = [a.s for a in node.annotations if a.name == "engine"][0]
+        self.assertEqual(op_def.engine, node_engine.decode("utf-8"))
+        node_control_input = [a.strings for a in node.annotations if a.name == "control_input"][0]
+        for x, y in zip(op_def.control_input, node_control_input):
+            self.assertEqual(x, y.decode("utf-8"))
+        self.assertEqual(op_def.debug_info, node.doc_string)
+        node_is_gradient_op = [a.i for a in node.annotations if a.name == "is_gradient_op"][0]
+        self.assertEqual(int(op_def.is_gradient_op), node_is_gradient_op)
+
+    def testEnd2End(self):
+        op_def = caffe2_pb2.OperatorDef()
+        op_def.type = "Add"
+        op_def.input.extend(["input1"])
+        op_def.input.extend(["input2"])
+        op_def.output.extend(["output1"])
+        node = convert.OperatorDefToNodeProto(op_def)
+
+        input1 = np.random.randn(1, 3, 1, 5).astype(np.float32)
+        input2 = np.random.randn(2, 1, 4, 1).astype(np.float32)
+        ref_output1 = input1 + input2
+        workspace.FeedBlob("input1", input1)
+        workspace.FeedBlob("input2", input2)
+        self.assertEqual(workspace.RunOperatorOnce(node.SerializeToString(), legacy_proto=False), True)
+
+        self.assertEqual(workspace.HasBlob("output1"), True)
+        fetched_back = workspace.FetchBlob("output1")
+        np.testing.assert_array_equal(fetched_back, ref_output1)
+
+    def testRoundTrip(self):
+        op_def = caffe2_pb2.OperatorDef()
+        op_def.type = "Add"
+        op_def.input.extend(["input1"])
+        op_def.input.extend(["input2"])
+        op_def.output.extend(["output1"])
+        node = convert.OperatorDefToNodeProto(op_def)
+        new_op_def = convert.NodeProtoToOperatorDef(node)
+
+        input1 = np.random.randn(1, 3, 1, 5).astype(np.float32)
+        input2 = np.random.randn(2, 1, 4, 1).astype(np.float32)
+        ref_output1 = input1 + input2
+        workspace.FeedBlob("input1", input1)
+        workspace.FeedBlob("input2", input2)
+        self.assertEqual(workspace.RunOperatorOnce(new_op_def.SerializeToString()), True)
+
+        self.assertEqual(workspace.HasBlob("output1"), True)
+        fetched_back = workspace.FetchBlob("output1")
+        np.testing.assert_array_equal(fetched_back, ref_output1)
+
+
 if __name__ == '__main__':
     unittest.main()
diff --git a/caffe2/python/pybind_state.cc b/caffe2/python/pybind_state.cc
index 7ebee57d49..7062ead045 100644
--- a/caffe2/python/pybind_state.cc
+++ b/caffe2/python/pybind_state.cc
@@ -1187,10 +1187,17 @@ void addGlobalMethods(py::module& m) {
     return true;
   });
   m.def("nets", []() { return gWorkspace->Nets(); });
-  m.def("run_operator_once", [](const py::bytes& op_def) {
+  m.def("run_operator_once", [](const py::bytes& op_def, bool legacy_proto=true) {
     CAFFE_ENFORCE(gWorkspace);
     OperatorDef def;
-    CAFFE_ENFORCE(ParseProtoFromLargeString(op_def.cast<std::string>(), &def));
+    if (legacy_proto) {
+      CAFFE_ENFORCE(ParseProtoFromLargeString(op_def.cast<std::string>(), &def));
+    } else {
+      ::torch::NodeProto node;
+      CAFFE_ENFORCE(
+          ParseProtoFromLargeString(op_def.cast<std::string>(), &node));
+      NodeProtoToOperatorDef(node, &def);
+    }
     py::gil_scoped_release g;
     CAFFE_ENFORCE(gWorkspace->RunOperatorOnce(def));
     return true;
@@ -1527,6 +1534,38 @@ void addGlobalMethods(py::module& m) {
     CAFFE_ENFORCE(blob);
     return BlobStat::sizeBytes(*blob);
   });
+  m.def("argument_to_attribute_proto", [](py::bytes arg_str) -> py::bytes {
+    Argument arg;
+    CAFFE_ENFORCE(
+      ParseProtoFromLargeString(arg_str.cast<std::string>(), &arg));
+    ::torch::AttributeProto attr;
+    ArgumentToAttributeProto(arg, &attr);
+    return attr.SerializeAsString();
+  });
+  m.def("attribute_proto_to_argument", [](py::bytes attr_str) -> py::bytes {
+    ::torch::AttributeProto attr;
+    CAFFE_ENFORCE(
+      ParseProtoFromLargeString(attr_str.cast<std::string>(), &attr));
+    Argument arg;
+    AttributeProtoToArgument(attr, &arg);
+    return arg.SerializeAsString();
+  });
+  m.def("operator_def_to_node_proto", [](py::bytes op_str) -> py::bytes {
+    OperatorDef op_def;
+    CAFFE_ENFORCE(
+      ParseProtoFromLargeString(op_str.cast<std::string>(), &op_def));
+    ::torch::NodeProto node;
+    OperatorDefToNodeProto(op_def, &node);
+    return node.SerializeAsString();
+  });
+  m.def("node_proto_to_operator_def", [](py::bytes node_str) -> py::bytes {
+    ::torch::NodeProto node_proto;
+    CAFFE_ENFORCE(
+      ParseProtoFromLargeString(node_str.cast<std::string>(), &node_proto));
+    OperatorDef op_def;
+    NodeProtoToOperatorDef(node_proto, &op_def);
+    return op_def.SerializeAsString();
+  });
   m.def("support_onnx_export", [](const std::string& op) -> bool {
     const OpSchema* schema = caffe2::OpSchemaRegistry::Schema(op);
     if (!schema) {
diff --git a/caffe2/python/workspace.py b/caffe2/python/workspace.py
index ef02f64dc9..a41cc15317 100644
--- a/caffe2/python/workspace.py
+++ b/caffe2/python/workspace.py
@@ -163,8 +163,8 @@ def GetOperatorCost(operator, blobs):
     return C.get_operator_cost(StringifyProto(operator), blobs)
 
 
-def RunOperatorOnce(operator):
-    return C.run_operator_once(StringifyProto(operator))
+def RunOperatorOnce(operator, legacy_proto=True):
+    return C.run_operator_once(StringifyProto(operator), legacy_proto)
 
 
 def RunOperatorsOnce(operators):
diff --git a/caffe2/utils/proto_convert.cc b/caffe2/utils/proto_convert.cc
index 1d69c8c80c..790bd27429 100644
--- a/caffe2/utils/proto_convert.cc
+++ b/caffe2/utils/proto_convert.cc
@@ -2,4 +2,185 @@
 #include "caffe2/core/logging.h"
 
 namespace caffe2 {
+
+C10_EXPORT void ArgumentToAttributeProto(
+    const Argument& arg,
+    ::torch::AttributeProto* attr) {
+  CAFFE_ENFORCE(arg.has_name());
+  attr->set_name(arg.name());
+  if (arg.has_f()) {
+    attr->set_f(arg.f());
+  } else if (arg.has_i()) {
+    attr->set_i(arg.i());
+  } else if (arg.has_s()) {
+    attr->set_s(arg.s());
+  } else if (arg.has_n()) {
+    // TODO
+    CAFFE_THROW("NetDef conversion is not implemented yet.");
+  } else if (arg.floats_size() > 0) {
+    attr->mutable_floats()->CopyFrom(arg.floats());
+  } else if (arg.ints_size() > 0) {
+    attr->mutable_ints()->CopyFrom(arg.ints());
+  } else if (arg.strings_size() > 0) {
+    attr->mutable_strings()->CopyFrom(arg.strings());
+  } else if (arg.nets_size() > 0) {
+    // TODO
+    CAFFE_THROW("NetDefs conversion is not implemented yet.");
+  }
+}
+
+C10_EXPORT void AttributeProtoToArgument(
+    const ::torch::AttributeProto& attr,
+    Argument* arg) {
+  CAFFE_ENFORCE(attr.has_name());
+  arg->set_name(attr.name());
+  CAFFE_ENFORCE(attr.has_type());
+  const auto type = attr.type();
+  if (type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_FLOAT) {
+    CAFFE_ENFORCE(attr.has_f());
+    arg->set_f(attr.f());
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::AttributeProto_AttributeType_INT) {
+    CAFFE_ENFORCE(attr.has_i());
+    arg->set_i(attr.i());
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_STRING) {
+    CAFFE_ENFORCE(attr.has_s());
+    arg->set_s(attr.s());
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_TENSOR) {
+    CAFFE_THROW("Caffe2's Argument does not support tensor as attribute.");
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_GRAPH) {
+    // TODO
+    CAFFE_THROW("GraphProto conversion is not implemented yet.");
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_FLOATS) {
+    arg->mutable_floats()->CopyFrom(attr.floats());
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_INTS) {
+    arg->mutable_ints()->CopyFrom(attr.ints());
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_STRINGS) {
+    arg->mutable_strings()->CopyFrom(attr.strings());
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_TENSORS) {
+    CAFFE_THROW("Caffe2's Argument does not support tensors as attribute.");
+  } else if (
+      type ==
+      ::torch::AttributeProto_AttributeType::
+          AttributeProto_AttributeType_GRAPHS) {
+    // TODO
+    CAFFE_THROW("GraphProtos conversion is not implemented yet.");
+  } else {
+    CAFFE_THROW("Unknow Attribute type.");
+  }
+}
+
+C10_EXPORT void OperatorDefToNodeProto(
+    const OperatorDef& def,
+    ::torch::NodeProto* node) {
+  node->mutable_input()->CopyFrom(def.input());
+  node->mutable_output()->CopyFrom(def.output());
+  if (def.has_name()) {
+    node->set_name(def.name());
+  }
+  CAFFE_ENFORCE(def.has_type());
+  node->set_op_type(def.type());
+  for (int i = 0; i < def.arg_size(); ++i) {
+    auto attr = node->add_attribute();
+    ArgumentToAttributeProto(def.arg(i), attr);
+  }
+  if (def.has_device_option()) {
+    node->mutable_device_option()->CopyFrom(def.device_option());
+  }
+  if (def.has_engine()) {
+    auto attr = node->add_annotations();
+    attr->set_name("engine");
+    attr->set_type(::torch::AttributeProto_AttributeType::
+                       AttributeProto_AttributeType_STRING);
+    attr->set_s(def.engine());
+  }
+  if (def.control_input_size() > 0) {
+    auto attr = node->add_annotations();
+    attr->set_name("control_input");
+    attr->set_type(::torch::AttributeProto_AttributeType::
+                       AttributeProto_AttributeType_STRINGS);
+    attr->mutable_strings()->CopyFrom(def.control_input());
+  }
+  if (def.has_is_gradient_op()) {
+    auto attr = node->add_annotations();
+    attr->set_name("is_gradient_op");
+    attr->set_type(::torch::AttributeProto_AttributeType::
+                       AttributeProto_AttributeType_INT);
+    if (def.is_gradient_op()) {
+      attr->set_i(1);
+    } else {
+      attr->set_i(0);
+    }
+  }
+  if (def.has_debug_info()) {
+    node->set_doc_string(def.debug_info());
+  }
+}
+
+C10_EXPORT void NodeProtoToOperatorDef(
+    const ::torch::NodeProto& node,
+    OperatorDef* def) {
+  def->mutable_input()->CopyFrom(node.input());
+  def->mutable_output()->CopyFrom(node.output());
+  if (node.has_name()) {
+    def->set_name(node.name());
+  }
+
+  CAFFE_ENFORCE(node.has_op_type());
+  def->set_type(node.op_type());
+  for (int i = 0; i < node.attribute_size(); ++i) {
+    auto arg = def->add_arg();
+    AttributeProtoToArgument(node.attribute(i), arg);
+  }
+  if (node.has_doc_string()) {
+    def->set_debug_info(node.doc_string());
+  }
+  for (int i = 0; i < node.annotations_size(); ++i) {
+    const auto& attr = node.annotations(i);
+    CAFFE_ENFORCE(attr.has_name());
+    if (attr.name() == "engine") {
+      CAFFE_ENFORCE(attr.has_s());
+      def->set_engine(attr.s());
+    } else if (attr.name() == "control_input") {
+      def->mutable_control_input()->CopyFrom(attr.strings());
+    } else if (attr.name() == "is_gradient_op") {
+      CAFFE_ENFORCE(attr.has_i());
+      if (i == 0) {
+        def->set_is_gradient_op(false);
+      } else {
+        def->set_is_gradient_op(true);
+      }
+    }
+    auto arg = def->add_arg();
+    AttributeProtoToArgument(node.annotations(i), arg);
+  }
+  if (node.has_device_option()) {
+    def->mutable_device_option()->CopyFrom(node.device_option());
+  }
+}
+
 } // namespace caffe2
diff --git a/caffe2/utils/proto_convert.h b/caffe2/utils/proto_convert.h
index 91bcf1bafa..a9ca9c3ad4 100644
--- a/caffe2/utils/proto_convert.h
+++ b/caffe2/utils/proto_convert.h
@@ -6,6 +6,20 @@
 #include "caffe2/proto/torch_pb.h"
 
 namespace caffe2 {
+
+CAFFE2_API void ArgumentToAttributeProto(
+    const Argument& arg,
+    ::torch::AttributeProto* attr);
+CAFFE2_API void AttributeProtoToArgument(
+    const ::torch::AttributeProto& attr,
+    Argument* arg);
+CAFFE2_API void OperatorDefToNodeProto(
+    const OperatorDef& def,
+    ::torch::NodeProto* node);
+CAFFE2_API void NodeProtoToOperatorDef(
+    const ::torch::NodeProto& node,
+    OperatorDef* def);
+
 } // namespace caffe2
 
 #endif // CAFFE2_UTILS_PROTO_CONVERT_H_