Add SignatureDefs to Go SavedModel API.

PiperOrigin-RevId: 281984017
Change-Id: Iefdf75ed88f54d97a0a7d210f5a42f3123205bf2
This commit is contained in:
Phil Stahlfeld 2019-11-22 09:32:02 -08:00 committed by TensorFlower Gardener
parent f0323528be
commit f1a3c8af3c
6 changed files with 356 additions and 4 deletions

View File

@ -24,6 +24,8 @@ sh_test(
"//tensorflow/c/eager:headers", # Eager C library header
"//tensorflow/cc/saved_model:saved_model_half_plus_two", # Testdata for LoadSavedModel
] + tf_shared_library_deps(),
# TODO: Enable this test again once protos are supported by bazel.
tags = ["manual"],
)
filegroup(

View File

@ -30,6 +30,7 @@ from source.
sudo apt-get install python swig python-numpy # Linux
brew install swig # OS X with homebrew
```
- [Protocol buffer compiler (protoc) 3.x](https://github.com/google/protobuf/releases/)
### Build
@ -74,6 +75,7 @@ from source.
4. Build and test:
```sh
go generate github.com/tensorflow/tensorflow/tensorflow/go/op
go test github.com/tensorflow/tensorflow/tensorflow/go
```

View File

@ -19,6 +19,10 @@ package tensorflow
import (
"runtime"
"unsafe"
"github.com/golang/protobuf/proto"
tfpb "github.com/tensorflow/tensorflow/tensorflow/go/genop/internal/proto/github.com/tensorflow/tensorflow/tensorflow/go/core/framework"
)
// #include <stdlib.h>
@ -28,8 +32,9 @@ import "C"
// SavedModel represents the contents of loaded SavedModel.
// TODO(jhseu): Add and document metagraphdef when we pregenerate protobufs.
type SavedModel struct {
Session *Session
Graph *Graph
Session *Session
Graph *Graph
Signatures map[string]Signature
}
// LoadSavedModel creates a new SavedModel from a model previously
@ -58,17 +63,35 @@ func LoadSavedModel(exportDir string, tags []string, options *SessionOptions) (*
cTags[i] = C.CString(tags[i])
}
graph := NewGraph()
metaGraphDefBuf := C.TF_NewBuffer()
defer C.TF_DeleteBuffer(metaGraphDefBuf)
// TODO(jhseu): Add support for run_options and meta_graph_def.
cSess := C.TF_LoadSessionFromSavedModel(cOpt, nil, cExportDir, (**C.char)(unsafe.Pointer(&cTags[0])), C.int(len(cTags)), graph.c, nil, status.c)
cSess := C.TF_LoadSessionFromSavedModel(cOpt, nil, cExportDir, (**C.char)(unsafe.Pointer(&cTags[0])), C.int(len(cTags)), graph.c, metaGraphDefBuf, status.c)
for i := range cTags {
C.free(unsafe.Pointer(cTags[i]))
}
C.free(unsafe.Pointer(cExportDir))
metaGraphDefBytes := C.GoBytes(metaGraphDefBuf.data, C.int(metaGraphDefBuf.length))
metaGraphDef := new(tfpb.MetaGraphDef)
if err := proto.Unmarshal(metaGraphDefBytes, metaGraphDef); err != nil {
return nil, err
}
signatures := generateSignatures(metaGraphDef.GetSignatureDef())
if err := status.Err(); err != nil {
return nil, err
}
s := &Session{c: cSess}
runtime.SetFinalizer(s, func(s *Session) { s.Close() })
return &SavedModel{Session: s, Graph: graph}, nil
return &SavedModel{Session: s, Graph: graph, Signatures: signatures}, nil
}
func generateSignatures(pb map[string]*tfpb.SignatureDef) map[string]Signature {
signatures := make(map[string]Signature)
for name, signature := range pb {
signatures[name] = signatureDefFromProto(signature)
}
return signatures
}

View File

@ -26,6 +26,7 @@ func TestSavedModel(t *testing.T) {
if op := bundle.Graph.Operation("y"); op == nil {
t.Fatalf("\"y\" not found in graph")
}
t.Logf("SavedModel: %+v", bundle)
// TODO(jhseu): half_plus_two has a tf.Example proto dependency to run. Add a
// more thorough test when the generated protobufs are available.
}

119
tensorflow/go/signature.go Normal file
View File

@ -0,0 +1,119 @@
/*
Copyright 2019 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package tensorflow
import tfpb "github.com/tensorflow/tensorflow/tensorflow/go/genop/internal/proto/github.com/tensorflow/tensorflow/tensorflow/go/core/framework"
// #include "tensorflow/c/c_api.h"
import "C"
// A Signature defines the signature of a computation supported by a TensorFlow
// graph.
//
// For example, a model with two loss computations, sharing a single input,
// might have the following signature_def map.
//
// Note that across the two Signatures "loss_A" and "loss_B", the input key,
// output key, and method_name are identical, and will be used by system(s) that
// implement or rely upon this particular loss method. The output tensor names
// differ, demonstrating how different outputs can exist for the same method.
//
// signature_def {
// key: "loss_A"
// value {
// inputs {
// key: "input"
// value {
// name: "input:0"
// dtype: DT_STRING
// tensor_shape: ...
// }
// }
// outputs {
// key: "loss_output"
// value {
// name: "loss_output_A:0"
// dtype: DT_FLOAT
// tensor_shape: ...
// }
// }
// }
// ...
// method_name: "some/package/compute_loss"
// }
// signature_def {
// key: "loss_B"
// value {
// inputs {
// key: "input"
// value {
// name: "input:0"
// dtype: DT_STRING
// tensor_shape: ...
// }
// }
// outputs {
// key: "loss_output"
// value {
// name: "loss_output_B:0"
// dtype: DT_FLOAT
// tensor_shape: ...
// }
// }
// }
// ...
// method_name: "some/package/compute_loss"
// }
type Signature struct {
Inputs, Outputs map[string]TensorInfo
MethodName string
}
// A TensorInfo contains the information about a Tensor necessary for feeding or retrieval.
type TensorInfo struct {
Name string
DType DataType
Shape Shape
}
func signatureDefFromProto(pb *tfpb.SignatureDef) Signature {
inputs := make(map[string]TensorInfo)
for name, input := range pb.GetInputs() {
inputs[name] = tensorInfoFromProto(input)
}
outputs := make(map[string]TensorInfo)
for name, output := range pb.GetOutputs() {
outputs[name] = tensorInfoFromProto(output)
}
return Signature{
Inputs: inputs,
Outputs: outputs,
MethodName: pb.GetMethodName(),
}
}
func tensorInfoFromProto(pb *tfpb.TensorInfo) TensorInfo {
var dims []int64
for _, d := range pb.GetTensorShape().GetDim() {
dims = append(dims, d.GetSize())
}
return TensorInfo{
Name: pb.GetName(),
DType: DataType(C.TF_DataType(pb.GetDtype())),
Shape: MakeShape(dims...),
}
}

View File

@ -0,0 +1,205 @@
/*
Copyright 2019 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package tensorflow
import (
"fmt"
"testing"
tfpb "github.com/tensorflow/tensorflow/tensorflow/go/genop/internal/proto/github.com/tensorflow/tensorflow/tensorflow/go/core/framework"
)
func TestSignatureFromProto(t *testing.T) {
got := signatureDefFromProto(&tfpb.SignatureDef{
Inputs: map[string]*tfpb.TensorInfo{
"input_1": &tfpb.TensorInfo{
Encoding: &tfpb.TensorInfo_Name{
Name: "tensor_1",
},
Dtype: tfpb.DataType_DT_INT8,
TensorShape: &tfpb.TensorShapeProto{
Dim: []*tfpb.TensorShapeProto_Dim{
{Size: 1},
{Size: 2},
{Size: 3},
},
},
},
"input_2": &tfpb.TensorInfo{
Encoding: &tfpb.TensorInfo_Name{
Name: "tensor_2",
},
Dtype: tfpb.DataType_DT_FLOAT,
TensorShape: &tfpb.TensorShapeProto{
Dim: []*tfpb.TensorShapeProto_Dim{
{Size: 4},
{Size: 5},
{Size: 6},
},
},
},
},
Outputs: map[string]*tfpb.TensorInfo{
"output_1": &tfpb.TensorInfo{
Encoding: &tfpb.TensorInfo_Name{
Name: "tensor_3",
},
Dtype: tfpb.DataType_DT_STRING,
TensorShape: &tfpb.TensorShapeProto{
Dim: []*tfpb.TensorShapeProto_Dim{
{Size: 1},
{Size: 2},
{Size: 3},
},
},
},
"output_2": &tfpb.TensorInfo{
Encoding: &tfpb.TensorInfo_Name{
Name: "tensor_4",
},
Dtype: tfpb.DataType_DT_BOOL,
TensorShape: &tfpb.TensorShapeProto{
Dim: []*tfpb.TensorShapeProto_Dim{
{Size: 4},
{Size: 5},
{Size: 6},
},
},
},
},
MethodName: "method",
})
want := Signature{
Inputs: map[string]TensorInfo{
"input_1": TensorInfo{
Name: "tensor_1",
DType: Int8,
Shape: MakeShape(1, 2, 3),
},
"input_2": TensorInfo{
Name: "tensor_2",
DType: Float,
Shape: MakeShape(4, 5, 6),
},
},
Outputs: map[string]TensorInfo{
"output_1": TensorInfo{
Name: "tensor_3",
DType: String,
Shape: MakeShape(1, 2, 3),
},
"output_2": TensorInfo{
Name: "tensor_4",
DType: Bool,
Shape: MakeShape(4, 5, 6),
},
},
MethodName: "method",
}
for k, input := range want.Inputs {
diff, err := diffTensorInfos(got.Inputs[k], input)
if err != nil {
t.Fatalf("Signature.Inputs[%s]: unable to diff TensorInfos: %v", k, err)
}
if diff != "" {
t.Errorf("Signature.Inputs[%s] diff:\n%s", k, diff)
}
}
for k, output := range want.Outputs {
diff, err := diffTensorInfos(got.Outputs[k], output)
if err != nil {
t.Fatalf("Signature.Outputs[%s]: unable to diff TensorInfos: %v", k, err)
}
if diff != "" {
t.Errorf("Signature.Outputs[%s] diff:\n%s", k, diff)
}
}
if got.MethodName != want.MethodName {
t.Errorf("Signature.MethodName: got %q, want %q", got.MethodName, want.MethodName)
}
}
func TestTensorInfoFromProto(t *testing.T) {
got := tensorInfoFromProto(&tfpb.TensorInfo{
Encoding: &tfpb.TensorInfo_Name{
Name: "tensor",
},
Dtype: tfpb.DataType_DT_INT8,
TensorShape: &tfpb.TensorShapeProto{
Dim: []*tfpb.TensorShapeProto_Dim{
{Size: 1},
{Size: 2},
{Size: 3},
},
},
})
want := TensorInfo{
Name: "tensor",
DType: Int8,
Shape: MakeShape(1, 2, 3),
}
diff, err := diffTensorInfos(got, want)
if err != nil {
t.Fatalf("Unable to diff TensorInfos: %v", err)
}
if diff != "" {
t.Errorf("tensorInfoFromProto produced a diff (got -> want): %s", diff)
}
}
func diffTensorInfos(a, b TensorInfo) (string, error) {
diff := ""
if a.Name != b.Name {
diff += fmt.Sprintf("Name: %q -> %q\n", a.Name, b.Name)
}
if a.DType != b.DType {
diff += fmt.Sprintf("DType: %v -> %v\n", a.DType, b.DType)
}
aShape, err := a.Shape.ToSlice()
if err != nil {
return "", err
}
bShape, err := b.Shape.ToSlice()
if err != nil {
return "", err
}
shapeLen := len(aShape)
if len(bShape) > shapeLen {
shapeLen = len(bShape)
}
for i := 0; i < shapeLen; i++ {
if i >= len(aShape) {
diff += fmt.Sprintf("+Shape[%d]: %d\n", i, bShape[i])
continue
}
if i >= len(bShape) {
diff += fmt.Sprintf("-Shape[%d]: %d\n", i, aShape[i])
continue
}
if aShape[i] != bShape[i] {
diff += fmt.Sprintf("Shape[%d]: %d -> %d\n", i, aShape[i], bShape[i])
}
}
return diff, nil
}