TensorFlow Frontend前端

TensorFlow Frontend前端

TensorFlow前端有助于将TensorFlow模型導入TVM。

Supported versions:

1.12 and below

Tested models:

Inception (V1/V2/V3/V4)

Resnet (All)

Mobilenet (V1/V2 All)

Vgg (16/19)

BERT (Base/3-layer)

導出過程将删除許多不需要進行推理的節點，但不幸的是會留下一些剩餘的節點。應該手動删除的節點：

Dropout, including Dropout and DropoutWrapper

Assert

TVM對動态張量形狀的支援最少。None應将尺寸替換為常量。例如，模型可以接受帶有shape的輸入(None,20)。這應轉換為的形狀(1,20)。應該相應地修改模型，以確定這些形狀在整個圖形中都比對。

Export

TensorFlow前端需要當機的protobuf（.pb）或儲存的模型作為輸入。不支援檢查點（.ckpt）。TensorFlow前端所需的graphdef，可以從活動會話中提取，可以使用TFParser幫助器類提取。

應該導出該模型并進行許多轉換，以準備模型進行推理。設定`add_shapes=True`也很重要，因為這會将每個節點的輸出形狀嵌入到圖形中。這是一個給定會話将模型導出為protobuf的函數：

import tensorflow as tf

from tensorflow.tools.graph_transforms import TransformGraph

def export_pb(session):

    with tf.gfile.GFile("myexportedmodel.pb", "wb") as f:

        inputs = ["myinput1", "myinput2"] # replace with your input names

        outputs = ["myoutput1"] # replace with your output names

        graph_def = session.graph.as_graph_def(add_shapes=True)

        graph_def = tf.graph.util.convert_variables_to_constants(session, graph_def, outputs)

        graph_def = TransformGraph(

            graph_def,

            inputs,

            outputs,

            [

                "remove_nodes(op=Identity, op=CheckNumerics, op=StopGradient)",

                "sort_by_execution_order", # sort by execution order after each transform to ensure correct node ordering

                "remove_attribute(attribute_name=_XlaSeparateCompiledGradients)",

                "remove_attribute(attribute_name=_XlaCompile)",

                "remove_attribute(attribute_name=_XlaScope)",

                "sort_by_execution_order",

                "remove_device",

                "fold_batch_norms",

                "fold_old_batch_norms",

                "sort_by_execution_order"

            ]

        )

        f.write(graph_def.SerializeToString())

Another method is to export and freeze the graph.

Import the Model

Explicit Shape:

確定可以在整個圖形中知道形狀，将`shape`參數傳遞給`from_tensorflow`。該詞典将輸入名稱映射到輸入形狀。

Data Layout

大多數TensorFlow模型以NHWC布局釋出。NCHW布局通常提供更好的性能，尤其是在GPU上。該TensorFlow前端可以通過傳遞參數自動轉換模型的資料布局`layout='NCHW'`到`from_tensorflow`。

Best Practices

使用靜态張量形狀代替動态形狀（删除`None`尺寸）。

`TensorArray`目前尚不支援使用靜态RNN代替動态RNN。

Supported Ops

Abs

Add

AddN

All

Any

ArgMax

ArgMin

AvgPool

BatchMatMul

BatchMatMulV2

BatchNormWithGlobalNormalization

BatchToSpaceND

BiasAdd

BroadcastTo

Cast

Ceil

CheckNumerics

ClipByValue

Concat

ConcatV2

Conv2D

Cos

Tan

CropAndResize

DecodeJpeg

DepthwiseConv2dNative

DepthToSpace

Dilation2D

Equal

Elu

Enter

Erf

Exit

Exp

ExpandDims

Fill

Floor

FloorDiv

FloorMod

FusedBatchNorm

FusedBatchNormV2

Gather

GatherNd

GatherV2

Greater

GreaterEqual

Identity

IsFinite

IsInf

IsNan

LeakyRelu

LeftShift

Less

LessEqual

Log

Log1p

LoopCond

LogicalAnd

LogicalOr

LogicalNot

LogSoftmax

LRN

LSTMBlockCell

MatMul

Max

MaxPool

Maximum

Mean

Merge

Min

Minimum

MirrorPad

Mod

Mul

Neg

NextIteration

NotEqual

OneHot

Pack

Pad

PadV2

Pow

Prod

Range

Rank

RealDiv

Relu

Relu6

Reshape

ResizeBilinear

ResizeBicubic

ResizeNearestNeighbor

ReverseV2

RightShift

Round

Rsqrt

Select

Selu

Shape

Sigmoid

Sign

Sin

Size

Slice

Softmax

Softplus

SpaceToBatchND

SpaceToDepth,

Split

SplitV

Sqrt

Square

SquareDifference

Squeeze

StridedSlice

Sub

Sum

Switch

Tanh

TensorArrayV3

TensorArrayScatterV3

TensorArrayGatherV3

TensorArraySizeV3

TensorArrayWriteV3

TensorArrayReadV3

TensorArraySplitV3

TensorArrayConcatV3

Tile

TopKV2

Transpose

TruncateMod

Unpack

UnravelIndex

Where

ZerosLike

人工智能晶片與自動駕駛