TensorFlow Frontend前端

TensorFlow Frontend前端

TensorFlow前端有助于将TensorFlow模型导入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)

导出过程将删除许多不需要进行推理的节点，但不幸的是会留下一些剩余的节点。应该手动删除的节点：

Dropout, including Dropout and DropoutWrapper

Assert

TVM对动态张量形状的支持最少。None应将尺寸替换为常量。例如，模型可以接受带有shape的输入(None,20)。这应转换为的形状(1,20)。应该相应地修改模型，以确保这些形状在整个图形中都匹配。

Export

TensorFlow前端需要冻结的protobuf（.pb）或保存的模型作为输入。不支持检查点（.ckpt）。TensorFlow前端所需的graphdef，可以从活动会话中提取，可以使用TFParser帮助器类提取。

应该导出该模型并进行许多转换，以准备模型进行推理。设置`add_shapes=True`也很重要，因为这会将每个节点的输出形状嵌入到图形中。这是一个给定会话将模型导出为protobuf的函数：

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:

确保可以在整个图形中知道形状，将`shape`参数传递给`from_tensorflow`。该词典将输入名称映射到输入形状。

Data Layout

大多数TensorFlow模型以NHWC布局发布。NCHW布局通常提供更好的性能，尤其是在GPU上。该TensorFlow前端可以通过传递参数自动转换模型的数据布局`layout='NCHW'`到`from_tensorflow`。

Best Practices

使用静态张量形状代替动态形状（删除`None`尺寸）。

`TensorArray`目前尚不支持使用静态RNN代替动态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

人工智能芯片与自动驾驶