[
https://issues.apache.org/jira/browse/SINGA-506?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
zhangzhaoqi updated SINGA-506:
------------------------------
Description:
*We are going to support these three NLP models, called, Bidirectional
Attention Flow, BERT-Squad and GPT-2.*
*Totally, there are still 19 operators that we need to add as following,*
*For details, these 19 operators belong to these three models separately:*
|{color:#000000}*Operator*{color}|{color:#000000}*Rank*{color}|{color:#000000}*Workload*{color}|{color:#000000}*Comments*{color}|{color:#000000}*Bidirectional
Attention
Flow*{color}|{color:#000000}*BERT-Squad*{color}|{color:#000000}*GPT-2*{color}|
|-{color:#000000}*Transpose*{color}-|{color:#000000}easy{color}|{color:#000000}1h{color}|{color:#000000}Transpose
the input tensor similar to numpy.transpose.
{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|-{color:#000000}*ConstantOfShape*{color}-|{color:#000000}easy{color}|{color:#000000}2h{color}|{color:#000000}Generate
a tensor with given value and shape.{color}|{color:#000000}T{color}|
|{color:#000000}T{color}|
|-{color:#000000}*Shape*{color}-|{color:#000000}easy{color}|{color:#000000}2h{color}|{color:#000000}Takes
a tensor as input and outputs an 1D int64 tensor containing the shape of the
input
tensor.{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|-{color:#000000}*Dropout*{color}-|{color:#000000}easy{color}|{color:#000000}3h{color}|{color:#000000}Dropout
takes an input floating-point tensor and an input ratio (floating-point
scalar), and produces two tensor outputs, output (floating-point tensor) and
mask (Tensor<bool>). {color}|{color:#000000}T{color}| | |
|-{color:#000000}*Ceil*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}y
= ceil(x){color}|{color:#000000}T{color}| | |
|{color:#000000}*ReduceMax*{color}|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
the max of the input tensor's element along the provided axes.
{color}|{color:#000000}T{color}| | |
|-{color:#000000}*ReduceMean*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
the mean of the input tensor's element along the provided axes.{color}|
|{color:#000000}T{color}|{color:#000000}T{color}|
|-{color:#000000}*ReduceSum*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
the sum of the input tensor's element along the provided
axes.{color}|{color:#000000}T{color}| | |
|-{color:#000000}*Slice*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Produces
a slice of the input tensor along multiple axes.
{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*Compress*{color}|{color:#000000}easy{color}|{color:#000000}6h{color}|{color:#000000}Selects
slices from an input tensor along a given axis where condition evaluates to
True for each axis index.{color}|{color:#000000}T{color}| | |
|{color:#000000}*Hardmax*{color}|{color:#000000}easy{color}|{color:#000000}6h{color}|{color:#000000}The
operator computes the hardmax (1 for the first maximum value, and 0 for all
others) values for each layer in the batch of the given
input.{color}|{color:#000000}T{color}| | |
|{color:#000000}*NonZero*{color}|{color:#000000}easy{color}|{color:#000000}12h{color}|{color:#000000}Returns
the indices of the elements that are non-zero (in row-major order - by
dimension).{color}| | |{color:#000000}T{color}|
|-{color:#000000}*Split*{color}-|{color:#000000}easy{color}|{color:#000000}12h{color}|{color:#000000}Split
a tensor into a list of tensors, along the specified 'axis'.{color}|
|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*Tile*{color}|{color:#000000}easy{color}|{color:#000000}1d{color}|{color:#000000}Constructs
a tensor by tiling a given tensor. This is the same as function tile in Numpy,
but no broadcast. For example A = [[1, 2], [3, 4]], B = [1, 2], tile(A, B) =
[[1, 2, 1, 2], [3, 4, 3, 4]]{color}| |{color:#000000}T{color}| |
|{color:#000000}*ArgMax*{color}|{color:#000000}complicated{color}|{color:#000000}2d{color}|{color:#000000}Computes
the indices of the max elements of the input tensor's element along the
provided axis. {color}|{color:#000000}T{color}| | |
|{color:#000000}*Gather*{color}|{color:#000000}complicated{color}|{color:#000000}3d{color}|{color:#000000}Given
data tensor of rank r >= 1, and indices tensor of rank q, gather entries of
the axis dimension of data (by default outer-most one as axis=0) indexed by
indices, and concatenates
them{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*Scan*{color}|{color:#000000}hard{color}|{color:#000000}2w{color}|{color:#000000}Scan
can be used to iterate over one or more scan_input tensors, constructing zero
or more scan_output tensors. It combines ideas from general recurrences,
functional programming constructs such as scan, fold, map, and zip and is
intended to enable generalizations of RNN-like constructs for
sequence-to-sequence processing.{color}|{color:#000000}T{color}| | |
|{color:#000000}*Cast*{color}|{color:#000000}hard{color}|{color:#000000}-{color}|{color:#000000}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 the
converted
type.{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*CategoryMapper*{color}|
|{color:#000000}-{color}|{color:#000000}not in onnx
document{color}|{color:#000000}T{color}| | |
*Bidirectional Attention Flow:*
ArgMax
Cast
CategoryMapper
Ceil
Compress
ConstantOfShape
Dropout
Gather
Hardmax
ReduceMax
ReduceSum
Scan
Shape
Slice
Transpose
*BERT-Squad:*
Slice
Shape
Gather
ReduceMean
Cast
Tile
Transpose
Split
*GPT-2:*
ConstantOfShape
Slice
Shape
Gather
ReduceMean
NonZero
Cast
Transpose
Split
was:
*We are going to support these three NLP models, called, Bidirectional
Attention Flow, BERT-Squad and GPT-2.*
*Totally, there are still 19 operators that we need to add as following,*
*For details, these 19 operators belong to these three models separately:*
|{color:#000000}*Operator*{color}|{color:#000000}*Rank*{color}|{color:#000000}*Workload*{color}|{color:#000000}*Comments*{color}|{color:#000000}*Bidirectional
Attention
Flow*{color}|{color:#000000}*BERT-Squad*{color}|{color:#000000}*GPT-2*{color}|
|{color:#000000}*Transpose*{color}|{color:#000000}easy{color}|{color:#000000}1h{color}|{color:#000000}Transpose
the input tensor similar to numpy.transpose.
{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*ConstantOfShape*{color}|{color:#000000}easy{color}|{color:#000000}2h{color}|{color:#000000}Generate
a tensor with given value and shape.{color}|{color:#000000}T{color}|
|{color:#000000}T{color}|
|{color:#000000}*Shape*{color}|{color:#000000}easy{color}|{color:#000000}2h{color}|{color:#000000}Takes
a tensor as input and outputs an 1D int64 tensor containing the shape of the
input
tensor.{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*Dropout*{color}|{color:#000000}easy{color}|{color:#000000}3h{color}|{color:#000000}Dropout
takes an input floating-point tensor and an input ratio (floating-point
scalar), and produces two tensor outputs, output (floating-point tensor) and
mask (Tensor<bool>). {color}|{color:#000000}T{color}| | |
|{color:#000000}*Ceil*{color}|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}y
= ceil(x){color}|{color:#000000}T{color}| | |
|{color:#000000}*ReduceMax*{color}|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
the max of the input tensor's element along the provided axes.
{color}|{color:#000000}T{color}| | |
|{color:#000000}*ReduceMean*{color}|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
the mean of the input tensor's element along the provided axes.{color}|
|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*ReduceSum*{color}|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
the sum of the input tensor's element along the provided
axes.{color}|{color:#000000}T{color}| | |
|{color:#000000}*Slice*{color}|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Produces
a slice of the input tensor along multiple axes.
{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*Compress*{color}|{color:#000000}easy{color}|{color:#000000}6h{color}|{color:#000000}Selects
slices from an input tensor along a given axis where condition evaluates to
True for each axis index.{color}|{color:#000000}T{color}| | |
|{color:#000000}*Hardmax*{color}|{color:#000000}easy{color}|{color:#000000}6h{color}|{color:#000000}The
operator computes the hardmax (1 for the first maximum value, and 0 for all
others) values for each layer in the batch of the given
input.{color}|{color:#000000}T{color}| | |
|{color:#000000}*NonZero*{color}|{color:#000000}easy{color}|{color:#000000}12h{color}|{color:#000000}Returns
the indices of the elements that are non-zero (in row-major order - by
dimension).{color}| | |{color:#000000}T{color}|
|{color:#000000}*Split*{color}|{color:#000000}easy{color}|{color:#000000}12h{color}|{color:#000000}Split
a tensor into a list of tensors, along the specified 'axis'.{color}|
|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*Tile*{color}|{color:#000000}easy{color}|{color:#000000}1d{color}|{color:#000000}Constructs
a tensor by tiling a given tensor. This is the same as function tile in Numpy,
but no broadcast. For example A = [[1, 2], [3, 4]], B = [1, 2], tile(A, B) =
[[1, 2, 1, 2], [3, 4, 3, 4]]{color}| |{color:#000000}T{color}| |
|{color:#000000}*ArgMax*{color}|{color:#000000}complicated{color}|{color:#000000}2d{color}|{color:#000000}Computes
the indices of the max elements of the input tensor's element along the
provided axis. {color}|{color:#000000}T{color}| | |
|{color:#000000}*Gather*{color}|{color:#000000}complicated{color}|{color:#000000}3d{color}|{color:#000000}Given
data tensor of rank r >= 1, and indices tensor of rank q, gather entries of
the axis dimension of data (by default outer-most one as axis=0) indexed by
indices, and concatenates
them{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*Scan*{color}|{color:#000000}hard{color}|{color:#000000}2w{color}|{color:#000000}Scan
can be used to iterate over one or more scan_input tensors, constructing zero
or more scan_output tensors. It combines ideas from general recurrences,
functional programming constructs such as scan, fold, map, and zip and is
intended to enable generalizations of RNN-like constructs for
sequence-to-sequence processing.{color}|{color:#000000}T{color}| | |
|{color:#000000}*Cast*{color}|{color:#000000}hard{color}|{color:#000000}-{color}|{color:#000000}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 the
converted
type.{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
|{color:#000000}*CategoryMapper*{color}|
|{color:#000000}-{color}|{color:#000000}not in onnx
document{color}|{color:#000000}T{color}| | |
*Bidirectional Attention Flow:*
ArgMax
Cast
CategoryMapper
Ceil
Compress
ConstantOfShape
Dropout
Gather
Hardmax
ReduceMax
ReduceSum
Scan
Shape
Slice
Transpose
*BERT-Squad:*
Slice
Shape
Gather
ReduceMean
Cast
Tile
Transpose
Split
*GPT-2:*
ConstantOfShape
Slice
Shape
Gather
ReduceMean
NonZero
Cast
Transpose
Split
> add autograd operators for NLP models
> -------------------------------------
>
> Key: SINGA-506
> URL: https://issues.apache.org/jira/browse/SINGA-506
> Project: Singa
> Issue Type: New Feature
> Reporter: zhangzhaoqi
> Priority: Major
> Time Spent: 10m
> Remaining Estimate: 0h
>
> *We are going to support these three NLP models, called, Bidirectional
> Attention Flow, BERT-Squad and GPT-2.*
> *Totally, there are still 19 operators that we need to add as following,*
>
> *For details, these 19 operators belong to these three models separately:*
> |{color:#000000}*Operator*{color}|{color:#000000}*Rank*{color}|{color:#000000}*Workload*{color}|{color:#000000}*Comments*{color}|{color:#000000}*Bidirectional
> Attention
> Flow*{color}|{color:#000000}*BERT-Squad*{color}|{color:#000000}*GPT-2*{color}|
> |-{color:#000000}*Transpose*{color}-|{color:#000000}easy{color}|{color:#000000}1h{color}|{color:#000000}Transpose
> the input tensor similar to numpy.transpose.
> {color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
> |-{color:#000000}*ConstantOfShape*{color}-|{color:#000000}easy{color}|{color:#000000}2h{color}|{color:#000000}Generate
> a tensor with given value and shape.{color}|{color:#000000}T{color}|
> |{color:#000000}T{color}|
> |-{color:#000000}*Shape*{color}-|{color:#000000}easy{color}|{color:#000000}2h{color}|{color:#000000}Takes
> a tensor as input and outputs an 1D int64 tensor containing the shape of the
> input
> tensor.{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
> |-{color:#000000}*Dropout*{color}-|{color:#000000}easy{color}|{color:#000000}3h{color}|{color:#000000}Dropout
> takes an input floating-point tensor and an input ratio (floating-point
> scalar), and produces two tensor outputs, output (floating-point tensor) and
> mask (Tensor<bool>). {color}|{color:#000000}T{color}| | |
> |-{color:#000000}*Ceil*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}y
> = ceil(x){color}|{color:#000000}T{color}| | |
> |{color:#000000}*ReduceMax*{color}|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
> the max of the input tensor's element along the provided axes.
> {color}|{color:#000000}T{color}| | |
> |-{color:#000000}*ReduceMean*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
> the mean of the input tensor's element along the provided axes.{color}|
> |{color:#000000}T{color}|{color:#000000}T{color}|
> |-{color:#000000}*ReduceSum*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Computes
> the sum of the input tensor's element along the provided
> axes.{color}|{color:#000000}T{color}| | |
> |-{color:#000000}*Slice*{color}-|{color:#000000}easy{color}|{color:#000000}4h{color}|{color:#000000}Produces
> a slice of the input tensor along multiple axes.
> {color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
> |{color:#000000}*Compress*{color}|{color:#000000}easy{color}|{color:#000000}6h{color}|{color:#000000}Selects
> slices from an input tensor along a given axis where condition evaluates to
> True for each axis index.{color}|{color:#000000}T{color}| | |
> |{color:#000000}*Hardmax*{color}|{color:#000000}easy{color}|{color:#000000}6h{color}|{color:#000000}The
> operator computes the hardmax (1 for the first maximum value, and 0 for all
> others) values for each layer in the batch of the given
> input.{color}|{color:#000000}T{color}| | |
> |{color:#000000}*NonZero*{color}|{color:#000000}easy{color}|{color:#000000}12h{color}|{color:#000000}Returns
> the indices of the elements that are non-zero (in row-major order - by
> dimension).{color}| | |{color:#000000}T{color}|
> |-{color:#000000}*Split*{color}-|{color:#000000}easy{color}|{color:#000000}12h{color}|{color:#000000}Split
> a tensor into a list of tensors, along the specified 'axis'.{color}|
> |{color:#000000}T{color}|{color:#000000}T{color}|
> |{color:#000000}*Tile*{color}|{color:#000000}easy{color}|{color:#000000}1d{color}|{color:#000000}Constructs
> a tensor by tiling a given tensor. This is the same as function tile in
> Numpy, but no broadcast. For example A = [[1, 2], [3, 4]], B = [1, 2],
> tile(A, B) = [[1, 2, 1, 2], [3, 4, 3, 4]]{color}| |{color:#000000}T{color}| |
> |{color:#000000}*ArgMax*{color}|{color:#000000}complicated{color}|{color:#000000}2d{color}|{color:#000000}Computes
> the indices of the max elements of the input tensor's element along the
> provided axis. {color}|{color:#000000}T{color}| | |
> |{color:#000000}*Gather*{color}|{color:#000000}complicated{color}|{color:#000000}3d{color}|{color:#000000}Given
> data tensor of rank r >= 1, and indices tensor of rank q, gather entries of
> the axis dimension of data (by default outer-most one as axis=0) indexed by
> indices, and concatenates
> them{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
> |{color:#000000}*Scan*{color}|{color:#000000}hard{color}|{color:#000000}2w{color}|{color:#000000}Scan
> can be used to iterate over one or more scan_input tensors, constructing
> zero or more scan_output tensors. It combines ideas from general recurrences,
> functional programming constructs such as scan, fold, map, and zip and is
> intended to enable generalizations of RNN-like constructs for
> sequence-to-sequence processing.{color}|{color:#000000}T{color}| | |
> |{color:#000000}*Cast*{color}|{color:#000000}hard{color}|{color:#000000}-{color}|{color:#000000}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 the
> converted
> type.{color}|{color:#000000}T{color}|{color:#000000}T{color}|{color:#000000}T{color}|
> |{color:#000000}*CategoryMapper*{color}|
> |{color:#000000}-{color}|{color:#000000}not in onnx
> document{color}|{color:#000000}T{color}| | |
> *Bidirectional Attention Flow:*
> ArgMax
> Cast
> CategoryMapper
> Ceil
> Compress
> ConstantOfShape
> Dropout
> Gather
> Hardmax
> ReduceMax
> ReduceSum
> Scan
> Shape
> Slice
> Transpose
> *BERT-Squad:*
> Slice
> Shape
> Gather
> ReduceMean
> Cast
> Tile
> Transpose
> Split
> *GPT-2:*
> ConstantOfShape
> Slice
> Shape
> Gather
> ReduceMean
> NonZero
> Cast
> Transpose
> Split
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
