Function Argument #1064

hedaoyuan · 2017-01-04T08:53:30Z

This PR is to achieve argument types of the Functions.

关于Function的BufferArg描述见issue 892
基于BufferArgs参数类型的Function:: calc实现和调用方式如下：

template <DeviceType DType>
void Function(const BufferArgs& arguments) {
  // get argument
  auto input = arguments[0].matrix<DType>();
  auto output = arguments[1].matrix<DType>();
  // call computation api
  FunctionApi<DType>(output, input);
}

TEST(BufferTest, Function) {
  CpuMatrix cpuInput = CpuMatrix(100, 200);
  CpuMatrix cpuOutput = CpuMatrix(100, 200);
  // prepare arguments
  BufferArgs cpuArgments;
  cpuArgments.addArg(cpuInput);
  cpuArgments.addArg(cpuOutput);
  // call Function
  Function<DEVICE_TYPE_CPU>(cpuArgments);
}

tianbingsz

Looks very good to me.

tianbingsz · 2017-01-06T19:23:37Z

paddle/function/BufferArg.h

+      : buf_(buf), valueType_(valueType) {}
+
+  BufferArg(const Matrix& matrix)
+      : buf_((void*)matrix.getData()),


can we use reinterpret_cast<void*>(matrix.getData())?

tianbingsz · 2017-01-06T19:24:16Z

paddle/function/BufferArg.h

+  }
+
+  BufferArg(const Matrix& matrix, const TensorShape& shape)
+      : buf_((void*)matrix.getData()),


same as line 59

tianbingsz · 2017-01-06T19:25:09Z

paddle/function/BufferArg.h

+  }
+
+  BufferArg(const Vector& vector)
+      : buf_((void*)vector.getData()),


same as line 59.

tianbingsz · 2017-01-06T19:25:20Z

paddle/function/BufferArg.h

+  }
+
+  BufferArg(const IVector& vector)
+      : buf_((void*)vector.getData()), valueType_(VALUE_TYPE_INT32), shape_(1) {


tianbingsz · 2017-01-06T19:50:51Z

paddle/function/BufferArg.h

+// sequence start positions in a mini-batch of sequences
+// shape_.ndims() == 1
+// valueType_ = int32
+// if a < b than value_.buf_[a] < value_.buf_[b]


if a < b then value_.buf_[a] < value_.buf_[b]

tianbingsz · 2017-01-06T23:03:53Z

paddle/function/TensorShape.h

+  TensorShape(std::initializer_list<size_t> dims) {
+    ndims_ = dims.size();
+    initDims(ndims_);
+    std::copy(dims.begin(), dims.end(), dims_.begin());


Can we use dims_.assign(dims.begin(), dims.end()) ?

tianbingsz · 2017-01-06T23:21:33Z

paddle/function/FunctionTest.cpp

+
+template <DeviceType DType>
+void Function(const BufferArgs& arguments) {
+  auto input = arguments[0].matrix<DType>();


const auto input = arguments[0].matrix();

tianbingsz · 2017-01-06T23:43:04Z

paddle/function/ContextProjectionOp.cpp

+    CHECK_EQ(inputs[0].shape()[0], outputs[0].shape()[0]);
+
+    auto out_mat = outputs[0].matrix<Device>();
+    auto in_mat = inputs[0].matrix<Device>();


const auto in_mat = inputs[0].matrix();

tianbingsz · 2017-01-06T23:43:18Z

paddle/function/ContextProjectionOp.cpp

+
+    auto out_mat = outputs[0].matrix<Device>();
+    auto in_mat = inputs[0].matrix<Device>();
+    auto w_mat = !inputs[1].data()


const auto w_mat

tianbingsz · 2017-01-06T23:48:43Z

paddle/function/ContextProjectionOp.cpp

+    auto w_grad_mat = !inputs[1].data()
+                          ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
+                          : inputs[1].matrix<Device>();
+    auto seq_vec = inputs[2].vector<int, Device>();


const auto seq_vec

hedaoyuan · 2017-01-09T15:27:39Z

commit 57e2521 对BufferArg增加一个ArgType argType_属性，同时去掉了Function的inouts参数。

Function的inouts参数原先的设计是用来区分outputs参数的；对于有些情况，Function计算结果是直接赋值到output中（比如，PoolLayer::forward等），对于有些情况Function计算结果是累加到output上（比如，MixedLaye和Layer::backward计算等）；对应assign to output的，是一个writeonly参数，用outputs来传参；add to output的，是一个read and write参数，用inouts来传参。

commit 57e2521修改的目的是，inouts参数的主要目的是为了描述Function计算结果是assign to output还是add to output，对BufferArg增加一个值为ASSIGN_TO或ADD_TO的属性更能清楚的表示这件事情；另外，对于有多个output的Function，可以对每个output单独标识ASSIGN_TO或ADD_TO的属性。

tianbingsz · 2017-01-09T20:20:33Z

@hedaoyuan, LGTM, please make the checks pass and will approve it.

* add sharding for gpt-3 * del debug * add sharding save model * update model save * fix seed func * set control in tensor parallel Co-authored-by: Zhong Hui <zhonghui.net@gmail.com>

hedaoyuan added 5 commits January 3, 2017 15:51

add TensorShape use to represent tensor of any dimension.

904eefa

add TensorType.h

0c4be7e

add BufferArg

c5c8051

Modify the argument type of Function

68156c8

Modify the argument type of ContextProjectionFunc

41c52d3

hedaoyuan requested review from emailweixu and tianbingsz January 5, 2017 12:42

hedaoyuan added 2 commits January 5, 2017 21:32

add some comments for Function.h

f3fdfd9

add FunctionTest.cpp

ccf0b1b

tianbingsz suggested changes Jan 6, 2017

View reviewed changes

hedaoyuan added 3 commits January 9, 2017 11:47

follow commit

d35ef9d

BufferArg add ArgType and Function remove inouts

57e2521

fix CrossMapNormalFunc and ContextProjectionFunc(remove inouts argument)

df9be2d

hedaoyuan assigned emailweixu and tianbingsz Jan 9, 2017

Merge branch 'develop' of https://github.com/baidu/Paddle into buffer

ee2da53

hedaoyuan force-pushed the buffer branch from a161f81 to 9bbe5ac Compare January 10, 2017 04:49

Bug fix for mac os

ae4400b

hedaoyuan force-pushed the buffer branch from 9bbe5ac to ae4400b Compare January 10, 2017 05:18

tianbingsz approved these changes Jan 10, 2017

View reviewed changes

hedaoyuan merged commit 7df67ba into PaddlePaddle:develop Jan 10, 2017

hedaoyuan mentioned this pull request Jan 12, 2017

Paddle Function #892

Closed

lizexu123 pushed a commit to lizexu123/Paddle that referenced this pull request Feb 23, 2024

support dict batch in dynabert (PaddlePaddle#1064)

6b8b683

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Function Argument #1064

Function Argument #1064

hedaoyuan commented Jan 4, 2017 •

edited

Loading

tianbingsz left a comment

tianbingsz Jan 6, 2017

hedaoyuan Jan 9, 2017

tianbingsz Jan 6, 2017

hedaoyuan Jan 9, 2017

tianbingsz Jan 6, 2017

hedaoyuan Jan 9, 2017

tianbingsz Jan 6, 2017

hedaoyuan Jan 9, 2017

tianbingsz Jan 6, 2017

hedaoyuan Jan 9, 2017

tianbingsz Jan 6, 2017 •

edited

Loading

tianbingsz Jan 6, 2017

tianbingsz Jan 6, 2017

tianbingsz Jan 6, 2017

tianbingsz Jan 6, 2017

hedaoyuan commented Jan 9, 2017

tianbingsz commented Jan 9, 2017

Function Argument #1064

Function Argument #1064

Conversation

hedaoyuan commented Jan 4, 2017 • edited Loading

tianbingsz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tianbingsz Jan 6, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hedaoyuan commented Jan 9, 2017

tianbingsz commented Jan 9, 2017

hedaoyuan commented Jan 4, 2017 •

edited

Loading

tianbingsz Jan 6, 2017 •

edited

Loading