Add ReorderLoDTensorByRank #6789
Conversation
reyoung
force-pushed
the
feature/add_reorder_lod_tensor
branch
from
fe473f9
to
70236b6
Compare
| AddInput("RankTable", | ||
| "(LoDRankTable) the rank table that input need follow"); | ||
| AddOutput("Out", "(LoDTensor) reordered lod tensor"); | ||
| AddComment(R"DOC(ReorderLoDTensorLoDRankTable |
There was a problem hiding this comment.
ReorderLoDTensorLoDRankTable --> ReorderLoDTensorByRankTable
| index [3, 0, 2, 1]. Input X will reorder its sequence, the third sequence of | ||
| X will be the first sequence of Output. | ||
|
|
||
| NOTE: The RankTable does not need to be calculated by X. |
There was a problem hiding this comment.
Please give a specific example to make things easier to understand.
| namespace paddle { | ||
| namespace operators { | ||
|
|
||
| class ReorderLoDTensorProtoMaker : public framework::OpProtoAndCheckerMaker { |
There was a problem hiding this comment.
Better to change to ReorderLoDTensorByRankTableProtoMaker ?
|
|
||
|
|
||
| def reorder_lod_tensor_by_rank(x, rank_table): | ||
| """ |
There was a problem hiding this comment.
I think we could use python to generate the docstring.
| size_t out_offset = 0; | ||
| out->mutable_lod()->clear(); | ||
| for (auto &item : rank_table.items()) { | ||
| out_offset = this->CopyTensorAndLod(dev_ctx, absolute_table[item.index], |
There was a problem hiding this comment.
Please add some assert to make sure absolute_table[item.index] valid
|
|
||
| out.Resize(x.dims()); | ||
| out.mutable_data(x.place(), x.type()); | ||
| this->process(dev_ctx, x, rank_table, &out); |
There was a problem hiding this comment.
Please make sure the size of rank_table and size of x' lod is match.
There was a problem hiding this comment.
It will be checked during Tensor::Slice.
|
|
||
|
|
||
| class TestReorderLoDTensor(unittest.TestCase): | ||
| def test_reorder(self): |
There was a problem hiding this comment.
It has been tested, since IG.sum() == 1 and LoD can be restored.
It is useful to reorder RNN memory block.
reyoung
force-pushed
the
feature/add_reorder_lod_tensor
branch
from
70236b6
to
b23982a
Compare
It is useful to reorder RNN memory block.