Enhance sequence_expand operator #9100
Conversation
|
|
pkuyym
changed the title
| AddComment(R"DOC( | ||
| Sequence Expand Operator. | ||
|
|
||
| This operator expands input(X) according to LOD of input(Y). | ||
| This operator expands `X` according to specified level lod of `Y`. Current | ||
| implementation constaints that lod level of `X` should be at most 1. Attribute |
There was a problem hiding this comment.
constaints -> requires
I feel the semantics of this op very difficult to understand. Is there an op in other framework that has the same effect?
There was a problem hiding this comment.
I don't think so, since LoDTensor is a peculiar concept in PaddlePaddle and this operator is dependent on the LoD of input variable.
| and input(Y) | ||
| Y.lod = [[0, 2, 4], | ||
| [0, 3, 6, 6, 8]] | ||
| ref_level: 0 |
There was a problem hiding this comment.
what would happen if this if -1?
There was a problem hiding this comment.
When set ref_level to -1, the LoD size would be inconsistent between [0, 1, 4] and [0, 3, 6, 6, 8].
| [0, 3, 6, 6, 8]] | ||
| ref_level: 0 | ||
| then we get 1-level LoDTensor | ||
| Out.lod = [[0, 2, 5, 8]] |
There was a problem hiding this comment.
Out.lod[0][2:8]是由X.lod[0][1:4] repeated 2 times 而来的？
那么， X.lod[0][0:1] 是不是应该repeated 2 times而成 ： Out.lod = [[0, 1， 2, 5, 8]] ？
| Out.lod = [[0, 2, 3, 6]] | ||
| Out.data = [a, a, b, c, c, c] | ||
| ref_level: -1 | ||
| then we a common Tensor |
There was a problem hiding this comment.
(LodTensor, default LoDTensor) Output LoDTensor which is generated from Input(X) by referring lod of Input(Y).
输出应该都是LoDTensor吧？同时给出Out.lod ?
There was a problem hiding this comment.
we get done.
Out is a Tensor.
| Out.lod = [[0, 2, 3, 6]] | ||
| Out.data = [[a,b], [a,b] [c,d], [e, f], [e, f], [e, f]] | ||
| ref_level: 0 | ||
| then we get a common LoDTensor |
There was a problem hiding this comment.
The LoD of output tensor should be given here?
| int ref_level = context.Attr<int>("ref_level"); | ||
| auto& x_lod = x->lod(); | ||
| auto& y_lod = y->lod(); | ||
| PADDLE_ENFORCE_GT(y_lod.size(), 0, |
There was a problem hiding this comment.
y_lod.size() and y_lod.size() have been checked in runtime InferShape
Resolves #9049