Design: infer_var_type #4795
Design: infer_var_type #4795
Conversation
|
Maybe we should also register InferShape to OpInfo @jacquesqiao |
|
|
||
| The variable in our design can hold variant types. Such as `LoDTensor` and `SelectedRows`. An operator should be able to inference the variable types of its output. | ||
|
|
||
| For example, a `lookup table` operator takes two `LoDTensor`; one is a float tensor as the embedding table, the other is an int tensor as word ID. The gradient operator of `lookup table` will generate a `SelectedRows` as its output. A `sum` operator can take both `LoDTensor` and `SelectedRows` as its inputs and will generate a `LoDTensor` if any of its inputs is `LoDTensor`, otherwise, the `sum` operator will generate `SelectedRows` as its output. |
There was a problem hiding this comment.
In my understanding, SelectedRows is a LoDTensor which stores word embeddings and a LoD which stores the sequence info and a seleted row ids (may be a vector<int>).
the LoDTensor is supported by other ops already.
In my option, the selected row ids is only used by lookup and do not need to be exposed to other ops such as sum, and no necessary to have a new data type that should be processed by ops besides LoDTensor;
There was a problem hiding this comment.
Had a discussion with reyang, the new type SelectedRow serves in scenarios such as sum the outputs of two lookup.
It seems ok and there is no better way to support this currently.
reyoung
changed the title
reyoung
changed the title
reyoung
force-pushed
the
feature/var_type_inferer
branch
from
7ed3ea6
to
f654810
Compare
@wangkuiyi
|
No description provided.