assign_value operator #7371
assign_value operator #7371
Conversation
paddle/operators/assign_value_op.cc
Outdated
| } | ||
|
|
||
| protected: | ||
| framework::OpKernelType GetActualKernelType( |
There was a problem hiding this comment.
The function name should be GetExpectedKernelType.
Related doc: https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/kernel_hint_design.md#final-choice
There was a problem hiding this comment.
Fixed.
The interface changed before I submitted the PR.
| framework::proto::DataType::FP32}); | ||
| AddAttr<std::vector<float>>("fp32_values", "store the float values") | ||
| .SetDefault({}); | ||
| AddAttr<std::vector<int>>("int32_values", "store the int values") |
There was a problem hiding this comment.
Op's attribute will be stored into the protobuf. However, the size of a protobuf object is limited(64M). I'm afraid there might be some problems if the tensor to be assigned is huge.
An alternative way is writing the tensor values to a file in Python, then the assign_value_op read the file and fill them into output variable.
There was a problem hiding this comment.
Yes. If the arry is too big, we may have to use a separate file. But if many cases where only a small matrix is used, it's quite inconvenient to use a separate file to array the matrix.
emailweixu
force-pushed
the
assign_value_op
branch
from
ed42430
to
9108785
Compare
paddle/operators/assign_value_op.h
Outdated
| break; | ||
| } | ||
| auto values = ctx.Attr<std::vector<T>>(value_name); | ||
| Copy(dst, values.data(), sizeof(T) * values.size(), ctx); |
There was a problem hiding this comment.
We have a uniform Copy interface of https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor_util.h#L101
We need this operator to assign value to a tensor and the values are stored in the program so that they can be used independent of python.
emailweixu
force-pushed
the
assign_value_op
branch
from
9108785
to
25ecd20
Compare
We need this operator to assign value to a tensor and the values are stored in the program so that they can be used independent of python.