Add math operator patches #7688
Conversation
Users can use `a+b`, `a*10`.
|
will this pr handle |
|
|
||
| def monkey_patch_variable(): | ||
| def new_name(): | ||
| return unique_name("tmp") |
There was a problem hiding this comment.
Any more meaningful name?
| tmp_name = new_name() | ||
| var = block.create_var(name=tmp_name, shape=shape, dtype=dtype) | ||
| block.append_op( | ||
| type="fill_constant", |
There was a problem hiding this comment.
Why do we use fill_constant_op in create_tensor while fill_op in create_scalar?
|
|
||
| def astype(self, dtype): | ||
| """ | ||
| Cast a variable to data type. |
There was a problem hiding this comment.
Cast a variable to a specified data type.
| # add fill_op to self.block | ||
| other_var = create_scalar( | ||
| self.block, value=other_var, dtype=lhs_dtype) | ||
|
|
There was a problem hiding this comment.
Maybe we shall add a type check here to make sure the rhs have been cast to Variable correctly. For some users may send np.array to the other_var.
There was a problem hiding this comment.
The create_scalar and create_tensor have checked the value can be cast to float.
| else: | ||
| other_var = create_tensor_with_batchsize( | ||
| self, other_var, lhs_dtype) | ||
| else: |
There was a problem hiding this comment.
Why does the other_var is cast to Variable by different methods in the two branches of if reverse:?
There was a problem hiding this comment.
The reverse operator in Python means the left-hand operand is a Python value and the right-hand operand is a Variable.
In the elementwise-operators, the right-hand operand can be broadcasted. Supposed we write an a - 10, every element in a will be minus 10. However, for 10 - a, we do not implement the broadcast in C++ operators. So we should explicitly broadcast 10 to the same shape of a.
Getting the shape of a variable has two situations.
- The variable is a parameter. Its shape is decided at compile time and not related to the batch size. So we use
fill_constant. - The variable is a layer output. Its shape is decided at runtime and related to the batch size. So we use
fill_constant_batch_size_like.
There was a problem hiding this comment.
Great idea! Impressive.
No, it won't. But following this PR, we can overload all operators in Python. We can give an API just like numpy. |
3 participants
Users can use
a+b,a*10.