【Hackathon 5th No.33】 reshape more dtype and test -part #58764
Conversation
luotao1
changed the title
python/paddle/tensor/manipulation.py
Outdated
| Args: | ||
| x (Tensor): An N-D Tensor. The data type is ``float32``, ``float64``, ``int32``, ``int64`` or ``bool`` | ||
| x (Tensor): An N-D Tensor. The data type is ``float16``, ``float32``, ``float64``, ``int16``, ``int32``, ``int64``, ``uint16``, ``int8``, ``uint8``, ``complex64``, ``complex128``, ``bfloat16`` or ``bool``. |
There was a problem hiding this comment.
no official support for uint16, should delete uint16?
There was a problem hiding this comment.
paddle use bfloat16 to hold uint16, so no official support for uint16?
In [42]: x = paddle.to_tensor(123, dtype='uint16')
In [43]: x
Out[43]:
Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=True,
123.)
In [46]: x = paddle.to_tensor(65535, dtype='uint16')
In [47]: x
Out[47]:
Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=True,
65280.)
| 'int32', | ||
| 'int64', | ||
| 'bool', | ||
| 'uint16', |
There was a problem hiding this comment.
numpy not support bfloat16, numpy uint16 is paddle bfloat16, so bfloat16 should be removed from check_variable_and_dtype, and keep uint16:
python/paddle/tensor/manipulation.py
check_variable_and_dtype(
x,
'x',
[
'float16',
'float32',
'float64',
'int16',
'int32',
'int64',
'bool',
'uint16',
'int8',
'uint8',
'complex64',
'complex128',
],
'reshape',
)But remove uint16 & keep bfloat6 in unittest:
test/legacy_test/test_reshape_op.py
def _test_static_dtype(self):
places = [paddle.CPUPlace()] + (
[paddle.CUDAPlace(0)] if base.core.is_compiled_with_cuda() else []
)
dtypes = [
'float16',
'float32',
'float64',
'int16',
'int32',
'int64',
'int8',
'uint8',
'complex64',
'complex128',
'bfloat16',
'bool',
]
...right?
There was a problem hiding this comment.
Yes, there is a historical issue with function check_variable_and_dtype, which temporarily use uint16 instead of bfloat16 in its parameters. We will make unified revisions in the future. So keep uint16 as parameter of check_variable_and_dtype in python/paddle/tensor/manipulation.py temporarily and use bfloat16 instead of uint16 in unittest
test/legacy_test/test_reshape_op.py
Outdated
| 'int16', | ||
| 'int32', | ||
| 'int64', | ||
| 'uint16', |
|
PR-CI-Codestyle-Check 需要过一下 |
@luotao1 我在本地 pre-commit 是过了的 ～ 我看 ci 日志，不像是格式的问题？ |
megemini
changed the title
luotao1
changed the title
…#58764) * [Change] reshape more dtype and test * [Fix] remove uint16 * [Change] use uint16 instead of bfloat16
…#58764) * [Change] reshape more dtype and test * [Fix] remove uint16 * [Change] use uint16 instead of bfloat16
3 participants
PR types
Function optimization
PR changes
APIs
Description
修改
reshape静态图中数据类型检查，增加支持：并增加相应单元测试。
涉及文件：
python/paddle/tensor/manipulation.py修改数据支持类型test/legacy_test/test_reshape_op.py增加单测关联 PR： #58323