Add diagflat op, test=develop #33334
Conversation
python/paddle/tensor/creation.py
Outdated
|
|
||
| import paddle | ||
|
|
||
| paddle.disable_static() |
There was a problem hiding this comment.
框架现在默认是动态图运行，这个应该可以去掉。下面也是
| @@ -771,6 +771,125 @@ def meshgrid(*args, **kwargs): | |||
| return out | |||
|
|
|||
|
|
|||
| def diagflat(x, offset=0, padding_value=0, name=None): | |||
| """ | |||
There was a problem hiding this comment.
这个接口和竞品对齐，没有padding_value=0, name=None这2个参数。
There was a problem hiding this comment.
已对齐，并去掉了unittest中涉及padding_value的测试用例。
There was a problem hiding this comment.
为了静态图的便利性，我们的api还是需要加上“name”参数的
python/paddle/tensor/creation.py
Outdated
| return core.ops.diag_v2(y, "offset", offset, "padding_value", | ||
| padding_value) | ||
|
|
||
| check_type(x, 'x', (Variable), 'diag_v2') |
There was a problem hiding this comment.
这是做diagflat的检查，如果检查未通过报错，用户看到的报错信息提示的是diag_v2会无法理解的。改为diagflat.下同
python/paddle/tensor/creation.py
Outdated
|
|
||
| Args: | ||
| x (Tensor): The input tensor. It can be any shape. Its data type should be float32, float64, int32, int64. | ||
| offset (int, optional): The diagonal offset. A positive value represents superdiagonal, 0 represents the main diagonal, and a negative value represents subdiagonal. |
There was a problem hiding this comment.
offset的说明上需要增加：Default: 0 (main diagonal).
| import paddle | ||
| import paddle.fluid as fluid | ||
| from paddle.fluid import core | ||
| from paddle.fluid import Program, program_guard |
There was a problem hiding this comment.
现在不推荐用fluid的API，如果要用program_guard换为paddle.static.program_guard
There was a problem hiding this comment.
已将所有涉及fluid的API替换成 paddle2.1 doc中的API。
| self.outputs = {'Out': self.out} | ||
|
|
||
| def init_config(self): | ||
| pass |
There was a problem hiding this comment.
这个接口没有对应的OP，你设置的self.op_type = "diagflat"是会报错的。不要用OpTest去检查了，换成unittest.TestCase，自己用numpy写一个函数计算期望值，和paddle.diagflat的实际结果做比较。
| self.expected12 = np.diagflat(self.input_np4, k=-1) | ||
|
|
||
| self.input_np5 = np.random.random(size=(2, 3, 4)).astype(np.int32) | ||
| self.expected13 = np.diagflat(self.input_np5, k=-1) |
There was a problem hiding this comment.
- 上面的这些case太重复了，不同dtype的计算逻辑并没有任何区别。留下float64的即可。单测尽量用最小的测试集合覆盖代码中不同分支。
- 如果计算逻辑中，仅有1-D和N-D的区别，那2-D还是3-D的输入也是重复的测试样例，留一种
There was a problem hiding this comment.
1.已修正，只保留了float64数据类型的测试。
2.已修正，只保留了1-D和2-D的测试样例。
| self.assertTrue(np.allclose(res10, self.expected10)) | ||
| self.assertTrue(np.allclose(res11, self.expected11)) | ||
| self.assertTrue(np.allclose(res12, self.expected12)) | ||
| self.assertTrue(np.allclose(res13, self.expected13)) |
There was a problem hiding this comment.
- 静态图还是动态图，底层调用的c++ op都是一样的，只是前端接口逻辑有差异。这里不必在静态图模式下重复验证计算精确度了。
- 可以留下一个静态图的测试样例，以保证代码覆盖率即可
There was a problem hiding this comment.
静态图和动态图在python前端组装算子（flatten+diag）时，代码实现方式不同。如果不对静态图做全面测试，若静态图组装错误，会检查不到。故最终决定留下了静态图的1-D和2-D两组测试用例。
limin2021
force-pushed
the
diagflat-api
branch
2 times, most recently
from
6f53eed to
6a43321
Compare
PR types
New features
PR changes
OPs
Describe
(1) Add diagflat op.
(2) Add unittest for diagflat.