【PaddlePaddle Hackathon 2】23、为 Paddle 新增 Softmax2D 组网API #40910
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
请先通过CI噢~ |
|
PR-CI-Windows这个CI是需要通过的哦~ |
|
@dingjiaweiww 你好，已通过 |
|
请注意代码PR中还需要和中文文档关联才可合入，可以参考这个PR：#40472 |
| @@ -0,0 +1,90 @@ | |||
| # Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. | |||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| from __future__ import print_function |
| from test_softmax_op import ref_softmax | ||
|
|
||
| paddle.enable_static() | ||
| np.random.seed(2022) |
| import paddle.fluid.core as core | ||
| from test_softmax_op import ref_softmax | ||
|
|
||
| paddle.enable_static() |
There was a problem hiding this comment.
这个可以放到 def test_static_api(self) 中
| else paddle.CPUPlace() | ||
|
|
||
| def test_extra_repr(self): | ||
| paddle.disable_static(self.place) |
| self.x_np = np.random.uniform(-1, 1, self.shape).astype('float64') | ||
| self.axis = -3 | ||
| self.place = paddle.CPUPlace() | ||
|
|
There was a problem hiding this comment.
增加GPU上单独测试、shape不是3或4维度的错误测试（可以使用assertRaises（error， func，input））
|
|
||
|
|
||
| class TestSoftmax2DAPI(unittest.TestCase): | ||
| # test paddle.nn.Softmax2D |
| # test paddle.nn.Softmax2D | ||
| def setUp(self): | ||
| self.shape = [2, 6, 5, 4] | ||
| self.dtype = 'float64' |
There was a problem hiding this comment.
输入默认需要支持float32，可以增加一个float32的测试
Asthestarsfalll
changed the title
|
已修改并添加中文文档 |
|
PR-CI-Kunlun和PR-CI-Windows-Inference这两个CI需要通过的噢~ |
|
@dingjiaweiww 你好，已通过 |
| x_np = np.random.randn(2, 3, 4, 2, 3) | ||
| x = paddle.to_tensor(x_np, dtype='float64') | ||
| m = paddle.nn.Softmax2D() | ||
| self.assertRaises(AssertionError, m, x) |
There was a problem hiding this comment.
因为代码中使用了assert
assert x.ndim == 3 or x.ndim == 4
There was a problem hiding this comment.
开发者在python层面实现避免输入维度不符合期望使用了assert，因此维度不匹配会触发AssertionError。
| def setUp(self): | ||
| self.shape = [2, 3, 4] | ||
| self.x_np = np.random.uniform(-1, 1, self.shape).astype('float32') | ||
| self.axis = -3 |
There was a problem hiding this comment.
axis可以覆盖不同的值，不需要都是-3，比如0， 正数
There was a problem hiding this comment.
对于softmax2d来说，需要将softmax的axis固定为-3，这里只是显式的写出来了，实际上不需要变动
There was a problem hiding this comment.
这里主要因为 softmax2D 的功能需求是固定了axis=-3的，如果不设置成-3，numpy的实现ref_softmax的就不是实现softmax2D的功能了。那么不能用于和softmax2D的对比测试了。
|
PR格式检查通过，你的PR将接受Paddle专家以及开源社区的review，请及时关注PR动态。 |
|
Sorry to inform you that d0484ba's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
PR types
Others
PR changes
APIs
Describe
解决了issue： #40310
增加了paddle.nn.Softmax2D。具体地，
paddle.nn.Softmax2D()在空间维度计算softmax，从而使输出 tensor 在每个空间维度(channels,hi,wj)的 tensor 求和为1，即表示空间位置上的某个通道向量求和为1。设计文档：PaddlePaddle/community#50
中文文档：PaddlePaddle/docs#4543