add log2 operator #28319

Joejiong · 2020-10-30T02:04:13Z

PR types

New features

PR changes

OPs

Describe

Add new op log2 and its API.
Doc preview:

Unitest:

class TestLog2(TestActivation):
    def setUp(self):
        self.op_type = "log2"
        self.init_dtype()

        x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype)
        out = np.log2(x)

        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
        self.outputs = {'Out': out}

    def test_check_grad(self):
        if self.dtype == np.float16:
            return
        self.check_grad(['X'], 'Out')

    def test_error(self):
        in1 = fluid.layers.data(
            name="in1", shape=[11, 17], append_batch_size=False, dtype="int32")
        in2 = fluid.layers.data(
            name="in2", shape=[11, 17], append_batch_size=False, dtype="int64")

        self.assertRaises(TypeError, fluid.layers.log2, in1)
        self.assertRaises(TypeError, fluid.layers.log2, in2)

    def test_api(self):
        with fluid.program_guard(fluid.Program(), fluid.Program()):
            input_x = np.random.uniform(0.1, 1, [11, 17]).astype("float64")
            data_x = fluid.layers.data(
                name="data_x",
                shape=[11, 17],
                append_batch_size=False,
                dtype="float64")

            out1 = fluid.layers.log2(data_x)
            exe = fluid.Executor(place=fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            res1 = exe.run(fluid.default_main_program(),
                           feed={"data_x": input_x},
                           fetch_list=[out1])
        expected_res = np.log2(input_x)
        self.assertTrue(np.allclose(res1, expected_res))

        # dygraph
        with fluid.dygraph.guard():
            np_x = np.random.uniform(0.1, 1, [11, 17]).astype("float64")
            data_x = fluid.dygraph.to_variable(np_x)
            z = fluid.layers.log2(data_x)
            np_z = z.numpy()
            z_expected = np.array(np.log2(np_x))
        self.assertTrue(np.allclose(np_z, z_expected))

… fix_log2

paddle-bot-old · 2020-10-30T02:04:18Z

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

… fix_log2

zhhsplendid

看到coverage挂了，可以像上周说的log验证那部分代码是否有跑到，如果有，报错到ci ce群

zhhsplendid · 2020-11-02T05:42:18Z

python/paddle/fluid/tests/unittests/test_activation_op.py

@@ -1895,10 +1974,12 @@ def test_errors(self):
            # The input type must be Variable.
            self.assertRaises(TypeError, F.softplus, 1)
            # The input dtype must be float16, float32, float64.
-            x_int32 = paddle.fluid.data(name='x_int32', shape=[12, 10], dtype='int32')
+            x_int32 = paddle.fluid.data(


2.0后用paddle.data是更推荐的用法，可否看看这部分能否改成paddle.data?

这个是Log1p, 我之后换一个pr，统一看看这个activate里面有多少要改的公共的这种需要迁移的，我统一迁移

python/paddle/fluid/tests/unittests/test_activation_op.py

zhhsplendid · 2020-11-02T05:49:42Z

paddle/fluid/operators/activation_op.h

+struct Log2Functor : public BaseActivationFunctor<T> {
+  template <typename Device, typename X, typename Out>
+  void operator()(Device d, X x, Out out) const {
+    out.device(d) = x.log() / static_cast<T>(log(2));


尽管数学上等价，但是计算机应该算以log2为底会更简单快速。比算log(x)/log(2)快。如果有空可以自己写写看这里有没有更快的实现。

这里看看有空调查和实现一下，没空时现在这样也能勉强接受。。

之后实现，谢谢

这个问题，除了性能，还有计算误差的问题，建议再调研下，看是否能优化。

换成tensor原生实现，thx
done；

… fix_log2

Joejiong

这个paddle2.0的activate我之后统一提一个解决，这个pr主要是和log2相关，之后统一迁移，谢谢

Joejiong · 2020-11-02T08:22:25Z

python/paddle/fluid/tests/unittests/test_activation_op.py

@@ -1895,10 +1974,12 @@ def test_errors(self):
            # The input type must be Variable.
            self.assertRaises(TypeError, F.softplus, 1)
            # The input dtype must be float16, float32, float64.
-            x_int32 = paddle.fluid.data(name='x_int32', shape=[12, 10], dtype='int32')
+            x_int32 = paddle.fluid.data(


这个是Log1p, 我之后换一个pr，统一看看这个activate里面有多少要改的公共的这种需要迁移的，我统一迁移

swtkiwi · 2020-11-02T14:21:18Z

python/paddle/fluid/layers/nn.py

+    """
+    :alias_main: paddle.log2
+	:alias: paddle.log2,paddle.tensor.log2,paddle.tensor.math.log2
+	:old_api: paddle.fluid.layers.log2


不需要再写alias_main\alias\old_api这三行了。文档会自动加，辛苦删掉

swtkiwi · 2020-11-02T14:22:48Z

python/paddle/fluid/layers/nn.py

+        Out = \\log_2x
+
+    Args:
+        x (Variable): Input LoDTensor or Tensor. Must be one of the following types: float32, float64.


Variable->Tensor

不推荐暴露LoDTensor，辛苦删掉

swtkiwi · 2020-11-02T14:24:30Z

python/paddle/fluid/layers/nn.py

+
+
+    Returns:
+        Variable: The log to the base 2 of the input LoDTensor or Tensor computed element-wise.


Variable->Tensor
删掉LoDTensor

swtkiwi · 2020-11-02T14:25:20Z

python/paddle/fluid/layers/nn.py

+            exe = paddle.static.Executor(fluid.CPUPlace())
+
+            # Execute
+            x_i = np.array([[1], [2]]).astype(np.float32)


不推荐用numpy的api，最好用paddle的相关api替代

swtkiwi · 2020-11-02T14:25:41Z

python/paddle/fluid/layers/nn.py

+
+            # Execute
+            x_i = np.array([[1], [2]]).astype(np.float32)
+            res_val, = exe.run(fluid.default_main_program(), feed={'x':x_i}, fetch_list=[res])


最好用动态图写示例代码哈

… fix_log2

swtkiwi · 2020-11-04T07:38:11Z

python/paddle/fluid/layers/nn.py

+        .. code-block:: python
+
+            import paddle
+            import paddle.fluid as fluid


删除import paddle.fluid as fluid

swtkiwi · 2020-11-04T07:38:21Z

python/paddle/fluid/layers/nn.py

+
+            # example 1: x is a float
+            x_i = paddle.to_tensor([[1.0], [2.0]])
+            res = paddle.log2(x_i)


增加结果注释

swtkiwi

LGTM

zhhsplendid

lgtm

XiaoguangHu01 · 2020-11-05T06:08:35Z

paddle/fluid/operators/activation_op.h

+struct Log2Functor : public BaseActivationFunctor<T> {
+  template <typename Device, typename X, typename Out>
+  void operator()(Device d, X x, Out out) const {
+    out.device(d) = x.log() / static_cast<T>(log(2));


这个问题，除了性能，还有计算误差的问题，建议再调研下，看是否能优化。

XiaoguangHu01 · 2020-11-05T06:10:15Z

python/paddle/fluid/layers/nn.py

+            x_i = paddle.full(shape=[1], fill_value=2, dtype='float32')
+            paddle.to_tensor(x_i)
+            res = paddle.log2(x_i)
+            print(res.numpy()) # [1.]


这里不需要加.numpy()， print(res)即可

XiaoguangHu01 · 2020-11-05T06:16:29Z

python/paddle/fluid/layers/nn.py

+    if in_dygraph_mode():
+        return core.ops.log2(x)
+
+    check_variable_and_dtype(x, 'x', ['float32', 'float64'], "log2")


XiaoguangHu01 · 2020-11-05T06:22:37Z

python/paddle/fluid/layers/nn.py

@@ -8731,6 +8732,62 @@ def log(x, name=None):
    return out


+def log2(x, name=None):


对于新增的API，源码只需要放到新的paddle目录下即可，不需要放到fluid下，原则上fluid下不新增API了。

… fix_log2

XiaoguangHu01

LGTM

zhhsplendid

LGTM，不过请修改PR标题，你这里是开发log2而不是修复了log2

swtkiwi

LGTM

… fix_log2

TCChenlong

LGTM

XiaoguangHu01

LGTM

Joejiong added 2 commits October 30, 2020 01:59

add new log2 operation

f6d9404

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

f1b4f73

… fix_log2

Joejiong added 7 commits October 30, 2020 03:18

fix sample code

cca2f9a

test fp16

3049700

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

665f827

… fix_log2

fix fp16_error_ratio

eebde3d

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

933dd5d

… fix_log2

fix latex

02fcf16

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

cd94a3a

… fix_log2

zhhsplendid reviewed Nov 2, 2020

View reviewed changes

Joejiong added 2 commits November 2, 2020 08:17

fix paddle2.0 api style

d1838bf

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

b72f42c

… fix_log2

Joejiong commented Nov 2, 2020

View reviewed changes

swtkiwi reviewed Nov 2, 2020

View reviewed changes

Joejiong added 4 commits November 3, 2020 04:59

add dygraph example code

e5a5c26

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

8ae9d2c

… fix_log2

fix doc gen

bfc4d79

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

e08c326

… fix_log2

swtkiwi reviewed Nov 4, 2020

View reviewed changes

clean doc fluid

9eadc71

swtkiwi previously approved these changes Nov 4, 2020

View reviewed changes

zhhsplendid previously approved these changes Nov 4, 2020

View reviewed changes

XiaoguangHu01 reviewed Nov 5, 2020

View reviewed changes

change directory

faedba2

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

f0151d0

… fix_log2

Joejiong dismissed stale reviews from zhhsplendid and swtkiwi via f0151d0 November 5, 2020 08:09

Joejiong added 6 commits November 5, 2020 12:58

optimize log2

bb143ad

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

6ca56f2

… fix_log2

clean code

c8b5fcf

fix float16

23e94c0

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

4b35414

… fix_log2

remove grad_atol

d508c4c

XiaoguangHu01 previously approved these changes Nov 9, 2020

View reviewed changes

zhhsplendid previously approved these changes Nov 9, 2020

View reviewed changes

Joejiong changed the title ~~Fix log2~~ add log2 operator Nov 9, 2020

Joejiong requested a review from swtkiwi November 9, 2020 08:47

swtkiwi previously approved these changes Nov 9, 2020

View reviewed changes

Joejiong added 2 commits November 9, 2020 09:31

fix example code

9a6d1bb

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

c84a99f

… fix_log2

Joejiong dismissed stale reviews from swtkiwi, zhhsplendid, and XiaoguangHu01 via c84a99f November 9, 2020 09:33

Joejiong added 2 commits November 10, 2020 06:27

clean example

c7023f9

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

816a086

… fix_log2

TCChenlong approved these changes Nov 11, 2020

View reviewed changes

XiaoguangHu01 approved these changes Nov 12, 2020

View reviewed changes

zhhsplendid merged commit 08d2413 into PaddlePaddle:develop Nov 12, 2020



		Returns:
		Variable: The log to the base 2 of the input LoDTensor or Tensor computed element-wise.

		@@ -8731,6 +8732,62 @@ def log(x, name=None):
		return out


		def log2(x, name=None):

add log2 operator #28319

add log2 operator #28319

Conversation

Joejiong commented Oct 30, 2020 • edited Loading

PR types

PR changes

Describe

paddle-bot-old bot commented Oct 30, 2020

zhhsplendid left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Joejiong left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

swtkiwi left a comment

Choose a reason for hiding this comment

zhhsplendid left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

zhhsplendid left a comment

Choose a reason for hiding this comment

swtkiwi left a comment

Choose a reason for hiding this comment

TCChenlong left a comment

Choose a reason for hiding this comment

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

Joejiong commented Oct 30, 2020 •

edited

Loading