Add intepolte_v2 for Paddle2.0 #26520

tink2123 · 2020-08-21T03:46:29Z

PR types

Performance optimization

PR changes

APIs

Describe

Update interpolate API

Support list in scale_factor
Modify parameter default values

新增interpolate_v2 原因：
修改了scale_factor参数的类型，没有办法和之前的OP做到兼容。旧的模型参数无法被新的OP加载，因此增加一个V2 版本。

paddle-bot-old · 2020-08-21T03:46:34Z

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

willthefrog · 2020-08-21T11:19:39Z

python/paddle/nn/functional/common.py

-              input : (N,C,H_in,W_in)
-              output: (N,C,H_out,W_out) where:
-              H_out = round(H_{in} * scale_{factor})
-              W_out = round(W_{in} * scale_{factor})

        Bilinear interpolation:


remove :alias from docstring

Variable -> Tensor

update code samples to 2.0 API

willthefrog · 2020-08-21T11:22:46Z

python/paddle/nn/functional/common.py

@@ -327,7 +322,10 @@ def interpolate(input,

    if align_mode != 0 and align_mode != 1:


add fast path for dygraph mode

willthefrog · 2020-08-21T11:23:28Z

python/paddle/nn/functional/common.py

@@ -327,7 +322,10 @@ def interpolate(input,

    if align_mode != 0 and align_mode != 1:
        raise ValueError("align_mode can only be 0 or 1")
-
+    if align_corners != 0 and resample == 'NEAREST':
+        raise ValueError(


also enforce in v2 op.

willthefrog · 2020-08-21T11:23:52Z

python/paddle/nn/functional/common.py

        else:
            raise TypeError(
-                "Attr(scale)'s type should be float, int or Variable.")
-
+                "Attr(scale)'s type should be float, int, list or Variable.")


Variable -> Tensor

willthefrog · 2020-08-21T11:24:47Z

python/paddle/nn/functional/common.py

+    https://en.wikipedia.org/wiki/Trilinear_interpolation.
+
+    Parameters:
+        input (Variable): 3-D, 4-D or 5-D Tensor, its data type is float32, float64, or uint8,


willthefrog · 2020-08-21T11:27:37Z

python/paddle/nn/functional/common.py

+        .. code-block:: python
+            import paddle
+            import numpy as np
+            import paddle.fluid.dygraph as dg


update to 2.0 API

willthefrog · 2020-08-21T11:27:58Z

python/paddle/nn/layer/common.py

+    This op upsamples a batch of images, using nearest neighbours' pixel values.
+    The input must be a 4-D Tensor of the shape (num_batches, channels, in_h, in_w), 
+    and the upsampling only applies on the two dimensions(height and width).
+    **Warning:** the parameter :attr:`actual_shape` will be deprecated in the


willthefrog · 2020-08-21T11:28:25Z

python/paddle/nn/layer/common.py

+                # [2L, 3L, 12L, 12L]
+    """
+
+    def __init__(self, size=None, scale_factor=None, data_format='NCHW'):


willthefrog · 2020-08-21T11:28:35Z

python/paddle/nn/layer/common.py

+        self.mode = mode.lower()
+        self.data_format = data_format
+
+    def forward(self, input):


input -> x

willthefrog · 2020-08-21T11:28:58Z

python/paddle/nn/functional/common.py

        inputs=inputs,
        outputs={"Out": out},
        attrs=attrs)
    return out


+def upsample(input,


input -> x

Above problems has been fixed, please review again

…nto interpolate

LDOUBLEV · 2020-08-24T09:45:36Z

python/paddle/nn/functional/common.py

+
+            input_data = np.random.rand(2,3,6,10).astype("float32")
+            x = paddle.to_tensor(input_data)
+            output = F.interpolate(x=x, size=[12,12])


是否需要多举个例子说明不同mode的区别

willthefrog · 2020-08-25T02:26:03Z

python/paddle/nn/functional/common.py

+    interpolating functions of three variables (e.g. D-direction,
+    H-direction and W-direction in this op) on a rectilinear 3D grid.
+    The linear interpolation is performed on three directions.
+    Align_corners and align_mode are optional parameters,the calculation method


Align -> align

jzhang533

lgtm

willthefrog

LGTM

jzhang533

lgtm

XiaoguangHu01

LGTM

luotao1 · 2020-08-27T07:31:26Z

2020-08-27 14:24:06 1. It is an Op accuracy problem, please take care of it. You must have one RD (juncaipeng (Recommend), zhangting2020 (Recommend) or luotao1) approval for the python/paddle/fluid/tests/unittests/white_list/op_accuracy_white_list.py, which manages the white list of upgrading the precision of op test to float64. For more information, please refer to: https://github.com/PaddlePaddle/Paddle/wiki/Upgrade-OP-Precision-to-Float64. 
2020-08-27 14:24:06 2. It is an Op accuracy problem, please take care of it. You must have one RD (juncaipeng (Recommend), zhangting2020 or luotao1) approval for the python/paddle/fluid/tests/unittests/white_list/op_threshold_white_list.py, which manages the white list of error threshold for op test with float64 precision. For more information, please refer to: https://github.com/PaddlePaddle/Paddle/wiki/Upgrade-OP-Precision-to-Float64. 
2020-08-27 14:24:06 5. Please use the default precision parameters of 'atol, rtol, eps, max_relative_error'. If you don't use the default value, you must have one RD (Xreki (Recommend), luotao1, lanxianghit or phlrain) approval for the usage of other values. The detailed information is in the link: https://github.cor/PaddlePaddle/Paddle/wiki/OP-test-accuracy-requirements. The error line is 
2020-08-27 14:24:06 +            self.check_output(atol=1e-7)
2020-08-27 14:24:06 +            self.check_output(atol=1e-5)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 +            self.check_output_with_place(place=core.CPUPlace(), atol=1e-7)
2020-08-27 14:24:06 +            self.check_output_with_place(place=core.CPUPlace(), atol=1e-5)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 There are 6 approved errors.
2020-08-27 14:24:06 ****************

@XiaoguangHu01 @lanxianghit 因为这个OP不符合精度规范，请问可以豁免么？
@tink2123 可以说下原因

tink2123 · 2020-08-27T07:40:02Z

2020-08-27 14:24:06 1. It is an Op accuracy problem, please take care of it. You must have one RD (juncaipeng (Recommend), zhangting2020 (Recommend) or luotao1) approval for the python/paddle/fluid/tests/unittests/white_list/op_accuracy_white_list.py, which manages the white list of upgrading the precision of op test to float64. For more information, please refer to: https://github.com/PaddlePaddle/Paddle/wiki/Upgrade-OP-Precision-to-Float64. 
2020-08-27 14:24:06 2. It is an Op accuracy problem, please take care of it. You must have one RD (juncaipeng (Recommend), zhangting2020 or luotao1) approval for the python/paddle/fluid/tests/unittests/white_list/op_threshold_white_list.py, which manages the white list of error threshold for op test with float64 precision. For more information, please refer to: https://github.com/PaddlePaddle/Paddle/wiki/Upgrade-OP-Precision-to-Float64. 
2020-08-27 14:24:06 5. Please use the default precision parameters of 'atol, rtol, eps, max_relative_error'. If you don't use the default value, you must have one RD (Xreki (Recommend), luotao1, lanxianghit or phlrain) approval for the usage of other values. The detailed information is in the link: https://github.cor/PaddlePaddle/Paddle/wiki/OP-test-accuracy-requirements. The error line is 
2020-08-27 14:24:06 +            self.check_output(atol=1e-7)
2020-08-27 14:24:06 +            self.check_output(atol=1e-5)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 +            self.check_output_with_place(place=core.CPUPlace(), atol=1e-7)
2020-08-27 14:24:06 +            self.check_output_with_place(place=core.CPUPlace(), atol=1e-5)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 There are 6 approved errors.
2020-08-27 14:24:06 ****************

@XiaoguangHu01 @lanxianghit 因为这个OP不符合精度规范，请问可以豁免么？
@tink2123 可以说下原因

这个问题应该是因为 numpy 和 c++ 在非整数计算插值的时候精度达不到要求。这次没有改旧的OP，之前的版本一直就在白名单里，新版本也需要加上才能过单测。对比过pytorch和paddle的输出结果，精度误差在1e-6可以对齐。

XiaoguangHu01 · 2020-08-27T08:51:44Z

2020-08-27 14:24:06 1. It is an Op accuracy problem, please take care of it. You must have one RD (juncaipeng (Recommend), zhangting2020 (Recommend) or luotao1) approval for the python/paddle/fluid/tests/unittests/white_list/op_accuracy_white_list.py, which manages the white list of upgrading the precision of op test to float64. For more information, please refer to: https://github.com/PaddlePaddle/Paddle/wiki/Upgrade-OP-Precision-to-Float64. 
2020-08-27 14:24:06 2. It is an Op accuracy problem, please take care of it. You must have one RD (juncaipeng (Recommend), zhangting2020 or luotao1) approval for the python/paddle/fluid/tests/unittests/white_list/op_threshold_white_list.py, which manages the white list of error threshold for op test with float64 precision. For more information, please refer to: https://github.com/PaddlePaddle/Paddle/wiki/Upgrade-OP-Precision-to-Float64. 
2020-08-27 14:24:06 5. Please use the default precision parameters of 'atol, rtol, eps, max_relative_error'. If you don't use the default value, you must have one RD (Xreki (Recommend), luotao1, lanxianghit or phlrain) approval for the usage of other values. The detailed information is in the link: https://github.cor/PaddlePaddle/Paddle/wiki/OP-test-accuracy-requirements. The error line is 
2020-08-27 14:24:06 +            self.check_output(atol=1e-7)
2020-08-27 14:24:06 +            self.check_output(atol=1e-5)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 +            self.check_output_with_place(place=core.CPUPlace(), atol=1e-7)
2020-08-27 14:24:06 +            self.check_output_with_place(place=core.CPUPlace(), atol=1e-5)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 +        self.check_output_with_place(place=core.CPUPlace(), atol=1)
2020-08-27 14:24:06 There are 6 approved errors.
2020-08-27 14:24:06 ****************

@XiaoguangHu01 @lanxianghit 因为这个OP不符合精度规范，请问可以豁免么？
@tink2123 可以说下原因

这个问题应该是因为 numpy 和 c++ 在非整数计算插值的时候精度达不到要求。这次没有改旧的OP，之前的版本一直就在白名单里，新版本也需要加上才能过单测。对比过pytorch和paddle的输出结果，精度误差在1e-6可以对齐。

确认可以先加入白名单。跟之前的op保持一致。@luotao1

raindrops2sea

增补一个pr注释说明一下新增v2 op的考量。

luotao1

LGTM

add intepolte_v2

0258262

tink2123 added 2 commits August 21, 2020 05:09

fix linear interp

4dc8c40

polish unittest, test=develop

b69fe4e

willthefrog requested changes Aug 21, 2020

View reviewed changes

tink2123 and others added 5 commits August 24, 2020 08:42

update code samples to 2.0 API, test=develop

e2e48ba

Merge branch 'develop' into interpolate_v2

44daf83

remove warning, test_develop

59f9568

Merge branch 'interpolate_v2' of https://github.com/tink2123/Paddle i…

55e6fd3

…nto interpolate

add name in attrs, test=develop

8a44f23

LDOUBLEV reviewed Aug 24, 2020

View reviewed changes

tink2123 added 2 commits August 24, 2020 11:20

polish code, test=develop

8498f66

merge develop, test=develop

25da67a

tink2123 force-pushed the interpolate_v2 branch from 9c80d4f to 25da67a Compare August 24, 2020 12:30

willthefrog reviewed Aug 25, 2020

View reviewed changes

willthefrog requested review from willthefrog and jzhang533 August 25, 2020 02:27

tink2123 added 2 commits August 25, 2020 02:53

change Align to align, test=develop

2d14177

fix unittest in py3,test=develop

6caa20c

jzhang533 previously approved these changes Aug 25, 2020

View reviewed changes

willthefrog previously approved these changes Aug 25, 2020

View reviewed changes

willthefrog self-requested a review August 25, 2020 11:30

fix coverage, test=develop

0e003f2

tink2123 dismissed stale reviews from willthefrog and jzhang533 via 0e003f2 August 25, 2020 11:51

tink2123 added 3 commits August 26, 2020 03:53

fix coverage, test=develop

2e048c9

fix for windows ci, test=develop

e99f7d5

fix coverage, test=develop

7d64df2

tink2123 requested a review from jzhang533 August 27, 2020 04:58

willthefrog approved these changes Aug 27, 2020

View reviewed changes

jzhang533 approved these changes Aug 27, 2020

View reviewed changes

XiaoguangHu01 approved these changes Aug 27, 2020

View reviewed changes

ceci3 requested a review from raindrops2sea August 27, 2020 09:14

raindrops2sea approved these changes Aug 27, 2020

View reviewed changes

luotao1 approved these changes Aug 27, 2020

View reviewed changes

ceci3 merged commit 89d7d86 into PaddlePaddle:develop Aug 27, 2020

tink2123 changed the title ~~add intepolte_v2~~ Add intepolte_v2 for Paddle2.0 Oct 19, 2020

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add intepolte_v2 for Paddle2.0 #26520

Add intepolte_v2 for Paddle2.0 #26520

tink2123 commented Aug 21, 2020 •

edited

Loading

paddle-bot-old bot commented Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

willthefrog Aug 21, 2020

tink2123 Aug 24, 2020

LDOUBLEV Aug 24, 2020

tink2123 Aug 24, 2020

willthefrog Aug 25, 2020

tink2123 Aug 25, 2020

jzhang533 left a comment

willthefrog left a comment

jzhang533 left a comment

XiaoguangHu01 left a comment

luotao1 commented Aug 27, 2020

tink2123 commented Aug 27, 2020

XiaoguangHu01 commented Aug 27, 2020

raindrops2sea left a comment

luotao1 left a comment

		@@ -327,7 +322,10 @@ def interpolate(input,

		if align_mode != 0 and align_mode != 1:

Add intepolte_v2 for Paddle2.0 #26520

Add intepolte_v2 for Paddle2.0 #26520

Conversation

tink2123 commented Aug 21, 2020 • edited Loading

PR types

PR changes

Describe

paddle-bot-old bot commented Aug 21, 2020

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

jzhang533 left a comment

Choose a reason for hiding this comment

willthefrog left a comment

Choose a reason for hiding this comment

jzhang533 left a comment

Choose a reason for hiding this comment

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

luotao1 commented Aug 27, 2020

tink2123 commented Aug 27, 2020

XiaoguangHu01 commented Aug 27, 2020

raindrops2sea left a comment

Choose a reason for hiding this comment

luotao1 left a comment

Choose a reason for hiding this comment

tink2123 commented Aug 21, 2020 •

edited

Loading