add MobileNetV3 #38653

SigureMo · 2021-12-31T17:54:25Z

PR types

New features

PR changes

APIs

Describe

向 paddle.vision.models 添加以下模型
- MobileNetV3Small
- MobileNetV3Large
- mobilenet_v3_small
- mobilenet_v3_large

Tasks

Performance ^updated

AI Studio 测试详情：https://aistudio.baidu.com/studio/project/partial/verify/3294252/037497a8fe694d89954b1c6e9bf274dd
基准参考：https://github.com/PaddlePaddle/PaddleClas/blob/release/2.2/docs/en/ImageNet_models_en.md

Model	Top-1	Top-5
mobilenet_v3_small_1_0	0.67668₍₌₎	0.87402₍₌₎
mobilenet_v3_large_1_0	0.74042₍₌₎	0.91340₍₌₎

括号中为以 torchvision 性能基准为参考的偏差值

虽然有很多不同 scale 的预训练模型，但这里参考了 paddle 的 mobilenet_v1 和 mobilenet_v2 使用参数 scale 调节 scale，并没有为每个 scale 暴露一个 API，同时也与 torchvision 暴露的 mobilenet_v3_small 和 mobilenet_v3_large 对齐～不过我不太清楚这样做是否合适，如果需要改成类似 PaddleClas 那样每个预训练模型都暴露一个 API 的话我再改一下～

paddle-bot-old · 2021-12-31T17:54:30Z

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

paddle-bot-old · 2022-01-13T02:39:31Z

Sorry to inform you that f747ccd's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

paddle-bot-old · 2022-01-29T02:39:25Z

Sorry to inform you that 3b22137's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

LielinJiang · 2022-02-17T15:31:34Z

python/paddle/vision/models/mobilenetv3.py

+        return paddle.multiply(x=identity, y=x)
+
+
+class ConvBNLayer(nn.Layer):


看看这些类能否复用已有的

ConvBNLayer 看起来是能够复用的，而且随着各式各样 Vision Models 的添加，这种重复也会越来越大，就目前这些模型而言，类似的代码已经出现在了：

inceptionv3.ConvBNLayer

mobilenetv1.ConvBNLayer

mobilenetv2.ConvBNLayer

mobilenetv3.ConvBNLayer (当前模型)

resnext.ConvBNLayer

shufflenetv2.ConvBNLayer

我们是否可以参考 torchvision 里那样，在 paddle.vision.ops 里添加一个容易复用的通用模块，比如像 torchvision 的 ConvNormActivation

https://github.com/pytorch/vision/blob/e13206d9749e81fd8b3aec5e664f697a73febf9f/torchvision/ops/misc.py#L68-L124

这里 Norm 没有限制成 BN 貌似是为了兼容 LN 等 Norm 的模块，比如 ConvNeXt

不过如果要复用同一代码的话，由于现有各个模型内部的 ConvBNLayer 内部参数命名各不相同，所以可能涉及到更新现有权重的问题，整体做起来工程量可能不小。

如果上述方案是可以的话，我可以在本 PR 中实施创建可复用的模块 ConvNormActivation，并在 mobilenetv3 引用它。然后我会在下一个 PR 中在其余 5 个模型中引用它，并逐步更新参数。

@LielinJiang 请问这样可以嘛？

python/paddle/vision/models/mobilenetv3.py

SigureMo · 2022-02-18T14:06:22Z

python/paddle/vision/models/mobilenetv3.py

+        self.avg_pool = nn.AdaptiveAvgPool2D(1)
+        self.conv1 = nn.Conv2D(
+            in_channels=channels,
+            out_channels=channels // reduction,


@LielinJiang

在对代码的修改和权重转换过程中，我发现 paddleclas 中目前的实现在 SE 模块中 squeeze 时通道数是直接除 4，而无论是 keras.applications 还是 torchvision 都在除 4 后进一步使用 _make_divisible 保证其大小能够被 8 整除。比如说 scale=1 时的 mobilenetv3_large 在第一个 SE 模块中第一个卷积时理应输入 72 输出 _make_divisible(72//4)=24，但目前 paddleclas 实现是输出 72//4=18。唔我不太清楚这是不是因为疏漏导致的，但确实和 keras 和 pytorch 实现是不一样的。

这样的话，我们是不是应该放弃使用 paddleclas 里的预训练权重（因为如果修正模型结构的话，SE 那部分 shape 完全对不上），转而使用从 torchvision 现有模型转换的权重呢？不过 torchvision 目前是 small 和 large 都只有 scale=1.0 的权重……比 paddleclas 少不少……

这部分最开始是和gluoncv对齐的。因为当时torch版本的实现无法训练出论文精度。我的建议是按照paddleclas的版本先合入，然后在文档中说明与torchvision版本的不同点

又和之前复现的同学探讨了下，确实clas的版本和其他框架版本都不一样（包括gluoncv）。所以感觉要不使用torchvision的版本吧，先把已有的权重转过来。缺少的权重我们后续再想办法

嗯，那我使用 torchvision 的试下～

好的好的，麻烦了~

TCChenlong

LGTM

XiaoguangHu01

LGTM

qingqing01 · 2022-11-16T05:31:00Z

python/paddle/vision/ops.py

+    """
+    Configurable block used for Convolution-Normalzation-Activation blocks.
+    This code is based on the torchvision code with modifications.
+    You can also see at https://github.com/pytorch/vision/blob/main/torchvision/ops/misc.py#L68


需要修改文档，不能重复，并且去掉参考链接~

「不能重复」是指？另外这个 API 不是一个会对外暴露的 API，目前不会在官网生成文档～

SigureMo mentioned this pull request Dec 31, 2021

add MobileNetV3 cn docs PaddlePaddle/docs#4176

Merged

SigureMo force-pushed the mobilenetv3 branch from 013f498 to ac21bab Compare December 31, 2021 18:39

SigureMo closed this Dec 31, 2021

SigureMo reopened this Dec 31, 2021

SigureMo force-pushed the mobilenetv3 branch 2 times, most recently from 66f44b7 to 4e0dc4b Compare January 1, 2022 05:00

SigureMo force-pushed the mobilenetv3 branch from 4e0dc4b to 33490fc Compare January 1, 2022 07:33

SigureMo force-pushed the mobilenetv3 branch from 33490fc to f747ccd Compare January 3, 2022 18:41

SigureMo force-pushed the mobilenetv3 branch 2 times, most recently from 58623fb to 3b22137 Compare January 21, 2022 18:35

SigureMo force-pushed the mobilenetv3 branch from 3b22137 to c11dbff Compare January 29, 2022 15:23

SigureMo force-pushed the mobilenetv3 branch from c11dbff to 4a59819 Compare February 13, 2022 16:26

SigureMo force-pushed the mobilenetv3 branch from 4a59819 to 7e021b4 Compare February 15, 2022 13:00

TCChenlong requested a review from LielinJiang February 17, 2022 03:11

LielinJiang reviewed Feb 17, 2022

View reviewed changes

python/paddle/vision/models/mobilenetv3.py Show resolved Hide resolved

SigureMo commented Feb 17, 2022

View reviewed changes

python/paddle/vision/models/mobilenetv3.py Outdated Show resolved Hide resolved

SigureMo commented Feb 18, 2022

View reviewed changes

SigureMo dismissed stale reviews from TCChenlong and LielinJiang via c1cbe93 March 8, 2022 07:24

SigureMo added 17 commits March 8, 2022 15:30

add mobilenetv3

60b5802

extract _make_divisible to paddle.vision.models._utils

0754658

extra comma

458bbca

typo

85c577a

using weights converted from torchvision

6d2b127

reformat code

b85c556

update tests

ce82108

trigger ci

948e8b9

trigger ci

aaf4f37

trigger ci

65d755c

pytorch -> torchvision

bd6f10f

wrap the long comments

5abca85

update pretrained weights link

9c81c23

trigger ci

8754f27

cancel the export of ConvNormActivation

b841de5

_utils.py -> utils.py

c2a7a1d

add pretrained model test for mobilenet_v3_large

9d1b054

SigureMo force-pushed the mobilenetv3 branch from c1cbe93 to 9d1b054 Compare March 8, 2022 07:31

TCChenlong approved these changes Mar 9, 2022

View reviewed changes

LielinJiang approved these changes Mar 9, 2022

View reviewed changes

XiaoguangHu01 approved these changes Mar 9, 2022

View reviewed changes

LielinJiang merged commit 68af310 into PaddlePaddle:develop Mar 9, 2022

SigureMo deleted the mobilenetv3 branch March 9, 2022 06:39

SigureMo mentioned this pull request Mar 10, 2022

reuse ConvNormActivation in some vision models #40431

Merged

5 tasks

SigureMo mentioned this pull request Sep 7, 2022

【社区治理】SigureMo 发起 Committer 身份申请 #45841

Closed

qingqing01 reviewed Nov 16, 2022

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add MobileNetV3 #38653

add MobileNetV3 #38653

SigureMo commented Dec 31, 2021 •

edited

Loading

paddle-bot-old bot commented Dec 31, 2021

paddle-bot-old bot commented Jan 13, 2022

paddle-bot-old bot commented Jan 29, 2022

LielinJiang Feb 17, 2022

SigureMo Feb 17, 2022 •

edited

Loading

LielinJiang Feb 18, 2022

SigureMo Feb 18, 2022 •

edited

Loading

LielinJiang Feb 22, 2022

SigureMo Feb 22, 2022

LielinJiang Feb 22, 2022

SigureMo Feb 22, 2022

LielinJiang Feb 22, 2022

TCChenlong left a comment

XiaoguangHu01 left a comment

qingqing01 Nov 16, 2022

SigureMo Nov 16, 2022

		return paddle.multiply(x=identity, y=x)


		class ConvBNLayer(nn.Layer):

add MobileNetV3 #38653

add MobileNetV3 #38653

Conversation

SigureMo commented Dec 31, 2021 • edited Loading

PR types

PR changes

Describe

Tasks

Performance updated

paddle-bot-old bot commented Dec 31, 2021

paddle-bot-old bot commented Jan 13, 2022

paddle-bot-old bot commented Jan 29, 2022

Choose a reason for hiding this comment

SigureMo Feb 17, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SigureMo Feb 18, 2022 • edited Loading

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

TCChenlong left a comment

Choose a reason for hiding this comment

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SigureMo commented Dec 31, 2021 •

edited

Loading

Performance ^updated

SigureMo Feb 17, 2022 •

edited

Loading

SigureMo Feb 18, 2022 •

edited

Loading