Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add IntermediateLayerGetter #47908

Merged
merged 18 commits into from
Dec 6, 2022
Merged

Add IntermediateLayerGetter #47908

merged 18 commits into from
Dec 6, 2022

Conversation

DrRyanHuang
Copy link
Member

PR types

New features

PR changes

APIs

Describe

Add IntermediateLayerGetter just as torchvision.models._utils.IntermediateLayerGetter

@paddle-bot
Copy link

paddle-bot bot commented Nov 12, 2022

你的PR提交成功,感谢你对开源项目的贡献!
请关注后续CI自动化测试结果,详情请参考Paddle-CI手册
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

@paddle-bot paddle-bot bot added contributor External developers status: proposed labels Nov 12, 2022
@luotao1
Copy link
Contributor

luotao1 commented Nov 14, 2022

需要补充单测来通过 PR-CI-Coverage

@luotao1 luotao1 self-assigned this Nov 14, 2022
@DrRyanHuang DrRyanHuang changed the title Add IntermediateLayerGetter del comment Nov 14, 2022
@DrRyanHuang
Copy link
Member Author

DrRyanHuang commented Nov 14, 2022

@luotao1 意思是提供一个这样的测试用例?

import paddle
m = paddle.vision.models.resnet18(pretrained=True)
# extract layer1 and layer3, giving as names `feat1` and feat2`
new_m = paddle.vision.models.utils.IntermediateLayerGetter(m,
    {'layer1': 'feat1', 'layer3': 'feat2'})
out = new_m(paddle.rand([1, 3, 224, 224]))
print([(k, v.shape) for k, v in out.items()])
[('feat1', [1, 64, 56, 56]), 
 ('feat2', [1, 256, 14, 14])]

@DrRyanHuang DrRyanHuang changed the title del comment Add IntermediateLayerGetter Nov 14, 2022
@@ -12,6 +12,9 @@
# See the License for the specific language governing permissions and
# limitations under the License.

import paddle.nn as nn
from collections import OrderedDict
from typing import Dict

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

#43611 可以参考这个pr增加单测

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

if name in return_layers:
del return_layers[name]
if not return_layers:
break
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这里的layers表示return layer前所有layer?能否给出示例代码

Copy link
Member Author

@DrRyanHuang DrRyanHuang Nov 14, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

是的

IntermediateLayerGetter 的目的就是运行 model 的前向传播,并将中间的某些层的输出拿出来

layers 之所以包括传入 model 的所有 sublayer,是因为要通过传入model 的前向传播

之后执行:

super(IntermediateLayerGetter, self).__init__(layers)

由于继承自 nn.LayerDict 则它包含的子层将被注册和添加

所以在 self.forward 中,能通过 for name, module in self.items(): 来迭代每一层操作

如果某一层的输出,是我们想要的,即if name in self.return_layers,则添加到返回out = OrderedDict()

self.return_layers 字典变量用来指示哪些层的输出需要返回,返回之后的名字叫什么

Examples:
.. code-block:: python

>>> m = paddle.vision.models.resnet18(pretrained=True)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

测试代码需要加import paddle,保证代码可以单独运行

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

2022-11-29 14:26:01 subprocess return code: 1
2022-11-29 14:26:01 Error Raised from Sample Code:
2022-11-29 14:26:01 stderr:   File "samplecode_temp/paddle.vision.models.utils.IntermediateLayerGetter_example.py", line 4
2022-11-29 14:26:01     >>> import paddle
2022-11-29 14:26:01      ^
2022-11-29 14:26:01 SyntaxError: invalid syntax
2022-11-29 14:26:01 
2022-11-29 14:26:01 stdout: 

https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/7275600/job/20709077
示例代码跑不过,把>>>都去掉吧

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done 😆

add `import paddle`
@jerrywgz
Copy link
Contributor

jerrywgz commented Nov 14, 2022

请在提交的代码前使用pre-commit 做下格式检查

pip install pre-commit==2.17.0
pre-commit install
pre-commit run --files python/paddle/vision/models/utils.py

@jerrywgz
Copy link
Contributor

还请下添加单测~

Copy link
Contributor

@luotao1 luotao1 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"/home/Paddle/build/python/paddle/fluid/tests/unittests/test_IntermediateLayerGetter.py", line 18, in <module>
2022-11-25 02:39:14     from paddle.vision.models import IntermediateLayerGetter
2022-11-25 02:39:14 ImportError: cannot import name 'IntermediateLayerGetter' from 'paddle.vision.models' 

现在CI上流水线都挂在这个单测上,请先修一下。需要在 utils.py里加一个__all__ 来加上这个新的API

@luotao1
Copy link
Contributor

luotao1 commented Nov 28, 2022

2022-11-28 14:20:37 	344 - test_IntermediateLayerGetter (Timeout)

PR-CI-Py3 和 PR-CI-Windows-OPENBLAS 上都是 CPU 环境,单测默认时间是15s,目前是超时的。
请减少下单测时间,或设置下单测时长,类似

if(WIN32)
set_tests_properties(test_static_save_load_large PROPERTIES TIMEOUT 900)
set_tests_properties(test_paddle_save_load PROPERTIES TIMEOUT 250)
else()
set_tests_properties(test_static_save_load_large PROPERTIES TIMEOUT 600)

if(NOT WITH_GPU)
   set_tests_properties(test_IntermediateLayerGetter PROPERTIES TIMEOUT xxx)
endif()

@DrRyanHuang
Copy link
Member Author

DrRyanHuang commented Nov 28, 2022

您好,打扰了,这是需要QA的同学的许可吧,我需要给他们发邮件吗?

You must have one QA (XieYunshen(Recommend) or chalsliu) approval for setting parameter RUN_TYPE as EXCLUSIVE, 
DIST, NIGHTLY, EXCLUSIVE:NIGHTLY or DISTNIGHTLY, or setting parameter SERIAL, or setting TIMEOUT properties.

@luotao1
Copy link
Contributor

luotao1 commented Nov 29, 2022

单测时间太长了,特别是windows这条要跑15分钟:

  • Py3流水线的时间140s左右
2022-11-28 22:59:47     Start 344: test_IntermediateLayerGetter
2022-11-28 23:02:11     Test #344: test_IntermediateLayerGetter .....   Passed  143.95 sec
2022-11-28 23:02:11     Start 344: test_IntermediateLayerGetter
2022-11-28 23:04:28     Test #344: test_IntermediateLayerGetter .....   Passed  136.54 sec
2022-11-28 23:04:28     Start 344: test_IntermediateLayerGetter
2022-11-28 23:06:52 1/1 Test #344: test_IntermediateLayerGetter .....   Passed  144.56 sec
  • Windows-openblas流水线的时间840s左右
2022-11-29 02:25:02 Test project C:/home/workspace/Paddle/build
2022-11-29 02:25:02     Start 306: test_IntermediateLayerGetter
2022-11-29 02:39:07 1/1 Test #306: test_IntermediateLayerGetter .....   Passed  844.64 sec

@paddle.no_grad()
def test_inter_result(self):

inp = paddle.randn([32, 3, 224, 224])
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

输入数据减少是否会降低单测时间?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

嗯嗯是的,在AIstudio CPU环境测试,由75s下降至6s 🥳🥳🥳

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

目前的单测时间

  • py3:12s左右
2022-11-29 23:08:58     Start 344: test_IntermediateLayerGetter
2022-11-29 23:09:11     Test #344: test_IntermediateLayerGetter .....   Passed   12.55 sec
2022-11-29 23:09:11     Start 344: test_IntermediateLayerGetter
2022-11-29 23:09:22     Test #344: test_IntermediateLayerGetter .....   Passed   11.40 sec
2022-11-29 23:09:22     Start 344: test_IntermediateLayerGetter
2022-11-29 23:09:34 1/1 Test #344: test_IntermediateLayerGetter .....   Passed   11.31 sec
  • windows-openblas: 50s左右
2022-11-29 19:00:37  325/1433 Test  #298: test_IntermediateLayerGetter ................................   Passed   51.91 sec

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

py3:

2022-11-30 14:25:47     Start 344: test_IntermediateLayerGetter
2022-11-30 14:25:58     Test #344: test_IntermediateLayerGetter .....   Passed   10.94 sec
2022-11-30 14:25:58     Start 344: test_IntermediateLayerGetter
2022-11-30 14:26:05     Test #344: test_IntermediateLayerGetter .....   Passed    7.32 sec
2022-11-30 14:26:05     Start 344: test_IntermediateLayerGetter
2022-11-30 14:26:13 1/1 Test #344: test_IntermediateLayerGetter .....   Passed    7.76 sec

windows-openblas

2022-11-30 17:17:55           Start  307: test_adam_op
2022-11-30 17:17:56  300/1433 Test  #298: test_IntermediateLayerGetter .......   Passed   14.71 sec

终于算是卡着边界通过了😂😂😂

luotao1
luotao1 previously approved these changes Dec 1, 2022
Copy link
Contributor

@luotao1 luotao1 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

jerrywgz
jerrywgz previously approved these changes Dec 1, 2022
@luotao1
Copy link
Contributor

luotao1 commented Dec 1, 2022

@jeff41404 的意见:
torchvision.models._utils.IntermediateLayerGetter 不是公开API,我们是否也写成内部API就行。如果加公开API需要和黑客松一样,在docs repo 中写设计文档并review,同时 paddle.vision.models.utils 这个路径之前是没有的,是否放这里也需要在设计文档中着重review

@jerrywgz
Copy link
Contributor

jerrywgz commented Dec 1, 2022

我们和torchvision保持一致吧,把utils.py改成_utils.py 注意示例代码也做下相应修改

@DrRyanHuang
Copy link
Member Author

DrRyanHuang commented Dec 1, 2022

@jerrywgz
请问下,utils.py 下还有_make_divisible 这个函数

def _make_divisible(v, divisor=8, min_value=None):
    """
    This function ensures that all layers have a channel number that is divisible by divisor
    You can also see at https://github.com/keras-team/keras/blob/8ecef127f70db723c158dbe9ed3268b3d610ab55/keras/applications/mobilenet_v2.py#L505
    Args:
        divisor (int): The divisor for number of channels. Default: 8.
        min_value (int, optional): The minimum value of number of channels, if it is None,
                the default is divisor. Default: None.
    """
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    # Make sure that round down does not go down by more than 10%.
    if new_v < 0.9 * v:
        new_v += divisor
    return new_v

请问 _make_divisible 也要一同迁移吗?还是只是新建一个 _utils.py 放入 IntermediateLayerGetter 即可

@luotao1
Copy link
Contributor

luotao1 commented Dec 1, 2022

请问 _make_divisible 也要一同迁移吗?

因为这个API只在内部使用,可以一同迁移

@DrRyanHuang DrRyanHuang dismissed stale reviews from jerrywgz and luotao1 via 2670de7 December 2, 2022 07:18
@DrRyanHuang
Copy link
Member Author

DrRyanHuang commented Dec 5, 2022

打扰了,还是有一个单测没有通过,请问是我操作不对吗?还是我需要给RD和TPM发邮件得到许可呢 ?

Please find RD for approval first, and then find TPM for approval.

......

2022-12-03 17:21:57 There are 3 approved errors.
2022-12-03 17:21:57 **************************************************************
2022-12-03 17:21:58 API Difference is: 
2022-12-03 17:21:58 + paddle.vision.models._utils.IntermediateLayerGetter (ArgSpec(), ('document', 'f743c6e12bae41b9da3175f725b62106'))
2022-12-03 17:21:58 + paddle.vision.models._utils.IntermediateLayerGetter.forward (ArgSpec(args=['self', 'x'], varargs=None, varkw=None, defaults=None, kwonlyargs=[], kwonlydefaults=None, annotations={}), ('document', 'a8313a8e60085775349df5535ec8174b'))
2022-12-03 17:21:58 + approval_error=6

jerrywgz
jerrywgz previously approved these changes Dec 5, 2022
import paddle
import paddle.nn as nn

__all__ = ["IntermediateLayerGetter"]
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

打扰了,还是有一个单测没有通过,请问是我操作不对吗?还是我需要给RD和TPM发邮件得到许可呢 ?

因为是非公开API,你试试去掉21行后:

  1. 流水线是否能通过
  2. 你能否使用paddle.vision.models._utils.IntermediateLayerGetter

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

Copy link
Contributor

@luotao1 luotao1 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
contributor External developers
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants