如何使用sensitive来确定yolov3 mobilenetv3的剪枝率呀？

[sg-goldrush]

我按照https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.1/static/slim/sensitive 这个指示来做；想要查看yolov3 mobilenetv3的敏感度，但直接报错

我发现只有PaddleDtection2.4以下的才有这个sensitive。我使用PaddlePaddle2.1/2.4 PaddleDetection2.1/2.4 都报这个错误。、 是怎么回事呢？ 就是如何知道yolov3 mobilenetv3的敏感度呀？？ 想得到yolov3 mobilenetv1那种图

[zzjjay]

这里的敏感度分析使用的是静态图代码，其中 ‘from ppdet.data.reader import create_reader’ 调用的是 static/ppdet/data/reader.py 中的函数，由于pip安装的是动态图paddledection，所以会找不到该函数。 解决方法：

• 可以pip uninstall paddledection后重新执行；
• 或者修改sensitive.py中 create_reader的调用路径；

[sg-goldrush]

是从重新安装paddledection版本可以吗 0.5版本？，那对应的paddle是多少版本呢？因为我单独改create_reader，后续很多方法调用路径也有问题。

[zzjjay]

安装0.5版本可以调用create_reader，但对应的paddle版本太低了，可能会有新的问题； 卸载paddledection后，直接执行sensitive.py就行了

[sg-goldrush]

paddlepaddle 2.1。paddleslim2.1 加上 PaddleDetedction工具包就可以执行吗？ 我会有另外的问题，

aistudio@jupyter-2373595-7476238:~/PaddleDetection$ python static/slim/sensitive/sensitive.py -c configs/yolov3/yolov3_mobilenet_v3_270e_voc.yml -p [02-22 11:04:27 MainThread @logger.py:242] Argv: static/slim/sensitive/sensitive.py -c configs/yolov3/yolov3_mobilenet_v3_270e_voc.yml -p [02-22 11:04:27 MainThread @utils.py:79] WRN paddlepaddle version: 2.1.2. The dynamic graph version of PARL is under development, not fully tested and supported Traceback (most recent call last): File "static/slim/sensitive/sensitive.py", line 36, in from paddleslim.prune import sensitivity File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/init.py", line 17, in from paddleslim import prune File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/prune/init.py", line 16, in from .pruner import * File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/prune/pruner.py", line 20, in from ..core import VarWrapper, OpWrapper, GraphWrapper File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/core/init.py", line 19, in from ..core import dygraph File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/core/dygraph.py", line 9, in from ..common import get_logger File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/common/init.py", line 21, in from .server import Server File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/common/server.py", line 28, in from .rl_controller.utils import add_grad, ConnectMessage File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleslim/common/rl_controller/init.py", line 19, in import parl File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/parl/init.py", line 47, in from parl.remote import remote_class, connect File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/parl/remote/init.py", line 19, in from parl.remote.remote_decorator import * File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/parl/remote/remote_decorator.py", line 19, in from parl.remote.remote_wrapper import RemoteWrapper File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/parl/remote/remote_wrapper.py", line 21, in from parl.remote.communication import loads_argument, loads_return,
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/parl/remote/communication.py", line 38, in context = pyarrow.default_serialization_context() File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pyarrow/init.py", line 317, in getattr "module 'pyarrow' has no attribute '{0}'".format(name) AttributeError: module 'pyarrow' has no attribute 'default_serialization_context' aistudio@jupyter-2373595-7476238:~/PaddleDetection$

[sg-goldrush]

能否有一个版本对应能跑出来的吗？就是目标检测网络明感度分析？ 使用paddledetection。

[sg-goldrush]

在剪枝的时候，我是这样配置的剪枝率。因为查不到其敏感度，我根据ppyolo mbv3 large的剪枝率来配置的。您们有关于这部分的配置吗？

[zzjjay]

这个剪枝率应该设置反了吧，0.75太大了，会导致较大的精度损失，应该是0.25比较合理。

我们没有做过ppyolo mbv3 large的剪枝实验，可以按照示例去尝试运行剪枝敏感度分析～

[sg-goldrush]

我就是设置的0.75，只剪枝了6个层，精度最后还行。voc上精度下降2-3%。 敏感度分析就是按照示例分析不出来。。。报错。。。。。。

我看你们文档说敏感度分析的定义，是按照裁剪某一层的10%，然后直接测试其精度，我这边直接像上面那样配置，如果只裁剪某一个卷积层的比如 conv2d_62.w_0的10%，这样配置。直接在原模型上去测试精度，精度几乎都没了.....

python tools/eval.py -c configs/yolov3/yolov3_mobilenet_v3_large_270e_voc.yml --slim_config configs/slim/prune/yolov3_mbv3_prune_fpgm.yml -o weights=yolov3_mobilenet_v3_large_270e_voc.pdparams

[zzjjay]

敏感度分析就是按照示例分析不出来。。。报错。。。。。。

这个报错可以按照上面的修改建议试试呢

直接在原模型上去测试精度，精度几乎都没了.....

可以尝试更换l2-norm指标看看，也可能就是说明该conv比较重要，裁剪容易导致较大精度损失，需要后续重新finetune。