tensorflow/models-v1.12.0中使用ssdlite_mobilenet_v3_large

喜欢ヅ旅行 2022-12-02 03:52 42阅读 0赞

为什么非要执着于tensorflow1.12.0呢，因为可以在cuda9.0上啊。而且经常会分布一些新的模型，不能总去迎合版本吧，毕竟服务器中cuda版本换的话，很麻烦。

## 0.环境 ##

ubuntu16.04
    python3.6
    cuda9.0
    cudnn7.6.1
    tensorflow-gpu==1.12.0
    models==1.12.0

## 1.准备 ##

### 1.1 下载ssdlite\_mobilenet\_v3\_large预训练模型 ###

[https://github.com/tensorflow/models/blob/master/research/object\_detection/g3doc/detection\_model\_zoo.md][https_github.com_tensorflow_models_blob_master_research_object_detection_g3doc_detection_model_zoo.md]

并解压复制到远程服务器我们放模型的路径：

![20200829144604326.png][]

### 1.2 下载models-master、修改自己config项目配置（需要其中的几个文件） ###

（1）将其中的models-master\\research\\object\_detection\\samples\\configs\\ssdlite\_mobilenet\_v3\_large\_320x320\_coco.config复制到Config\\ssdlite\_mobilenet\_v3\_large\_pretrain\_16\_dataAug\_320\_mergedata目录下；

（2）修改ssdlite\_mobilenet\_v3\_large\_320x320\_coco.config文件适配到我们自己的数据（输入数据、类别、预训练模型路径）；

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

![20200829144920895.png][]

## 2.修改 ##

### 2.1 添加文件models中 ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2][]

复制三个文件到models-1.12.0对应的目录下。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3][]

Step2

参考models-master修改/research/object\_detection/builders/model\_builder.py

添加以下模型导入

from object_detection.models.ssd_mobilenet_v3_feature_extractor import SSDMobileNetV3LargeFeatureExtractor
    from object_detection.models.ssd_mobilenet_v3_feature_extractor import SSDMobileNetV3SmallFeatureExtractor

Step3

在/research/object_detection/builders/model_builder.py添加
    
        'ssd_mobilenet_v3_large':
              SSDMobileNetV3LargeFeatureExtractor,
          'ssd_mobilenet_v3_small':
              SSDMobileNetV3SmallFeatureExtractor,

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4][]

### 2.2 修改/models/ssd\_mobilenet\_v3\_feature\_extractor.py等 ###

step1

import tensorflow.compat.v1 as tf
    为
    import tensorflow as tf

![20200829150252353.png][]

Step2

添加

slim = tf.contrib.slim

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 5][]

step3

注释

# import tf_slim as slim

![20200829150454882.png][]

### 2.3 添加mobilenet\_v3文件 ###

Step1

\\research\\slim\\nets\\mobilenet目录下面两个文件复制到对应目录下，将

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 6][]

Step2

修改mobilenet\_v3.py

![20200829150648530.png][]

为：

![20200829150704837.png][]

Step3

添加slim = tf.contrib.slim

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 7][]

Step4：

修改\\research\\slim\\nets\\mobilenet\\conv\_blocks.py文件中squeeze\_excite添加到对应目录下

![20200829150824733.png][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 8][]

修改\\research\\slim\\nets\\mobilenet\\conv\_blocks.py文件中expanded\_conv添加到对应目录下

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 9][]

## 3.训练正常 ##

![20200829151437749.png][]

## 4.推理正常 ##

![20200829151451169.png][]

[https_github.com_tensorflow_models_blob_master_research_object_detection_g3doc_detection_model_zoo.md]: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
[20200829144604326.png]: /images/20221123/ba78f58c3ae84a37a52ce5bae1f4bdae.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221123/f3fb9cca6ac643cb8c4c7a88490eb703.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: /images/20221123/cda5b80404bf48ec8c71e40a724a0662.png
[20200829144920895.png]: /images/20221123/74abe7c9b4394db9b7c7c2c49eae8891.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2]: /images/20221123/92cd2021fc284530b180a80f83252369.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20200829150024824.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4]: /images/20221123/d82cd438fb2e4498bb8a096d744ac035.png
[20200829150252353.png]: /images/20221123/6970b6ed67c547fdaacb95a52331f47e.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 5]: /images/20221123/7da53c3bef5f4026a0ee5fc496aae269.png
[20200829150454882.png]: /images/20221123/658f7ab0abd6464390e0db68ca393e30.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 6]: /images/20221123/cb40a1f3bc3d41a1929df99b1400f0f2.png
[20200829150648530.png]: /images/20221123/b81fbc1bec66497b9cf53957f473095a.png
[20200829150704837.png]: /images/20221123/45e976245a24403c892fa96efdb2d848.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 7]: /images/20221123/9e325adcf4bb43efbf3450f126d9602f.png
[20200829150824733.png]: /images/20221123/f1596e2a48304d808d4c5ecf65c4139b.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 8]: /images/20221123/6483bc41bdcc4d09a695a1b03c725daf.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 9]: /images/20221123/c8e01fd265994f13ab6619792b0b7256.png
[20200829151437749.png]: /images/20221123/b49987753bad46a6a97267fa9a9826a8.png
[20200829151451169.png]: /images/20221123/e3b8e50089e846e9902df6fae1dcb60f.png