PaddlePaddle

深度学习
文章目录
  1. Docker install
  2. 图像分类
  3. ResNet

在2016百度世界大会上开源的分布式深度学习框架Paddle, 这篇文章是PaddlePaddle的官方文档里关于图像处理方面的学习记录。

Paddle开发文档

看了极客头条的 基于Spark的异构分布式深度学习平台文章特别期待Spark on Paddle, 相关视频

Paddle源码浏览

Docker install

毫无疑问用docker安装最为方便, 由于我的云主机支持AVX, 所以就安装cpu-demo-latest版的Paddle

  1. 首先从 hub.docker上将镜像pull到本地

    sudo docker pull paddledev/paddle:cpu-demo-latest

  2. 创建Paddle容器

    docker run -i -t -d -P --name paddle -v /home/shang:/home/ paddle:shangyan

  3. 在容器中安装anaconda, 然后更新镜像

    docker commit -m"add anaconda" 8fec2af6ccf8

    • 8fec2af6ccf8是容器id

如果是GPU版本的docker, 使用下面的把bash脚本启动容器:

1
2
3
export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
sudo docker run ${CUDA_SO} ${DEVICES} --privileged=true -i -t -d -P --name paddle_gpu_shangyan -v /home/shang:/home/ paddle:shangyan
  • 将我的/home/shang文件夹映射到容器里的/home文件夹, 方便传送数据

图像分类

该图像分类的任务以CIFAR-10为数据集, Image Classification Tutorial-Paddle

分类任务:

  1. 进入paddle/demo/image_classification/data, 运行./download_cifar.sh下载数据, 然后运行./preprocess.sh对数据进行预处理, 成功后:

    root@7b0f9a66f43c:~/paddle/demo/image_classification/data/cifar-out/test# ls
    airplane automobile bird cat deer dog frog horse ship truck

    • preprocess.sh 调用了./demo/image_classification/preprocess.py 去处理下载的原始图片

  2. vgg_16_cifar.py是模型配置文件, 在训练模型之前要定义, 然后运行./train.sh训练模型, 输出一些信息:

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    I1221 09:20:47.534273  1044 Trainer.cpp:170] trainer mode: Normal
    I1221 09:20:47.839715  1044 PyDataProvider2.cpp:257] loading dataprovider image_provider::processData
    [INFO 2016-12-21 09:20:47,855 image_provider.py:51] Image size: 32
    [INFO 2016-12-21 09:20:47,855 image_provider.py:52] Meta path: data/cifar-out/batches/batches.meta
    [INFO 2016-12-21 09:20:47,855 image_provider.py:58] DataProvider Initialization finished
    I1221 09:20:47.855934  1044 PyDataProvider2.cpp:257] loading dataprovider image_provider::processData
    [INFO 2016-12-21 09:20:47,856 image_provider.py:51] Image size: 32
    [INFO 2016-12-21 09:20:47,856 image_provider.py:52] Meta path: data/cifar-out/batches/batches.meta
    [INFO 2016-12-21 09:20:47,856 image_provider.py:58] DataProvider Initialization finished
    I1221 09:20:47.856549  1044 GradientMachine.cpp:134] Initing parameters..
    I1221 09:20:48.485172  1044 GradientMachine.cpp:141] Init parameters done.
    .........
    I1221 09:23:26.772271  1044 TrainerInternal.cpp:165]  Batch=100 samples=12800 AvgCost=2.38281 CurrentCost=2.38281 Eval: classification_error_evaluator=0.8325  CurrentEval: classification_error_evaluator=0.8325
    • 如果没有GPU, 将shell中的use_gpu=0

  3. ./predict.sh预测: 训练好的参数保存在./cifar_vgg_model/

    1
    2
    3
    4
    model=cifar_vgg_model/pass-00175/
    image=data/cifar-out/test/airplane/seaplane_s_000978.png
    use_gpu=1
    python prediction.py $model $image $use_gpu

    输出结果为:

    1
    2
    I1222 00:49:44.455191   464 GradientMachine.cpp:123] Loading parameters from cifar_vgg_model/pass-00175/
    [INFO 2016-12-22 00:49:44,553 prediction.py:137] Label of data/cifar-out/test/airplane/seaplane_s_000978.png is: 2

ResNet

ResNet就不多介绍了, 请看ResNet介绍-Paddle

  1. 运行get_model.sh获得ResNet的paddle模型文件, 如果下载不了移步 百度云, 然后根据shell文件中的命令解压, 完成后的文件目录如下:

    root@7b0f9a66f43c:~/paddle/demo/model_zoo/resnet/model# ls
    mean_meta_224 resnet_101 resnet_152 resnet_50

    • resnet_50: model of 50 layers.
    • resnet_101: model of 101 layers.
    • resnet_152: model of 152 layers.
    • mean_meta_224: mean file with 3 x 224 x 224 size in BGR order. You also can use three mean values: 103.939, 116.779, 123.68.

  2. 可视化ResNet_50网络结构, 运行./net_diagram.sh, 生成ResNet_50.dot文件

    • 使用dot -T<type> -o<outfile> <infile.dot>命令将dot文件转为png, 如下:

    dot -T png -o ResNet_50.png ResNet_50.dot

    生成50层的ResNet网络结构如下:

  3. 用ResNet抽取特征, 运行./extract_fea_py.sh, 最终输出:[INFO 2016-12-21 08:29:45,764 classify.py:219] Done: make image feature batch, 生成features文件夹:

    root@7b0f9a66f43c:~/paddle/demo/model_zoo/resnet/features# ls
    batch_0
    python load_feature.py ./features

    1
    2
    3
    4
    5
    6
    7
    8
    {
    'cat.jpg': {'res5_3_branch2c_conv': array([-0.11962075, -0.10988633, -0.12072118, ..., -0.01200525,
          0.02182791, -0.00304621], dtype=float32), 'res5_3_branch2c_bn': array([-1.33475351, -1.20347679, -1.34959376, ..., -1.50335991,
         -1.01844168, -1.37495327], dtype=float32)},
    'dog.jpg': {'res5_3_branch2c_conv': array([-0.11691939, -0.11028718, -0.08920972, ...,  0.00291708,
          0.0172889 , -0.08754676], dtype=float32), 'res5_3_branch2c_bn': array([-1.29832339, -1.20888269, -0.92463595, ..., -1.28948379,
         -1.08349776, -2.58606839], dtype=float32)}
    }

  4. 对测试图片进行预测分类, ./predict.sh, 最终输出如下:

    1
    2
    3
    4
    5
    6
    7
    [INFO 2016-12-21 08:37:47,909 layers.py:1985] output size for avgpool is 1*1
    I1221 08:37:47.910166   928 Util.cpp:155] commandline:  --use_gpu=1
    I1221 08:37:50.655699   928 Util.cpp:130] Calling runInitFunctions
    I1221 08:37:50.656028   928 Util.cpp:143] Call runInitFunctions done.
    I1221 08:37:52.103798   928 GradientMachine.cpp:123] Loading parameters from model/resnet_50
    [INFO 2016-12-21 08:37:55,102 classify.py:179] Label of example/dog.jpg is: 156
    [INFO 2016-12-21 08:37:57,137 classify.py:179] Label of example/cat.jpg is: 282
    • predict.sh调用了classify.py

具体的命令可以vim这些shell文件查看

打个广告,一起来: