10
11

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 5 years have passed since last update.

dlib CNNベースの物体検出

Last updated at Posted at 2017-01-17

dlibの物体検出は、簡単に使えるし、素晴らしい性能ですが、時々誤検出が発生することもあります。このブログで紹介されているdeep learningを利用した物体検出は、HOGの検出器よりも性能よさそうなので試してみました。

dnn_mmod_ex.cpp

学習データからモデルをトレーニングするサンプルです。ビルド方法はこちらを参考に。

さて、このサンプルを実行した時、私のようにGPUがしょぼい環境だと、以下のようなエラーが出ると思います。

Error while calling cudaMalloc(&data, new_size*sizeof(float)) in file
/Users/mkisono/work/dlib/dlib/dnn/gpu_data.cpp:191. code: 2, 
reason: out of memory

このサンプルを動かすためには、5GB以上のRAMが必要みたいです。私が使っているiMacのGPUはメモリが1GBしかありません。

/Users/mkisono/NVIDIA_CUDA-8.0_Samples/1_Utilities/deviceQuery/deviceQuery Starting...

CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "GeForce GT 755M"
  CUDA Driver Version / Runtime Version          8.0 / 8.0
  CUDA Capability Major/Minor version number:    3.0
  Total amount of global memory:                 1024 MBytes (1073283072 bytes)
  ( 2) Multiprocessors, (192) CUDA Cores/MP:     384 CUDA Cores
  GPU Max Clock rate:                            1085 MHz (1.09 GHz)
  Memory Clock rate:                             2500 Mhz
  Memory Bus Width:                              128-bit
  L2 Cache Size:                                 262144 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
  Maximum Layered 1D Texture Size, (num) layers  1D=(16384), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(16384, 16384), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  2048
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 1 copy engine(s)
  Run time limit on kernels:                     Yes
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  Device supports Unified Addressing (UVA):      Yes
  Device PCI Domain ID / Bus ID / location ID:   0 / 1 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 8.0, CUDA Runtime Version = 8.0, NumDevs = 1, Device0 = GeForce GT 755M
Result = PASS

バッチサイズを小さくして対応します。cropperで指定している150を5に変更しました(これ以上の値だとエラーになった・・)

dnn_mmod_ex.cpp
    while(trainer.get_learning_rate() >= 1e-4)
    {
        // cropper(150, images_train, face_boxes_train, mini_batch_samples, mini_batch_labels);
        cropper(5, images_train, face_boxes_train, mini_batch_samples, mini_batch_labels);
        // We can also randomly jitter the colors and that often helps a detector
        // generalize better to new images.
        for (auto&& img : mini_batch_samples)
            disturb_colors(img, rnd);

        trainer.train_one_step(mini_batch_samples, mini_batch_labels);
    }

一つのステップで処理するバッチがぐっと小さくなったので、learning rateを下げる閾値は大きくしておきます。この値を変更しないと、直ぐにlearning rateが小さくなってしまい、モデルの学習がうまくいきません。元は 300 だった値を 3000 に変更しました(もっと大きい値でもいいかも)。

dnn_mmod_ex.cpp
    trainer.set_iterations_without_progress_threshold(3000);

これでとりあえず学習は始まりました。とはいえ、どれくらい時間がかかるのか分からないので中断しました。サンプルでは5分ごとにモデルが保存されているので、次に学習を再開した時はそこから続きが出来ます。

% ./dnn_mmod_ex ../faces
num training images: 4
num testing images:  5
detection window width,height:      40,40
overlap NMS IOU thresh:             0.0781701
overlap NMS percent covered thresh: 0.257122
step#: 0     learning rate: 0.1   average loss: 0           steps without apparent progress: 0
step#: 312   learning rate: 0.1   average loss: 3.70172     steps without apparent progress: 81
step#: 625   learning rate: 0.1   average loss: 1.93546     steps without apparent progress: 122
step#: 941   learning rate: 0.1   average loss: 1.72469     steps without apparent progress: 325
step#: 1242  learning rate: 0.1   average loss: 1.6436      steps without apparent progress: 336
step#: 1547  learning rate: 0.1   average loss: 1.55475     steps without apparent progress: 262
step#: 1859  learning rate: 0.1   average loss: 1.55434     steps without apparent progress: 594
step#: 2171  learning rate: 0.1   average loss: 1.52154     steps without apparent progress: 121
Saved state to mmod_sync
step#: 2482  learning rate: 0.1   average loss: 1.41587     steps without apparent progress: 244
step#: 2792  learning rate: 0.1   average loss: 1.30095     steps without apparent progress: 313
step#: 3105  learning rate: 0.1   average loss: 1.13682     steps without apparent progress: 259
step#: 3401  learning rate: 0.1   average loss: 0.979448    steps without apparent progress: 186
step#: 3712  learning rate: 0.1   average loss: 0.906737    steps without apparent progress: 273
step#: 4018  learning rate: 0.1   average loss: 0.809688    steps without apparent progress: 194
step#: 4322  learning rate: 0.1   average loss: 0.781587    steps without apparent progress: 224
Saved state to mmod_sync
step#: 4620  learning rate: 0.1   average loss: 0.727887    steps without apparent progress: 553
step#: 4936  learning rate: 0.1   average loss: 0.654706    steps without apparent progress: 145
step#: 5249  learning rate: 0.1   average loss: 0.588801    steps without apparent progress: 180
step#: 5560  learning rate: 0.1   average loss: 0.580081    steps without apparent progress: 574
step#: 5872  learning rate: 0.1   average loss: 0.599059    steps without apparent progress: 909
step#: 6182  learning rate: 0.1   average loss: 0.504902    steps without apparent progress: 395
step#: 6495  learning rate: 0.1   average loss: 0.537297    steps without apparent progress: 753
step#: 6808  learning rate: 0.1   average loss: 0.539641    steps without apparent progress: 1104
Saved state to mmod_sync
step#: 7118  learning rate: 0.1   average loss: 0.503599    steps without apparent progress: 1350
step#: 7428  learning rate: 0.1   average loss: 0.486274    steps without apparent progress: 578
step#: 7746  learning rate: 0.1   average loss: 0.479272    steps without apparent progress: 892
step#: 8059  learning rate: 0.1   average loss: 0.448152    steps without apparent progress: 548
step#: 8374  learning rate: 0.1   average loss: 0.462102    steps without apparent progress: 519
step#: 8684  learning rate: 0.1   average loss: 0.460537    steps without apparent progress: 1184
step#: 8996  learning rate: 0.1   average loss: 0.474958    steps without apparent progress: 1592
Saved state to mmod_sync
step#: 9312  learning rate: 0.1   average loss: 0.424878    steps without apparent progress: 1453
step#: 9627  learning rate: 0.1   average loss: 0.421029    steps without apparent progress: 86
step#: 9943  learning rate: 0.1   average loss: 0.445149    steps without apparent progress: 956
step#: 10257  learning rate: 0.1   average loss: 0.407989    steps without apparent progress: 1087
step#: 10570  learning rate: 0.1   average loss: 0.44248     steps without apparent progress: 1576
step#: 10884  learning rate: 0.1   average loss: 0.46317     steps without apparent progress: 2187
step#: 11194  learning rate: 0.1   average loss: 0.431704    steps without apparent progress: 2360
step#: 11502  learning rate: 0.1   average loss: 0.404676    steps without apparent progress: 2509

dnn_mmod_face_detection_ex.cpp

顔検出のサンプルです。学習済みのデータをダウンロードして使います。

% ./dnn_mmod_face_detection_ex
Call this program like this:
./dnn_mmod_face_detection_ex mmod_human_face_detector.dat faces/*.jpg

You can get the mmod_human_face_detector.dat file from:
http://dlib.net/files/mmod_human_face_detector.dat.bz2

実行してみると、またメモリ不足。

% ./dnn_mmod_face_detection_ex mmod_human_face_detector.dat ../faces/*.jpg
Error while calling cudaMalloc(&backward_filters_workspace, backward_filters_workspace_size_in_bytes) in file /Users/mkisono/work/dlib/dlib/dnn/cudnn_dlibapi.cpp:948. code: 2, reason: out of memory

諦めて、CUDAを外してサンプルをビルド仕直し(cmakeで -DDLIB_USE_CUDA=OFF を追加)実行しました。
その場合も、画像の拡大率をやや抑えないと実行できませんでした。

dnn_mmod_face_detection_ex.cpp
        while(img.size() < 1000*1000)
            pyramid_up(img);

実行例
スクリーンショット 2017-01-17 16.14.09.png

ちなみに、dnn_mmod_ex.cppで学習させたモデルを dnn_mmod_face_detection_ex.cpp で使う場合は、モデルの定義が違うのでそのままでは動きません。dnn_mmod_face_detection_ex.cpp に記載がある通りです。

    TRAINING THE MODEL
        Finally, users interested in how the face detector was trained should
        read the dnn_mmod_ex.cpp example program.  It should be noted that the
        face detector used in this example uses a bigger training dataset and
        larger CNN architecture than what is shown in dnn_mmod_ex.cpp, but
        otherwise training is the same.  If you compare the net_type statements
        in this file and dnn_mmod_ex.cpp you will see that they are very similar
        except that the number of parameters has been increased.

感想

GPUがあれば非常に高速に処理できるし、性能もバッチリな感じがします。
これがPythonから使えたらどんなに便利か・・・ 対応予定は無さそうですが、モデルを使ったpredictがPythonから出来るだけでもうれしいな。

Pythonから使う (追記)

boost::python 使えばできるかなと思い、試作を始めたところでこれを見つけました。pybind11なるものを使ってPythonバインディングしている先人がおりました。これを参考にしてやってみたら出来ました。
ちなみに、dlibのPythonバインディング自体をpybind11にしようかという話題もあります。
pybind11は初めて使ってみましたが、boost:pythonより良さそうに思いました。

参考リンク pybind11を使ってPythonからC++コードを実行する方法

10
11
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
10
11

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?