Google Cloud Platform

Googleが提供するクラウドサービスの総称。（https://cloud.google.com）

機械学習データセット

cifar10を使った(ローカルCPUとGPUでの処理速度比べたい)
cifar10:10のクラスにラベル付けされた，50,000枚の32x32訓練用カラー画像，10,000枚のテスト用画像のデータセット．

環境

ローカル

OS:Windows10
エディタ: Atom
Anaconda

GPUのみこれを追加

GCP OS:ubuntu16.04 LS
GPU:Cuda 9.0
device name:Tesla K80

trainingコード

train_cifar10.py


import numpy as np
import matplotlib.pyplot as plt
import keras 
from keras.datasets import cifar10
from keras.models import Sequential
from keras.layers import Conv2D,Activation,MaxPooling2D,Dense,Dropout,Flatten
from keras.optimizers import Adam
from keras.utils import np_utils

# 下のコードでcifar10のデータ取得
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
x_train=np.asarray(x_train).astype("float")/255.0
x_test=np.asarray(x_test).astype("float")/255.0
y_train=np_utils.to_categorical(y_train,10)
y_test=np_utils.to_categorical(y_test,10)
epoch=30
def train_model():
    model=Sequential()
    model.add(Conv2D(32,(3,3),padding="same",input_shape=x_train.shape[1:]))
    model.add(Activation("relu"))
    model.add(Conv2D(64,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(0.25))
        
    model.add(Conv2D(128,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(Conv2D(256,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(0.25))

    model.add(Conv2D(256,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(Conv2D(256,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(0.25))

    model.add(Flatten())
    model.add(Dense(1024))
    model.add(Activation("relu"))
    model.add(Dropout(0.5))
    model.add(Dense(y_train.shape[1]))
    model.add(Activation("softmax"))
    model.compile(loss="categorical_crossentropy",optimizer="Adam",metrics=["accuracy"])
    result=model.fit(x_train,y_train,batch_size=120,epochs=epoch,validation_split=0.2,shuffle=True)

    #評価
　　score=model.evaluate(x_test,y_test,verbose=0)
    print("Test Loss:",score[0])
    print("Test Accuracy:",score[1])

　　#モデルの重み保存(predictやevalusteの時に使う)
    model.save("./cifar10_train.h5")
    return model

if __name__=="__main__":
    train_model()

Training

・GPU(GCP train)

Epoch 1/30
2019-07-28 09:48:47.961033: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0
2019-07-28 09:48:51.145668: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7
50000/50000 [==============================] - 43s 863us/step - loss: 1.7028 - acc: 0.3595 - val_loss: 1.3475 - val_acc: 0.4981
Epoch 2/30
50000/50000 [==============================] - 32s 640us/step - loss: 1.1785 - acc: 0.5752 - val_loss: 0.9698 - val_acc: 0.6595
Epoch 3/30
50000/50000 [==============================] - 32s 639us/step - loss: 0.9198 - acc: 0.6775 - val_loss: 0.7873 - val_acc: 0.7246
Epoch 4/30
50000/50000 [==============================] - 32s 635us/step - loss: 0.7709 - acc: 0.7287 - val_loss: 0.7076 - val_acc: 0.7577
Epoch 5/30
50000/50000 [==============================] - 32s 637us/step - loss: 0.6777 - acc: 0.7638 - val_loss: 0.6583 - val_acc: 0.7758
Epoch 6/30
50000/50000 [==============================] - 32s 642us/step - loss: 0.6102 - acc: 0.7859 - val_loss: 0.6247 - val_acc: 0.7891
Epoch 7/30
50000/50000 [==============================] - 32s 644us/step - loss: 0.5575 - acc: 0.8044 - val_loss: 0.6225 - val_acc: 0.7926
Epoch 8/30
50000/50000 [==============================] - 32s 634us/step - loss: 0.5113 - acc: 0.8218 - val_loss: 0.5550 - val_acc: 0.8114
Epoch 9/30
50000/50000 [==============================] - 32s 640us/step - loss: 0.4731 - acc: 0.8329 - val_loss: 0.5555 - val_acc: 0.8117
Epoch 10/30
50000/50000 [==============================] - 32s 632us/step - loss: 0.4471 - acc: 0.8429 - val_loss: 0.5509 - val_acc: 0.8175
Epoch 11/30
50000/50000 [==============================] - 32s 636us/step - loss: 0.4145 - acc: 0.8558 - val_loss: 0.5768 - val_acc: 0.8170
Epoch 12/30
50000/50000 [==============================] - 32s 643us/step - loss: 0.3883 - acc: 0.8638 - val_loss: 0.5925 - val_acc: 0.8089
Epoch 13/30
50000/50000 [==============================] - 32s 639us/step - loss: 0.3715 - acc: 0.8695 - val_loss: 0.5669 - val_acc: 0.8156
Epoch 14/30
50000/50000 [==============================] - 32s 634us/step - loss: 0.3479 - acc: 0.8750 - val_loss: 0.6267 - val_acc: 0.8086
Epoch 15/30
50000/50000 [==============================] - 32s 641us/step - loss: 0.3352 - acc: 0.8806 - val_loss: 0.5758 - val_acc: 0.8277
Epoch 16/30
50000/50000 [==============================] - 31s 628us/step - loss: 0.3191 - acc: 0.8869 - val_loss: 0.5729 - val_acc: 0.8259
Epoch 17/30
50000/50000 [==============================] - 32s 642us/step - loss: 0.3068 - acc: 0.8912 - val_loss: 0.5528 - val_acc: 0.8331
Epoch 18/30
50000/50000 [==============================] - 32s 639us/step - loss: 0.2950 - acc: 0.8951 - val_loss: 0.5619 - val_acc: 0.8330
Epoch 19/30
50000/50000 [==============================] - 32s 640us/step - loss: 0.2897 - acc: 0.8986 - val_loss: 0.5595 - val_acc: 0.8319
Epoch 20/30
50000/50000 [==============================] - 32s 631us/step - loss: 0.2761 - acc: 0.9039 - val_loss: 0.6063 - val_acc: 0.8251
Epoch 21/30
50000/50000 [==============================] - 32s 631us/step - loss: 0.2736 - acc: 0.9039 - val_loss: 0.5647 - val_acc: 0.8375
Epoch 22/30
50000/50000 [==============================] - 31s 630us/step - loss: 0.2587 - acc: 0.9092 - val_loss: 0.5871 - val_acc: 0.8344
Epoch 23/30
50000/50000 [==============================] - 31s 629us/step - loss: 0.2578 - acc: 0.9109 - val_loss: 0.6191 - val_acc: 0.8303
Epoch 24/30
50000/50000 [==============================] - 31s 627us/step - loss: 0.2420 - acc: 0.9166 - val_loss: 0.5932 - val_acc: 0.8346
Epoch 25/30
50000/50000 [==============================] - 31s 628us/step - loss: 0.2473 - acc: 0.9135 - val_loss: 0.5977 - val_acc: 0.8358
Epoch 26/30
50000/50000 [==============================] - 31s 628us/step - loss: 0.2406 - acc: 0.9168 - val_loss: 0.5824 - val_acc: 0.8373
Epoch 27/30
50000/50000 [==============================] - 32s 630us/step - loss: 0.2355 - acc: 0.9179 - val_loss: 0.6245 - val_acc: 0.8322
Epoch 28/30
50000/50000 [==============================] - 31s 628us/step - loss: 0.2381 - acc: 0.9181 - val_loss: 0.5909 - val_acc: 0.8368
Epoch 29/30
50000/50000 [==============================] - 31s 629us/step - loss: 0.2292 - acc: 0.9211 - val_loss: 0.5889 - val_acc: 0.8408
Epoch 30/30
50000/50000 [==============================] - 32s 633us/step - loss: 0.2221 - acc: 0.9239 - val_loss: 0.5917 - val_acc: 0.8368
Test Loss: 0.591716582775116
Test Accuracy: 0.8368

・CPU
Epoch 1/30
50000/50000 [==============================] - 216s 4ms/step - loss: 1.6905 - acc: 0.3657 - val_loss: 1.3237 - val_acc: 0.5209
Epoch 2/30
50000/50000 [==============================] - 215s 4ms/step - loss: 1.1577 - acc: 0.5846 - val_loss: 1.0158 - val_acc: 0.6408
Epoch 3/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.9370 - acc: 0.6677 - val_loss: 0.8442 - val_acc: 0.7018
Epoch 4/30
50000/50000 [==============================] - 216s 4ms/step - loss: 0.8047 - acc: 0.7174 - val_loss: 0.7127 - val_acc: 0.7519
Epoch 5/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.7146 - acc: 0.7494 - val_loss: 0.6724 - val_acc: 0.7659
Epoch 6/30
50000/50000 [==============================] - 217s 4ms/step - loss: 0.6437 - acc: 0.7736 - val_loss: 0.6371 - val_acc: 0.7783
Epoch 7/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.5817 - acc: 0.7962 - val_loss: 0.6174 - val_acc: 0.7872
Epoch 8/30
50000/50000 [==============================] - 216s 4ms/step - loss: 0.5347 - acc: 0.8120 - val_loss: 0.5697 - val_acc: 0.8086
Epoch 9/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.4989 - acc: 0.8255 - val_loss: 0.5885 - val_acc: 0.8058
Epoch 10/30
50000/50000 [==============================] - 216s 4ms/step - loss: 0.4632 - acc: 0.8362 - val_loss: 0.5526 - val_acc: 0.8129
Epoch 11/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.4324 - acc: 0.8483 - val_loss: 0.5568 - val_acc: 0.8198
Epoch 12/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.4154 - acc: 0.8542 - val_loss: 0.5906 - val_acc: 0.8065
Epoch 13/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.3894 - acc: 0.8634 - val_loss: 0.5491 - val_acc: 0.8212
Epoch 14/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.3727 - acc: 0.8697 - val_loss: 0.5525 - val_acc: 0.8212
Epoch 15/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.3477 - acc: 0.8784 - val_loss: 0.5607 - val_acc: 0.8241
Epoch 16/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.3404 - acc: 0.8804 - val_loss: 0.5476 - val_acc: 0.8287
Epoch 17/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.3208 - acc: 0.8861 - val_loss: 0.5676 - val_acc: 0.8302
Epoch 18/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.3099 - acc: 0.8898 - val_loss: 0.5521 - val_acc: 0.8274
Epoch 19/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.2986 - acc: 0.8927 - val_loss: 0.5752 - val_acc: 0.8226
Epoch 20/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.2948 - acc: 0.8978 - val_loss: 0.5946 - val_acc: 0.8218
Epoch 21/30
50000/50000 [==============================] - 216s 4ms/step - loss: 0.2755 - acc: 0.9036 - val_loss: 0.6167 - val_acc: 0.8211
Epoch 22/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.2726 - acc: 0.9050 - val_loss: 0.5721 - val_acc: 0.8363
Epoch 23/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.2623 - acc: 0.9081 - val_loss: 0.5658 - val_acc: 0.8354
Epoch 24/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.2633 - acc: 0.9080 - val_loss: 0.5759 - val_acc: 0.8347
Epoch 25/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.2575 - acc: 0.9105 - val_loss: 0.6118 - val_acc: 0.8337
Epoch 26/30
50000/50000 [==============================] - 215s 4ms/step - loss: 0.2530 - acc: 0.9121 - val_loss: 0.5677 - val_acc: 0.8370
Epoch 27/30
50000/50000 [==============================] - 222s 4ms/step - loss: 0.2432 - acc: 0.9155 - val_loss: 0.6069 - val_acc: 0.8317
Epoch 28/30
50000/50000 [==============================] - 233s 5ms/step - loss: 0.2387 - acc: 0.9163 - val_loss: 0.5798 - val_acc: 0.8387
Epoch 29/30
50000/50000 [==============================] - 232s 5ms/step - loss: 0.2392 - acc: 0.9163 - val_loss: 0.6137 - val_acc: 0.8265
Epoch 30/30
50000/50000 [==============================] - 230s 5ms/step - loss: 0.2395 - acc: 0.9169 - val_loss: 0.5662 - val_acc: 0.8392
Test Loss: 0.5662494749784469
Test Accuracy: 0.8392

1epochあたりの時間
GPU:30s前後
CPU:220~230s
大体、6倍くらい変わった。

もっといろんなデータを試したい。

追記

別の画像を検証したとき
検証画像

ディレクトリ:下記のコードを書いたディレクトリ(カレントディレクトリ)にpicsというファイルを作成しそこに格納
"./pics/{各々で決定}.jpg"

predict.py


import numpy as np 
from PIL import Image 
import keras,sys,os 
from keras.datasets import cifar10
from keras.models import Sequential,load_model 
from keras.layers import Conv2D,MaxPooling2D,Dropout,Dense,Flatten,Activation
from keras.optimizers import Adam 
from keras.utils import np_utils
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
x_train=np.asarray(x_train).astype("float")/255.0
x_test=np.asarray(x_test).astype("float")/255.0
y_train=np_utils.to_categorical(y_train,10)
y_test=np_utils.to_categorical(y_test,10)
def model_build():
    model=Sequential()
    model.add(Conv2D(32,(3,3),padding="same",input_shape=x_train.shape[1:]))
    model.add(Activation("relu"))
    model.add(Conv2D(64,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(0.25))
        
    model.add(Conv2D(128,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(Conv2D(256,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(0.25))

    model.add(Conv2D(256,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(Conv2D(256,(3,3),padding="same"))
    model.add(Activation("relu"))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(0.25))

    model.add(Flatten())
    model.add(Dense(1024))
    model.add(Activation("relu"))
    model.add(Dropout(0.5))
    model.add(Dense(y_train.shape[1]))
    model.add(Activation("softmax"))

    model.compile(loss="categorical_crossentropy",optimizer="Adam",metrics=["accuracy"])
    model=load_model("cifar10_train.h5")#保存した重みをロード(読み込み)
    return model 

def main():
    X=[]
    img_dir="./pics/airplane.jpg"
    img=Image.open(img_dir)
    img=img.resize((32,32))
    img=np.asarray(img)
    X.append(img)
    X=np.array(X)
    model=model_build()
    result=model.predict(X)[0]
    pre=result.argmax()
    percent=int(result[pre]*100)
    img_pred=model.predict_classes(X)
    print("pred:{0}:{1}%".format(img_pred,percent))

if __name__=="__main__":
    main()

・出力
print(result):[1. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
print("pred:{0}:{1}%".format(img_pred,percent)):pred:[0]:100%

print(result)では各ラベルにおいて入力画像データがどれに対応しているかを出力し、result.argmax()でその中の最大値のみ抜き出した。ここではresult[0]が対応している。

cifar10ではラベル0は飛行機だから正しくすh津力で来ている。