// Keras // MNIST를 이용한 계산시간 비교

프랑소와 숄레의 Deep Learning with Python에 나와있는 MNIST 예제를 PC와 Colab으로 계산시간 비교

Source Code     https://github.com/keras-team/keras/blob/master/examples/mnist_cnn.py

keras-team/keras

 

'''Trains a simple convnet on the MNIST dataset.
Gets to 99.25% test accuracy after 12 epochs
(there is still a lot of margin for parameter tuning).
16 seconds per epoch on a GRID K520 GPU.
'''
 
from __future__ import print_function
import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D
from keras import backend as K
 
batch_size = 128
num_classes = 10
epochs = 12
 
# input image dimensions
img_rows, img_cols = 28, 28
 
# the data, split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()
 
if K.image_data_format() == 'channels_first':
    x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
    x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
    input_shape = (1, img_rows, img_cols)
else:
    x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
    x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
    input_shape = (img_rows, img_cols, 1)
 
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
 
# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
 
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
 
model.compile(loss=keras.losses.categorical_crossentropy,
              optimizer=keras.optimizers.Adadelta(),
              metrics=['accuracy'])
 
model.summary()
 
import time
start = time.time()  # 시작 시간 저장
 
model.fit(x_train, y_train,
          batch_size=batch_size,
          epochs=epochs,
          verbose=1,
          validation_data=(x_test, y_test))
 
score = model.evaluate(x_test, y_test, verbose=0)
 
print("time :", time.time() - start)  # 현재시각 - 시작시간 = 실행 시간
print('Test loss:', score[0])
print('Test accuracy:', score[1])

 

PC i7-7700 @ 3.60GHz, 32GB x64 (CPU)

Train on 60000 samples, validate on 10000 samples
Epoch 1/12
60000/60000 [==============================] - 71s 1ms/step - loss: 0.2752 - acc: 0.9160 - val_loss: 0.0547 - val_acc: 0.9826
Epoch 2/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0909 - acc: 0.9730 - val_loss: 0.0396 - val_acc: 0.9859
Epoch 3/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0676 - acc: 0.9797 - val_loss: 0.0325 - val_acc: 0.9887
Epoch 4/12
60000/60000 [==============================] - 68s 1ms/step - loss: 0.0550 - acc: 0.9838 - val_loss: 0.0300 - val_acc: 0.9901
Epoch 5/12
60000/60000 [==============================] - 68s 1ms/step - loss: 0.0476 - acc: 0.9854 - val_loss: 0.0299 - val_acc: 0.9900
Epoch 6/12
60000/60000 [==============================] - 68s 1ms/step - loss: 0.0417 - acc: 0.9876 - val_loss: 0.0273 - val_acc: 0.9911
Epoch 7/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0365 - acc: 0.9887 - val_loss: 0.0262 - val_acc: 0.9915
Epoch 8/12
60000/60000 [==============================] - 71s 1ms/step - loss: 0.0349 - acc: 0.9889 - val_loss: 0.0333 - val_acc: 0.9885
Epoch 9/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0327 - acc: 0.9899 - val_loss: 0.0300 - val_acc: 0.9907
Epoch 10/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0292 - acc: 0.9907 - val_loss: 0.0271 - val_acc: 0.9911
Epoch 11/12
60000/60000 [==============================] - 71s 1ms/step - loss: 0.0279 - acc: 0.9913 - val_loss: 0.0286 - val_acc: 0.9905
Epoch 12/12
60000/60000 [==============================] - 73s 1ms/step - loss: 0.0270 - acc: 0.9920 - val_loss: 0.0277 - val_acc: 0.9912
time : 839.7007076740265
Test loss: 0.02767643234600473
Test accuracy: 0.9912

Colab (GPU)

Train on 60000 samples, validate on 10000 samples
Epoch 1/12
60000/60000 [==============================] - 5s 79us/step - loss: 0.0246 - acc: 0.9924 - val_loss: 0.0320 - val_acc: 0.9903
Epoch 2/12
60000/60000 [==============================] - 5s 76us/step - loss: 0.0238 - acc: 0.9922 - val_loss: 0.0251 - val_acc: 0.9920
Epoch 3/12
60000/60000 [==============================] - 5s 76us/step - loss: 0.0232 - acc: 0.9931 - val_loss: 0.0276 - val_acc: 0.9916
Epoch 4/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0226 - acc: 0.9931 - val_loss: 0.0293 - val_acc: 0.9917
Epoch 5/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0197 - acc: 0.9939 - val_loss: 0.0318 - val_acc: 0.9907
Epoch 6/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0199 - acc: 0.9937 - val_loss: 0.0269 - val_acc: 0.9917
Epoch 7/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0200 - acc: 0.9935 - val_loss: 0.0339 - val_acc: 0.9903
Epoch 8/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0189 - acc: 0.9938 - val_loss: 0.0279 - val_acc: 0.9922
Epoch 9/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0203 - acc: 0.9935 - val_loss: 0.0294 - val_acc: 0.9919
Epoch 10/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0177 - acc: 0.9948 - val_loss: 0.0270 - val_acc: 0.9923
Epoch 11/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0183 - acc: 0.9942 - val_loss: 0.0307 - val_acc: 0.9915
Epoch 12/12
60000/60000 [==============================] - 5s 76us/step - loss: 0.0184 - acc: 0.9941 - val_loss: 0.0261 - val_acc: 0.9919
time : 55.00918984413147
Test loss: 0.02614340115247869
Test accuracy: 0.9919

 

* PC 그래픽카드가 허접한거라 CPU와 Colab GPU를 비교한거지만... 확실히 GPU가 빠르긴하네...

* Colab TPU도 그닥 빠르지는 않은 듯?