 View on TensorFlow.org
|
 Run in Google Colab
|
 View source on GitHub
|
This tutorial trains a TensorFlow model to classify the CIFAR-10 dataset, and we compile it using XLA.
Load and normalize the dataset using the Keras API:
import tensorflow as tf
# Check that GPU is available: cf. https://colab.research.google.com/notebooks/gpu.ipynb
assert(tf.test.is_gpu_available())
tf.keras.backend.clear_session()
tf.config.optimizer.set_jit(False) # Start with XLA disabled.
def load_data():
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
x_train = x_train.astype('float32') / 256
x_test = x_test.astype('float32') / 256
# Convert class vectors to binary class matrices.
y_train = tf.keras.utils.to_categorical(y_train, num_classes=10)
y_test = tf.keras.utils.to_categorical(y_test, num_classes=10)
return ((x_train, y_train), (x_test, y_test))
(x_train, y_train), (x_test, y_test) = load_data()
WARNING:tensorflow:From <ipython-input-1-f643ac1e83e4>:4: is_gpu_available (from tensorflow.python.framework.test_util) is deprecated and will be removed in a future version.
Instructions for updating:
Use `tf.config.list_physical_devices('GPU')` instead.
Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170500096/170498071 [==============================] - 11s 0us/step
Keras CIFAR-10 예제를 기초로 모델을 정의합니다.
def generate_model():
return tf.keras.models.Sequential([
tf.keras.layers.Conv2D(32, (3, 3), padding='same', input_shape=x_train.shape[1:]),
tf.keras.layers.Activation('relu'),
tf.keras.layers.Conv2D(32, (3, 3)),
tf.keras.layers.Activation('relu'),
tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
tf.keras.layers.Dropout(0.25),
tf.keras.layers.Conv2D(64, (3, 3), padding='same'),
tf.keras.layers.Activation('relu'),
tf.keras.layers.Conv2D(64, (3, 3)),
tf.keras.layers.Activation('relu'),
tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
tf.keras.layers.Dropout(0.25),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(512),
tf.keras.layers.Activation('relu'),
tf.keras.layers.Dropout(0.5),
tf.keras.layers.Dense(10),
tf.keras.layers.Activation('softmax')
])
model = generate_model()
RMSprop 옵티마이저를 사용하여 모델을 훈련합니다.
def compile_model(model):
opt = tf.keras.optimizers.RMSprop(lr=0.0001, decay=1e-6)
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
return model
model = compile_model(model)
def train_model(model, x_train, y_train, x_test, y_test, epochs=25):
model.fit(x_train, y_train, batch_size=256, epochs=epochs, validation_data=(x_test, y_test), shuffle=True)
def warmup(model, x_train, y_train, x_test, y_test):
# Warm up the JIT, we do not wish to measure the compilation time.
initial_weights = model.get_weights()
train_model(model, x_train, y_train, x_test, y_test, epochs=1)
model.set_weights(initial_weights)
warmup(model, x_train, y_train, x_test, y_test)
%time train_model(model, x_train, y_train, x_test, y_test)
scores = model.evaluate(x_test, y_test, verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])
196/196 [==============================] - 5s 12ms/step - loss: 2.1496 - accuracy: 0.1983 - val_loss: 1.8149 - val_accuracy: 0.3625 Epoch 1/25 196/196 [==============================] - 2s 11ms/step - loss: 2.1004 - accuracy: 0.2203 - val_loss: 1.8906 - val_accuracy: 0.3238 Epoch 2/25 196/196 [==============================] - 2s 9ms/step - loss: 1.8013 - accuracy: 0.3475 - val_loss: 1.6799 - val_accuracy: 0.3996 Epoch 3/25 196/196 [==============================] - 2s 9ms/step - loss: 1.6819 - accuracy: 0.3920 - val_loss: 1.5903 - val_accuracy: 0.4248 Epoch 4/25 196/196 [==============================] - 2s 9ms/step - loss: 1.6081 - accuracy: 0.4157 - val_loss: 1.5107 - val_accuracy: 0.4530 Epoch 5/25 196/196 [==============================] - 2s 9ms/step - loss: 1.5503 - accuracy: 0.4380 - val_loss: 1.4521 - val_accuracy: 0.4804 Epoch 6/25 196/196 [==============================] - 2s 9ms/step - loss: 1.4977 - accuracy: 0.4594 - val_loss: 1.4120 - val_accuracy: 0.4896 Epoch 7/25 196/196 [==============================] - 2s 9ms/step - loss: 1.4496 - accuracy: 0.4799 - val_loss: 1.3468 - val_accuracy: 0.5175 Epoch 8/25 196/196 [==============================] - 2s 9ms/step - loss: 1.4125 - accuracy: 0.4945 - val_loss: 1.3819 - val_accuracy: 0.5060 Epoch 9/25 196/196 [==============================] - 2s 9ms/step - loss: 1.3695 - accuracy: 0.5115 - val_loss: 1.3210 - val_accuracy: 0.5385 Epoch 10/25 196/196 [==============================] - 2s 9ms/step - loss: 1.3332 - accuracy: 0.5241 - val_loss: 1.2489 - val_accuracy: 0.5537 Epoch 11/25 196/196 [==============================] - 2s 9ms/step - loss: 1.3050 - accuracy: 0.5356 - val_loss: 1.2619 - val_accuracy: 0.5451 Epoch 12/25 196/196 [==============================] - 2s 9ms/step - loss: 1.2763 - accuracy: 0.5474 - val_loss: 1.2221 - val_accuracy: 0.5702 Epoch 13/25 196/196 [==============================] - 2s 9ms/step - loss: 1.2433 - accuracy: 0.5601 - val_loss: 1.1858 - val_accuracy: 0.5812 Epoch 14/25 196/196 [==============================] - 2s 9ms/step - loss: 1.2230 - accuracy: 0.5686 - val_loss: 1.1337 - val_accuracy: 0.6005 Epoch 15/25 196/196 [==============================] - 2s 9ms/step - loss: 1.1973 - accuracy: 0.5775 - val_loss: 1.1198 - val_accuracy: 0.6048 Epoch 16/25 196/196 [==============================] - 2s 9ms/step - loss: 1.1727 - accuracy: 0.5896 - val_loss: 1.0952 - val_accuracy: 0.6142 Epoch 17/25 196/196 [==============================] - 2s 9ms/step - loss: 1.1454 - accuracy: 0.5988 - val_loss: 1.0855 - val_accuracy: 0.6163 Epoch 18/25 196/196 [==============================] - 2s 9ms/step - loss: 1.1315 - accuracy: 0.6043 - val_loss: 1.0694 - val_accuracy: 0.6237 Epoch 19/25 196/196 [==============================] - 2s 9ms/step - loss: 1.1075 - accuracy: 0.6126 - val_loss: 1.0425 - val_accuracy: 0.6340 Epoch 20/25 196/196 [==============================] - 2s 9ms/step - loss: 1.0865 - accuracy: 0.6180 - val_loss: 1.0308 - val_accuracy: 0.6390 Epoch 21/25 196/196 [==============================] - 2s 9ms/step - loss: 1.0657 - accuracy: 0.6275 - val_loss: 1.0263 - val_accuracy: 0.6370 Epoch 22/25 196/196 [==============================] - 2s 9ms/step - loss: 1.0526 - accuracy: 0.6320 - val_loss: 1.0057 - val_accuracy: 0.6511 Epoch 23/25 196/196 [==============================] - 2s 9ms/step - loss: 1.0388 - accuracy: 0.6359 - val_loss: 1.0015 - val_accuracy: 0.6513 Epoch 24/25 196/196 [==============================] - 2s 10ms/step - loss: 1.0209 - accuracy: 0.6443 - val_loss: 0.9913 - val_accuracy: 0.6491 Epoch 25/25 196/196 [==============================] - 2s 9ms/step - loss: 1.0075 - accuracy: 0.6481 - val_loss: 0.9850 - val_accuracy: 0.6558 CPU times: user 58.9 s, sys: 7.5 s, total: 1min 6s Wall time: 44.5 s 313/313 [==============================] - 1s 3ms/step - loss: 0.9850 - accuracy: 0.6558 Test loss: 0.9849947094917297 Test accuracy: 0.6557999849319458
이제 XLA 컴파일러를 사용하여 모델을 다시 훈련하겠습니다. 애플리케이션 중간에 컴파일러를 활성화하려면 Keras 세션을 재설정해야 합니다.
# We need to clear the session to enable JIT in the middle of the program.
tf.keras.backend.clear_session()
tf.config.optimizer.set_jit(True) # Enable XLA.
model = compile_model(generate_model())
(x_train, y_train), (x_test, y_test) = load_data()
warmup(model, x_train, y_train, x_test, y_test)
%time train_model(model, x_train, y_train, x_test, y_test)
196/196 [==============================] - 6s 14ms/step - loss: 2.1409 - accuracy: 0.1995 - val_loss: 1.8821 - val_accuracy: 0.3151 Epoch 1/25 196/196 [==============================] - 4s 20ms/step - loss: 2.0976 - accuracy: 0.2184 - val_loss: 1.8789 - val_accuracy: 0.3493 Epoch 2/25 196/196 [==============================] - 2s 8ms/step - loss: 1.8061 - accuracy: 0.3482 - val_loss: 1.6674 - val_accuracy: 0.4048 Epoch 3/25 196/196 [==============================] - 2s 8ms/step - loss: 1.6806 - accuracy: 0.3910 - val_loss: 1.6368 - val_accuracy: 0.4117 Epoch 4/25 196/196 [==============================] - 2s 8ms/step - loss: 1.6100 - accuracy: 0.4194 - val_loss: 1.5094 - val_accuracy: 0.4560 Epoch 5/25 196/196 [==============================] - 2s 8ms/step - loss: 1.5538 - accuracy: 0.4379 - val_loss: 1.5283 - val_accuracy: 0.4387 Epoch 6/25 196/196 [==============================] - 2s 8ms/step - loss: 1.5004 - accuracy: 0.4594 - val_loss: 1.4120 - val_accuracy: 0.4909 Epoch 7/25 196/196 [==============================] - 2s 8ms/step - loss: 1.4554 - accuracy: 0.4767 - val_loss: 1.3769 - val_accuracy: 0.4973 Epoch 8/25 196/196 [==============================] - 2s 8ms/step - loss: 1.4165 - accuracy: 0.4896 - val_loss: 1.4065 - val_accuracy: 0.4951 Epoch 9/25 196/196 [==============================] - 2s 8ms/step - loss: 1.3810 - accuracy: 0.5045 - val_loss: 1.3334 - val_accuracy: 0.5228 Epoch 10/25 196/196 [==============================] - 2s 8ms/step - loss: 1.3466 - accuracy: 0.5188 - val_loss: 1.2982 - val_accuracy: 0.5375 Epoch 11/25 196/196 [==============================] - 2s 8ms/step - loss: 1.3145 - accuracy: 0.5296 - val_loss: 1.2729 - val_accuracy: 0.5512 Epoch 12/25 196/196 [==============================] - 2s 9ms/step - loss: 1.2859 - accuracy: 0.5451 - val_loss: 1.2059 - val_accuracy: 0.5742 Epoch 13/25 196/196 [==============================] - 2s 8ms/step - loss: 1.2639 - accuracy: 0.5504 - val_loss: 1.1961 - val_accuracy: 0.5799 Epoch 14/25 196/196 [==============================] - 2s 8ms/step - loss: 1.2353 - accuracy: 0.5621 - val_loss: 1.1558 - val_accuracy: 0.5908 Epoch 15/25 196/196 [==============================] - 2s 8ms/step - loss: 1.2114 - accuracy: 0.5694 - val_loss: 1.2587 - val_accuracy: 0.5546 Epoch 16/25 196/196 [==============================] - 2s 8ms/step - loss: 1.1874 - accuracy: 0.5799 - val_loss: 1.1096 - val_accuracy: 0.6105 Epoch 17/25 196/196 [==============================] - 2s 8ms/step - loss: 1.1668 - accuracy: 0.5879 - val_loss: 1.1292 - val_accuracy: 0.6048 Epoch 18/25 196/196 [==============================] - 2s 8ms/step - loss: 1.1434 - accuracy: 0.5977 - val_loss: 1.0774 - val_accuracy: 0.6197 Epoch 19/25 196/196 [==============================] - 2s 8ms/step - loss: 1.1258 - accuracy: 0.6040 - val_loss: 1.0642 - val_accuracy: 0.6274 Epoch 20/25 196/196 [==============================] - 2s 8ms/step - loss: 1.1079 - accuracy: 0.6107 - val_loss: 1.0274 - val_accuracy: 0.6383 Epoch 21/25 196/196 [==============================] - 2s 8ms/step - loss: 1.0856 - accuracy: 0.6206 - val_loss: 1.0485 - val_accuracy: 0.6329 Epoch 22/25 196/196 [==============================] - 2s 8ms/step - loss: 1.0711 - accuracy: 0.6238 - val_loss: 1.0369 - val_accuracy: 0.6358 Epoch 23/25 196/196 [==============================] - 2s 8ms/step - loss: 1.0565 - accuracy: 0.6293 - val_loss: 1.0801 - val_accuracy: 0.6224 Epoch 24/25 196/196 [==============================] - 2s 8ms/step - loss: 1.0353 - accuracy: 0.6358 - val_loss: 1.1216 - val_accuracy: 0.6056 Epoch 25/25 196/196 [==============================] - 2s 8ms/step - loss: 1.0229 - accuracy: 0.6412 - val_loss: 1.0463 - val_accuracy: 0.6339 CPU times: user 49 s, sys: 6.5 s, total: 55.4 s Wall time: 43.2 s
Titan V GPU 및 Intel Xeon E5-2690 CPU를 탑재한 시스템에서 속도 향상은 약 1.17배입니다.