# はじめに
正式リリース前ですがTensorflow 2.0少しだけ試してみます。
ソースコード
Tensorflow 2.0で変わること
2.0で大きく変わることは
- Eager executionがdefaultになる
- 重複したAPIを統一(
tf.layers
とtf.keras.layers
など) - Contribが一掃される
で、一番の目玉は「Eager executionがdefaultになる」だと思います。
今回はEager executionとkeras実装での学習を比較してみます。
基本的な実装は公式のチュートリアルを参考にしています。
Eager executionとは
Tensorflowはdefine and runという方式で動くライブラリでしたが
Eager executionでは、pytorchやchainerのようにdefine by runで実行されます。
計算グラフを定義しながら実行するdefine by runではモデルを柔軟に定義することができるためRNNなどの実装では好まれていますし、最近のpytorchの勢いからして今後の主流になっていきそうです。
やること
- Tensorflow 2.0.0-alpha0 を使ってみる
- Mnistのサンプルコードを動かす
- Eager executionとkeras実装で学習を比較する
実行環境
- tensorflow-datasets==1.0.1
- tensorflow==2.0.0-alpha0
モデルの定義
from tensorflow.keras.layers import Dense, Flatten, Conv2D
from tensorflow.keras import Model
class MyModel(Model):
def __init__(self):
super(MyModel, self).__init__()
self.conv1 = Conv2D(32, 3, activation='relu')
self.flatten = Flatten()
self.d1 = Dense(128, activation='relu')
self.d2 = Dense(10, activation='softmax')
def call(self, x):
x = self.conv1(x)
x = self.flatten(x)
x = self.d1(x)
return self.d2(x)
モデルは簡単なCNNです。kerasのModelクラスを継承してモデルを定義することができるようになったみたいです。(kerasの少し前のバージョンからできたみたいですが全然気づかなかった)
pytorchやchainerを使ったことがある人は馴染みのある書き方で、個人的にもわかりやすいと思うのでこの書き方に慣れた方がいいと思います。
Trainer
EagerTrainer
class EagerTrainer(object):
def __init__(self):
self.model = MyModel()
self.loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
self.optimizer = tf.keras.optimizers.Adam()
self.train_loss = tf.keras.metrics.Mean(name='train_loss')
self.train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
self.test_loss = tf.keras.metrics.Mean(name='test_loss')
self.test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')
@tf.function
def train_step(self, image, label):
with tf.GradientTape() as tape:
predictions = self.model(image) # 順伝播の計算
loss = self.loss_object(label, predictions) # lossの計算
gradients = tape.gradient(loss, self.model.trainable_variables) # 勾配の計算
self.optimizer.apply_gradients(zip(gradients, self.model.trainable_variables)) # パラメータの更新
self.train_loss(loss)
self.train_accuracy(label, predictions)
@tf.function
def test_step(self, image, label):
predictions = self.model(image)
t_loss = self.loss_object(label, predictions)
self.test_loss(t_loss)
self.test_accuracy(label, predictions)
def train(self, epochs, training_data, test_data):
template = 'Epoch {}, Loss: {:.5f}, Accuracy: {:.5f}, Test Loss: {:.5f}, Test Accuracy: {:.5f}, elapsed_time {:.5f}'
for epoch in range(epochs):
start = time.time()
for image, label in tqdm(training_data):
self.train_step(image, label)
elapsed_time = time.time() - start
for test_image, test_label in test_data:
self.test_step(test_image, test_label)
print(template.format(epoch + 1,
self.train_loss.result(),
self.train_accuracy.result() * 100,
self.test_loss.result(),
self.test_accuracy.result() * 100,
elapsed_time))
EagerTrainer
はEager executionのtrainerです。
コードを見るとだいたいどんな操作をしているかわかると思います。
KerasTrainer
class KerasTrainer(object):
def __init__(self):
self.model = MyModel()
self.model.compile(optimizer=tf.keras.optimizers.Adam(),
loss=tf.keras.losses.SparseCategoricalCrossentropy(),
metrics=['accuracy'])
def train(self, epochs, training_data, test_data):
self.model.fit(training_data, epochs=epochs, validation_data=test_data)
KerasTrainer
はEagerTrainer
と全く同じ学習をするように書いています。
シンプルなモデルだと圧倒的にコードの量が少なくできるのでkerasの偉大さがわかります。
Training
import argparse
import tensorflow as tf
import tensorflow_datasets as tfds
from trainer import EagerTrainer, KerasTrainer
def convert_types(image, label):
image = tf.cast(image, tf.float32)
image /= 255
return image, label
def main():
parser = argparse.ArgumentParser(description='Train Example')
parser.add_argument('--trainer', type=str, default='eager')
args = parser.parse_args()
dataset, info = tfds.load('mnist', with_info=True, as_supervised=True)
mnist_train, mnist_test = dataset['train'], dataset['test']
mnist_train = mnist_train.map(convert_types).shuffle(10000).batch(32)
mnist_test = mnist_test.map(convert_types).batch(32)
if args.trainer.lower() == 'eager':
trainer = EagerTrainer()
else:
trainer = KerasTrainer()
trainer.train(epochs=5, training_data=mnist_train, test_data=mnist_test)
if __name__ == '__main__':
main()
--trainer
のオプションでどちらのTrainerを使うか選べるようにしています。
実際に実行してみるとCPU上では若干Eager executionの方が速い結果になりました。
Tensorflow 1系ではEager executionだと激おそになるという噂がありましたが、改善されているかもしれません。(GPUでもっと実用的なモデルを動かしてみないとわかりませんが)
実行結果(Eager execution)
$ python main.py
2019-03-23 14:27:42.698962: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
0it [00:00, ?it/s]2019-03-23 14:27:43.063703: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875it [00:34, 54.75it/s]
Epoch 1, Loss: 0.13644, Accuracy: 95.91666, Test Loss: 0.06534, Test Accuracy: 97.75000, elapsed_time 34.25085
0it [00:00, ?it/s]2019-03-23 14:28:19.570435: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875it [00:32, 56.85it/s]
Epoch 2, Loss: 0.08982, Accuracy: 97.28416, Test Loss: 0.06335, Test Accuracy: 97.89000, elapsed_time 32.98253
0it [00:00, ?it/s]2019-03-23 14:28:54.483842: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875it [00:34, 75.77it/s]
Epoch 3, Loss: 0.06759, Accuracy: 97.94389, Test Loss: 0.06064, Test Accuracy: 98.03667, elapsed_time 34.02746
0it [00:00, ?it/s]2019-03-23 14:29:30.609697: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875it [00:34, 54.66it/s]
Epoch 4, Loss: 0.05404, Accuracy: 98.35125, Test Loss: 0.05985, Test Accuracy: 98.14000, elapsed_time 34.30574
0it [00:00, ?it/s]2019-03-23 14:30:06.853807: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875it [00:32, 70.69it/s]
Epoch 5, Loss: 0.04531, Accuracy: 98.61067, Test Loss: 0.05936, Test Accuracy: 98.18999, elapsed_time 32.76600
実行結果(keras)
$ python main.py --trainer keras
2019-03-23 14:35:22.782574: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Epoch 1/5
2019-03-23 14:35:23.047235: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875/1875 [==============================] - 49s 26ms/step - loss: 0.1304 - accuracy: 0.9271 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00
Epoch 2/5
2019-03-23 14:36:12.005254: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875/1875 [==============================] - 42s 23ms/step - loss: 0.0415 - accuracy: 0.9861 - val_loss: 0.0530 - val_accuracy: 0.9828
Epoch 3/5
2019-03-23 14:36:54.346357: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875/1875 [==============================] - 42s 22ms/step - loss: 0.0219 - accuracy: 0.9927 - val_loss: 0.0632 - val_accuracy: 0.9811
Epoch 4/5
2019-03-23 14:37:36.479987: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875/1875 [==============================] - 39s 21ms/step - loss: 0.0124 - accuracy: 0.9959 - val_loss: 0.0633 - val_accuracy: 0.9826
Epoch 5/5
2019-03-23 14:38:15.134248: W ./tensorflow/core/framework/model.h:202] Encountered a stop event that was not preceded by a start event.
1875/1875 [==============================] - 39s 21ms/step - loss: 0.0089 - accuracy: 0.9966 - val_loss: 0.0665 - val_accuracy: 0.9836