(株)クラスキャットの記事「AutoKeras 1.0 : Tutorials : 画像分類」[1]で, AutoML(Automated Machine Learning)のOSS実装であるAutoKeras[2]がバージョン1.0になったことを知った.
AutoMLは機械学習プロセスの自動化を目的とした技術で, GoogleのCloud AutoMLがよく知られている.
ちなみに, Cloud AutoML Visionは, CODE for YAMATOKORIYAMAで開発を進めている「Kingyo AI Navi」[3]の金魚の品種識別でも使用しているサービスだ.
以前からAutoKerasが開発されていることは知っていたが, 今回バージョンが1.0になったので画像分類を少し試してみることにした.
1. 学習
画像分類を試す場合, よく手書き数字の画像データセット「MNIST」を使用する場合が多いのだが, CNNなどを使用するとaccuracyが0.99以上になるなどモデルの差が分かりにくい.
そこで, 今回は衣料品画像の画像データセット「Fashion-MNIST」[4]を使用して試してみる.
[コード]
from PIL import Image # autokeras, tensorflow import autokeras as ak from tensorflow.keras.datasets import fashion_mnist (train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data() x_train = train_images / 255.0 x_test = test_images / 255.0 y_train = train_labels y_test = test_labels # インスタンス生成 clf = ak.ImageClassifier(num_classes=10, loss='categorical_crossentropy', metrics='accuracy', max_trials=5, directory='./outputs/', objective='val_accuracy', seed=7) # 学習(ベストなモデル, ハイパーパラメータの探索) clf.fit(x_train, y_train, epochs=10) # 予測 predicted_y = clf.predict(x_test) # 評価 print(clf.evaluate(x_test, y_test)) # モデルのExport model = clf.export_model() model.save('./model/fmnist_trial5+epoch10.h5')
さすが, モデル構造やハイパーパラメータなどを自動的に決定するだけあって, コードも非常に簡単だ.
Linux PC(CPU:Core i7 7700, RAM:16GB, GPU:GTX-1060, OS:Ubuntu18.04)で実行すること約20分で一応学習終了...
2. 識別
先の学習済モデルを用いて識別を行う.
[コード]
from PIL import Image # autokeras, tensorflow import autokeras as ak from tensorflow.keras.models import load_model from tensorflow.keras.utils import to_categorical from tensorflow.keras.datasets import fashion_mnist (train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data() test_images = test_images.reshape(test_images.shape + (1,)) x_test = test_images / 255.0 y_test = to_categorical(test_labels) # モデルの読み込み model = load_model('./model/fmnist_trial5+epoch10.h5', custom_objects=ak.CUSTOM_OBJECTS) # 評価 result = model.evaluate(x_test, y_test) print(result)
[実行結果]
(tensorflow2) aska@moonlight:~/work/Fashion-MNIST$ python evaluate_autokeras.py 2020-03-29 00:14:23.270877: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libnvinfer.so.6
:
10000/10000 [==============================] - 1s 125us/sample - loss: 0.2337 - accuracy: 0.9196
[0.23365052340626716, 0.9196]accuracyは0.920で, もっと高い精度がでるかと思っていたが, それほどではなかった.
ちなみに, Keras DocumentationのMnist cnn[5]のサンプルコードでFashion-MNISTを試すと, 0.925であった.
(tensorflow2) aska@moonlight:~/work/Fashion-MNIST$ python mnist_cnn.py
Using TensorFlow backend.
2020-03-28 23:39:10.844111: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libnvinfer.so.6
:
60000/60000 [==============================] - 5s 83us/step - loss: 0.1718 - accuracy: 0.9380 - val_loss: 0.2127 - val_accuracy: 0.9254
Test loss: 0.21271519023776053
Test accuracy: 0.9254000186920166
3. AutoKerasを使う上での注意点
3.1 tf.kerasを使用
kerasのAPIを使用する場合には, tf.kerasを使用する.
[不具合例]
・keras.models.load_modelを利用すると, 以下のようなエラーが発生する.
(tensorflow2) aska@moonlight:~/work/Fashion-MNIST$ python evaluate_autokeras.py 2020-03-29 00:26:45.084879: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libnvinfer.so.6
2020-03-29 00:26:45.085942: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libnvinfer_plugin.so.6
Using TensorFlow backend.
Traceback (most recent call last):
File "evaluate_autokeras.py", line 17, in <module>
model = load_model('./model/fmnist_trial5+epoch10.h5', custom_objects=ak.CUSTOM_OBJECTS)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/engine/saving.py", line 492, in load_wrapper
return load_function(*args, **kwargs)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/engine/saving.py", line 584, in load_model
model = _deserialize_model(h5dict, custom_objects, compile)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/engine/saving.py", line 274, in _deserialize_model
model = model_from_config(model_config, custom_objects=custom_objects)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/engine/saving.py", line 627, in model_from_config
return deserialize(config, custom_objects=custom_objects)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/layers/__init__.py", line 168, in deserialize
printable_module_name='layer')
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/utils/generic_utils.py", line 147, in deserialize_keras_object
list(custom_objects.items())))
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/engine/network.py", line 1056, in from_config
process_layer(layer_data)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/engine/network.py", line 1042, in process_layer
custom_objects=custom_objects)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/layers/__init__.py", line 168, in deserialize
printable_module_name='layer')
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/utils/generic_utils.py", line 149, in deserialize_keras_object
return cls.from_config(config['config'])
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/engine/base_layer.py", line 1179, in from_config
return cls(**config)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
return func(*args, **kwargs)
TypeError: __init__() got an unexpected keyword argument 'ragged'
3.2 autokeras.CUSTOM_OBJECTSの修正
load_modelを使用して学習済モデルを読み込む場合, CUSTOM_OBJECTSの修正が必要.
2020/3/29時点では, issue #1031[6]が反映されてないため, 以下のようなエラーが発生する.
(tensorflow2) aska@moonlight:~/work/Fashion-MNIST$ python evaluate_autokeras.py
2020-03-29 00:38:48.560065: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libnvinfer.so.6
2020-03-29 00:38:48.561123: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libnvinfer_plugin.so.6
Traceback (most recent call last):
File "evaluate_autokeras.py", line 17, in <module>
model = load_model('./model/fmnist_trial5+epoch10.h5', custom_objects=ak.CUSTOM_OBJECTS)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/saving/save.py", line 146, in load_model
return hdf5_format.load_model_from_hdf5(filepath, custom_objects, compile)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/saving/hdf5_format.py", line 168, in load_model_from_hdf5
custom_objects=custom_objects)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/saving/model_config.py", line 55, in model_from_config
return deserialize(config, custom_objects=custom_objects)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/layers/serialization.py", line 106, in deserialize
printable_module_name='layer')
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/utils/generic_utils.py", line 303, in deserialize_keras_object
list(custom_objects.items())))
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/network.py", line 937, in from_config
config, custom_objects)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/network.py", line 1893, in reconstruct_from_config
process_layer(layer_data)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/network.py", line 1875, in process_layer
layer = deserialize_layer(layer_data, custom_objects=custom_objects)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/layers/serialization.py", line 106, in deserialize
printable_module_name='layer')
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/utils/generic_utils.py", line 292, in deserialize_keras_object
config, module_objects, custom_objects, printable_module_name)
File "/home/aska/anaconda3/envs/tensorflow2/lib/python3.7/site-packages/tensorflow_core/python/keras/utils/generic_utils.py", line 250, in class_and_config_for_serialized_keras_object
raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)
ValueError: Unknown layer: Normalization[修正方法]
{autokerasのインストール先}/autokeras/keras_layers.py内のCUSTOM_OBJECTSを以下のように修正する.
CUSTOM_OBJECTS = {
'CategoricalEncoding': CategoricalEncoding,
'Sigmoid': Sigmoid,
'Normalization':preprocessing.Normalization
}
今回, AutoKerasを少し試してみて一応動作は確認できたが, Accuracyだけをみると比較したMnist cnnの方が性能がよかったので, まだAutoKerasのすごさは実感できていない.
AutoKerasでモデル構造やハイパーパラメータを自動決定してくれ, 識別精度もよくなってくるといろいろと使い道がありそうなので, 引き続き調査していく予定!!
----
[1] AutoKeras 1.0 : Tutorials : 画像分類 - TensorFlow 2.x
[2] AutoKeras
[3] 「Kingyo AI Navi」のアプリ化を考える(1)～(6)
[4] zalandoresearch/fashion-mnist: A MNIST-like fashion product database. Benchmark
[5] Mnist cnn - Keras Documentation
[6] fix bug 'ValueError: Unknown layer: Normalization' #1031
