HyからKerasを利用してMNISTを分類する
10 July 2017
以前HyからChainerを利用してMNISTを分類する例を上げたので、今度はKerasでやってみる。
参考
インストール
pip install tensorflow-gpu
pip install keras
データの用意とモデル定義
(import [keras.models [Sequential]]
[keras.layers [Dense Dropout Activation]]
[keras.optimizers [Adam]]
[keras.datasets [mnist]]
[keras.utils [np-utils]])
;;; データセット(MNIST)の取得と前処理
(def mnist-dataset (mnist.load-data))
(def X-train (first (first mnist-dataset)))
(def Y-train (second (first mnist-dataset)))
(def X-test (first (second mnist-dataset)))
(def Y-test (second (second mnist-dataset)))
;; 画像を1次元配列化
(setv X-train (.reshape X-train 60000 784)
X-test (.reshape X-test 10000 784))
;; 画素を0.0-1.0の範囲に変換
(setv X-train (.astype X-train "float32")
X-test (.astype X-test "float32"))
(/= X-train 255)
(/= X-test 255)
(print X-train.shape "train samples")
(print X-test.shape "test samples")
;; one-hot-encoding
(setv Y-train (np-utils.to-categorical Y-train 10)
Y-test (np-utils.to-categorical Y-test 10))
;;; モデル定義
(def model (Sequential))
(.add model (Dense :input-dim 784 :units 512))
(.add model (Activation "relu"))
(.add model (Dropout 0.2))
(.add model (Dense :units 512))
(.add model (Activation "relu"))
(.add model (Dropout 0.2))
(.add model (Dense :units 10))
(.add model (Activation "softmax"))
(.compile model
:loss "categorical_crossentropy"
:optimizer (Adam)
:metrics ["accuracy"])
(.summary model)
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 512) 401920
_________________________________________________________________
activation_1 (Activation) (None, 512) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 512) 0
_________________________________________________________________
dense_2 (Dense) (None, 512) 262656
_________________________________________________________________
activation_2 (Activation) (None, 512) 0
_________________________________________________________________
dropout_2 (Dropout) (None, 512) 0
_________________________________________________________________
dense_3 (Dense) (None, 10) 5130
_________________________________________________________________
activation_3 (Activation) (None, 10) 0
=================================================================
Total params: 669,706
Trainable params: 669,706
Non-trainable params: 0
_________________________________________________________________
マクロを使うとモデル定義を多少すっきりさせられる。
(defmacro define-sequential [model layers-def compile-options-dict]
(setv layers
(list (map (fn (elem)
`(.add ~model ~elem))
layers-def)))
`(do (def ~model (Sequential))
~@layers
(.compile ~model ~@compile-options-dict)
~model))
(define-sequential model
;; layer
[(Dense :input-dim 784 :units 512)
(Activation "relu")
(Dropout 0.2)
(Dense :units 512)
(Activation "relu")
(Dropout 0.2)
(Dense :units 10)
(Activation "softmax")]
;; options
{:loss "categorical_crossentropy"
:optimizer (Adam)
:metrics ["accuracy"]})
訓練、テスト、予測
;; 訓練を実行
(def history (.fit model X-train Y-train :nb-epoch 5 :batch-size 100))
;; テスト
(def loss-metrics (.evaluate model X-test Y-test :batch-size 100))
(for [i (range 4)]
(.fit model X-train Y-train :nb-epoch 5 :batch-size 100)
(print "\n" (.evaluate model X-test Y-test :batch-size 100)))
;; 予測(分類結果)
(def classes (.predict-classes model X-test :batch-size 100))
;; 予測(分布)
(def proba (.predict-proba model X-test :batch_size 100))
Epoch 1/10
2017-06-20 13:31:26.460964: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 13:31:26.460990: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 13:31:26.460997: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 13:31:26.461001: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 13:31:26.461007: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 13:31:26.683495: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:893] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2017-06-20 13:31:26.683701: I tensorflow/core/common_runtime/gpu/gpu_device.cc:940] Found device 0 with properties:
name: GeForce GTX 750 Ti
major: 5 minor: 0 memoryClockRate (GHz) 1.0845
pciBusID 0000:01:00.0
Total memory: 1.95GiB
Free memory: 1.60GiB
2017-06-20 13:31:26.683719: I tensorflow/core/common_runtime/gpu/gpu_device.cc:961] DMA: 0
2017-06-20 13:31:26.683724: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] 0: Y
2017-06-20 13:31:26.683733: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 750 Ti, pci bus id: 0000:01:00.0)
60000/60000 [==============================] - 76s - loss: 0.2144 - acc: 0.9345
Epoch 2/10
60000/60000 [==============================] - 6s - loss: 0.1040 - acc: 0.9682
Epoch 3/10
60000/60000 [==============================] - 6s - loss: 0.0821 - acc: 0.9744
Epoch 4/10
60000/60000 [==============================] - 6s - loss: 0.0664 - acc: 0.9789
Epoch 5/10
60000/60000 [==============================] - 6s - loss: 0.0574 - acc: 0.9820
Epoch 6/10
60000/60000 [==============================] - 6s - loss: 0.0523 - acc: 0.9840
Epoch 7/10
60000/60000 [==============================] - 6s - loss: 0.0454 - acc: 0.9858
Epoch 8/10
60000/60000 [==============================] - 6s - loss: 0.0440 - acc: 0.9867
Epoch 9/10
60000/60000 [==============================] - 6s - loss: 0.0388 - acc: 0.9879
Epoch 10/10
60000/60000 [==============================] - 6s - loss: 0.0351 - acc: 0.9890
<keras.callbacks.History object at 0x7f7882ace6d8>
=> (.fit model X-train Y-train :nb-epoch 10 :batch_size 32)
Epoch 1/10
60000/60000 [==============================] - 6s - loss: 0.0341 - acc: 0.9896
Epoch 2/10
60000/60000 [==============================] - 6s - loss: 0.0350 - acc: 0.9897
Epoch 3/10
60000/60000 [==============================] - 6s - loss: 0.0332 - acc: 0.9900
Epoch 4/10
60000/60000 [==============================] - 6s - loss: 0.0326 - acc: 0.9912
Epoch 5/10
60000/60000 [==============================] - 6s - loss: 0.0325 - acc: 0.9907
Epoch 6/10
60000/60000 [==============================] - 6s - loss: 0.0293 - acc: 0.9917
Epoch 7/10
60000/60000 [==============================] - 6s - loss: 0.0277 - acc: 0.9921
Epoch 8/10
60000/60000 [==============================] - 6s - loss: 0.0294 - acc: 0.9919
Epoch 9/10
60000/60000 [==============================] - 6s - loss: 0.0257 - acc: 0.9928
Epoch 10/10
60000/60000 [==============================] - 6s - loss: 0.0291 - acc: 0.9919
<keras.callbacks.History object at 0x7f7893573a20>
