keras2とchainerの使い方をCNNでくらべてみる - 空飛ぶロボットのつくりかた

目的：keras2とchainerの使い方の違いを知る

まとめ：

keras2はmodelの最初の層以外の入力は記述しなくても良い。バックエンドがtheanoとtensorflowで入力の配列が異なる。

chainerはmodelの層追加時、入力と出力の数を記入。入力でNoneと記述すると自動的に計算してくれる。Conv->Linearの際に便利。

参考：

出力ノード数を手計算せずにConvolutionの出力をLinearに入力する（chainer 1.15） - Qiita

わかりやすい↓

Chainer: ビギナー向けチュートリアル Vol.1 - Qiita

わかりやすい↓

Keras tips: 様々な画像の前処理をカンタンにやってくれるkeras.preprocessingのまとめ - MATHGRAM

keras2

mnistをcnnで学習するサンプル

'''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, shuffled and split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()

# backendがtensorflowとtheanoで配列のshapeが異なる
if K.image_data_format() == 'channels_first':
    x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols) # theano
    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) # tensorflow
    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), # 32:出力を与える
                 activation='relu',
                 input_shape=input_shape)) # input_shape = x_train.shape[1:]としてもよい。最初のみ入力の値(img_rows, img_cols, 1)が必要。
model.add(Conv2D(64, (3, 3), activation='relu')) # 畳み込み。32:出力だけ与える
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))　# 全結合相。2:出力だけ与える
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))  # 活性化関数

model.compile(loss=keras.losses.categorical_crossentropy,　# 損失関数
              optimizer=keras.optimizers.Adadelta(),　# 最適化
              metrics=['accuracy'])

model.fit(x_train, y_train,
          batch_size=batch_size,
          epochs=epochs,
          verbose=1,
          validation_data=(x_test, y_test))　# 学習
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])

model.save('./model.hdf5')

参考： keras/mnist_cnn.py at master · fchollet/keras · GitHub

予測するサンプル * 写真を一枚指定しそのクラスを出力

# coding: UTF-8

from __future__ import print_function
import keras
from keras.models import load_model
from keras.preprocessing import image

import numpy as np
import sys


model = load_model('model.hdf5')

if len(sys.argv) != 2:
    print("usage: python [image file]")
    sys.exit(1)

filename = sys.argv[1]

img = image.load_img(filename, target_size=(32, 32))
x = image.img_to_array(img)
#x = x / 255.0 #この操作を行うと正しく予測されない(TBD) 

# rows, cols, channels) -> (samples, rows, cols, channels) 
x = np.expand_dims(x, axis=0)

pred = model.predict(x)
print(pred)

参考： KerasでVGG16を使う - 人工知能に関する断創録

chainer

mnistをcnnで学習するサンプル

# -*- coding: utf-8 -*
import numpy as np
import matplotlib.pylab as plt
from sklearn.datasets import fetch_mldata
import chainer
import chainer.links as L
import chainer.functions as F
from chainer import Chain, optimizers, Variable, serializers
 
mnist = fetch_mldata('MNIST original', data_home='.')
mnist.data = mnist.data.astype(np.float32) # image data 784*70000 [[0-255, 0-255, ...], [0-255, 0-255, ...], ... ]
mnist.data /= 255 # to 0-1
mnist.target = mnist.target.astype(np.int32) # label data 70000
N = 60000
x_train, x_test = np.split(mnist.data,   [N])
t_train, t_test = np.split(mnist.target, [N])
 
# to (n_sample, channel, height, width)
x_train = x_train.reshape((len(x_train), 1, 28, 28))
x_test = x_test.reshape((len(x_test), 1, 28, 28))
 
class CNN(Chain):
    def __init__(self):
        super(CNN, self).__init__(
            conv1 = L.Convolution2D(1, 20, 5), # 入力:1 , 出力:20。両方記述。filter 5
            conv2 = L.Convolution2D(20, 50, 5), # 入力:1 , 出力:20。両方記述。filter 5
            l1 = L.Linear(800, 500),  # ここの入力の計算は以下の参考HPを参照。
            l2 = L.Linear(500, 500),
            l3 = L.Linear(500, 10, initialW=np.zeros((10, 500), dtype=np.float32))
        )
    def forward(self, x):
        h = F.max_pooling_2d(F.relu(self.conv1(x)), 2)
        h = F.max_pooling_2d(F.relu(self.conv2(h)), 2)
        h = F.relu(self.l1(h))
        h = F.relu(self.l2(h))
        h = self.l3(h)
        return h
 
model = CNN()
optimizer = optimizers.Adam() # 最適化関数
optimizer.setup(model)
 
n_epoch = 5
batch_size = 1000
for epoch in range(n_epoch):
    sum_loss = 0
    sum_accuracy = 0
    perm = np.random.permutation(N)
    for i in range(0, N, batch_size):
        x = Variable(x_train[perm[i:i+batch_size]])
        t = Variable(t_train[perm[i:i+batch_size]])
        y = model.forward(x)
        model.zerograds()
        loss = F.softmax_cross_entropy(y, t)  # 活性化関数と損失関数
        acc = F.accuracy(y, t)
        loss.backward()
        optimizer.update()
        sum_loss += loss.data*batch_size
        sum_accuracy += acc.data*batch_size
    print("epoch: {}, mean loss: {}, mean accuracy: {}".format(epoch, sum_loss/N, sum_accuracy/N))
 
cnt = 0
for i in range(10000):
    x = Variable(np.array([x_test[i]], dtype=np.float32))
    t = t_test[i]
    y = model.forward(x)
    y = np.argmax(y.data[0])
    if t == y:
        cnt += 1
 
print("accuracy: {}".format(cnt/10000))

serializers.save_npz("model.npz", model)

予測するサンプル1(TBD)

import numpy as np
%matplotlib inline
import matplotlib.pylab as plt
from sklearn.datasets import fetch_mldata
import chainer
import chainer.links as L
import chainer.functions as F
from chainer import Chain, optimizers, Variable, serializers
 
mnist = fetch_mldata('MNIST original', data_home='.')
mnist.data = mnist.data.astype(np.float32) # image data 784*70000 [[0-255, 0-255, ...], [0-255, 0-255, ...], ... ]
mnist.data /= 255 # to 0-1
mnist.target = mnist.target.astype(np.int32) # label data 70000
N = 60000
x_train, x_test = np.split(mnist.data,   [N])
t_train, t_test = np.split(mnist.target, [N])
 
# to (n_sample, channel, height, width)
x_train = x_train.reshape((len(x_train), 1, 28, 28))
x_test = x_test.reshape((len(x_test), 1, 28, 28))
 
class CNN(Chain):
    def __init__(self):
        super(CNN, self).__init__(
            conv1 = L.Convolution2D(1, 20, 5), # 入力:1 , 出力:20。両方記述。filter 5
            conv2 = L.Convolution2D(20, 50, 5), # 入力:1 , 出力:20。両方記述。filter 5
            l1 = L.Linear(800, 500),　# ここの入力の計算は以下の参考HPを参照。
            l2 = L.Linear(500, 500),
            l3 = L.Linear(500, 10, initialW=np.zeros((10, 500), dtype=np.float32))
        )
    def forward(self, x):
        h = F.max_pooling_2d(F.relu(self.conv1(x)), 2)
        h = F.max_pooling_2d(F.relu(self.conv2(h)), 2)
        h = F.relu(self.l1(h))
        h = F.relu(self.l2(h))
        h = self.l3(h)
        return h

model = CNN()
serializers.load_npz("model.npz", model)

image_data = (TBD)

x = Variable(np.array([image_data], dtype=np.float32))
y = model.forward(x)
y = np.argmax(y.data[0])

予測するサンプル2 : 機械学習フレームワークchainerを使って1からコードをかいてみる(mnist編) - Qiita

def forwardではなくdef callにすると

class MLP(Chain):
    def __init__(self):
        super(MLP, self).__init__(
            l1=L.Linear(784, 100),
            l2=L.Linear(100, 100),
            l3=L.Linear(100, 10),
        )

    def __call__(self, x):
        h1 = F.relu(self.l1(x))
        h2 = F.relu(self.l2(h1))
        y = self.l3(h2)
        return y

以下のようにpredictorで予測することができる

def predict(model, x_data):
    x = Variable(x_data.astype(np.float32))
    y = model.predictor(x)
    return np.argmax(y.data, axis = 1)

答え合わせ。モデルの保存

def predict(model, x_data):
    x = Variable(x_data.astype(np.float32))
    y = model.predictor(x)
    return np.argmax(y.data, axis = 1)

予測サンプル3 : Chainer: ビギナー向けチュートリアル Vol.1 - Qiita

import numpy as np
from chainer import serializers
from chainer.cuda import to_gpu
from chainer.cuda import to_cpu

model = MLP()
serializers.load_npz('mnist_result/model_epoch-10', model)
model.to_gpu(gpu_id)

%matplotlib inline
import matplotlib.pyplot as plt

x, t = test[0]
plt.imshow(x.reshape(28, 28), cmap='gray')
plt.show()
print('label:', t)

x = to_gpu(x[None, ...])
y = model(x)
y = to_cpu(y.data)

print('predicted_label:', y.argmax(axis=1)[0])