2019年9月15日 - 流行の科学技術を勝手に追いかけるブログ人工知能仮想現実ブロックチェーン

Web Application: 第６回　VGG16を構築する

VGG16をやってみます。

オックスフォード大学のVisual Geometry Groupが作った、畳み込み13層＋全結合３層の　合計16層のニューラルネットワークです。

VGG16用の入力データを作成

VGG16の入力データの画像サイズは 224 x 224 である必要がありますので　generate_inputdata.pyを改造して224×224のnpyデータを生成するgenerate_inputdata_224.pyを作りましょう。

また、それだと元々150角サイズが224角サイズになるわけですから、容量が増えてしまいます。（実際に600MBから1300MBに増えました。）

そこで、このスクリプトでは正規化するのをやめて、浮動小数点が発生するのを防ぎ容量を落とします。

代わりに、VGG16をやる直前に正規化を行う方針でいきます。

スクリプト：generate_inputdata_224.py

変更した部分は

入力サイズ変更：image_size = 224
保存名を変更：np.save(‘./image_files_224.npy’, xy)
標準化中止　：data = data / 255.0 の行をコメントアウト

from PIL import Image # image operation
import os             # get file list
import glob           # image treatment
import numpy as np    # culculation
from sklearn.model_selection import train_test_split # split data

# initialize parameters
classes = ['car', 'motorbike'] # define 2 classses 
num_classes = len(classes)     # the number of classes
image_size = 224               # pixels of width or height

# read images and cnvert numpy array
X = [] # ready to make list
Y = [] # ready to make list

# numbering and 
for index, class_label in enumerate(classes):
    
    # create class_label directories (car or motorbike)
    photos_dir ='./' + class_label

    # search jpeg fileb and create files object
    files = glob.glob(photos_dir + '/*.jpg')

    for i, file in enumerate(files):

        # open image files as instance
        image = Image.open(file)

        # convert image into RGB value data
        image = image.convert('RGB')

        # align to the same size (just in case)
        image = image.resize((image_size, image_size))

        # convert RGB value into numpy array 
        data = np.asarray(image)

        # do not normalize
        # so, data object nees to be normalize other script later 
        #data = data / 255.0

        # append value to list
        X.append(data)
        Y.append(index)

# convert list into numpy array
X = np.array(X)
Y = np.array(Y)

# split data (training and test)
X_train, X_test, y_train, y_test = train_test_split(X, Y)

# replace 1 variable
xy =(X_train, X_test, y_train, y_test)

# save values as npy file 
np.save('./image_files_224.npy', xy)

from PIL import Image # image operation

import os # get file list

import glob # image treatment

import numpy as np # culculation

from sklearn.model_selection import train_test_split # split data

# initialize parameters

classes = ['car', 'motorbike'] # define 2 classses

num_classes = len(classes) # the number of classes

image_size = 224 # pixels of width or height

# read images and cnvert numpy array

X = [] # ready to make list

Y = [] # ready to make list

# numbering and

for index, class_label in enumerate(classes):

# create class_label directories (car or motorbike)

photos_dir ='./' + class_label

# search jpeg fileb and create files object

files = glob.glob(photos_dir + '/*.jpg')

for i, file in enumerate(files):

# open image files as instance

image = Image.open(file)

# convert image into RGB value data

image = image.convert('RGB')

# align to the same size (just in case)

image = image.resize((image_size, image_size))

# convert RGB value into numpy array

data = np.asarray(image)

# do not normalize

# so, data object nees to be normalize other script later

#data = data / 255.0

# append value to list

X.append(data)

Y.append(index)

# convert list into numpy array

X = np.array(X)

Y = np.array(Y)

# split data (training and test)

X_train, X_test, y_train, y_test = train_test_split(X, Y)

# replace 1 variable

xy =(X_train, X_test, y_train, y_test)

# save values as npy file

np.save('./image_files_224.npy', xy)

転移学習のモデル作成

では、データができたところでVGG16モデルを作成しましょう。

モジュール追加：from keras.applications import VGG16

ロードデータ名変更；

X_train, X_test, y_train, y_test = np.load(‘./image_files_224.npy’, allow_pickle=True)

正規化操作の追加：

X_train = X_train.astype(‘float’) / 255.0
X_test = X_train.astype(‘float’) / 255.0

モデル定義の変更

model = VGG16( weights=’imagenet’, include_top=False, input_shape=(image_size,image_size, 3) )
print(‘Model loaded’)
model.summary()

import numpy as np
import tensorflow as tf
import keras
import matplotlib.pyplot as plt

from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D

from keras.optimizers import SGD, Adam
from keras.utils import np_utils

from keras.applications import VGG16

# initialize parameters
classes = ['car', 'motorbike'] # define 2 classses 
num_classes = len(classes)     # the number of classes
image_size = 224               # pixels of width or height

# load npy file
# need 'allow_pickle=True'
X_train, X_test, y_train, y_test = np.load('./image_files_224.npy', allow_pickle=True)

# one hot encording
y_train = np_utils.to_categorical(y_train, num_classes)
y_test = np_utils.to_categorical(y_test, num_classes)

# normalize
X_train = X_train.astype('float') / 255.0
X_test = X_train.astype('float') / 255.0

# define model
model = VGG16(
    weights='imagenet', include_top=False,
    input_shape=(image_size,image_size, 3)
    )
print('Model loaded')
model.summary()

import numpy as np

import tensorflow as tf

import keras

import matplotlib.pyplot as plt

from keras.models import Sequential

from keras.layers import Dense, Dropout, Flatten

from keras.layers import Conv2D, MaxPooling2D

from keras.optimizers import SGD, Adam

from keras.utils import np_utils

from keras.applications import VGG16

# initialize parameters

classes = ['car', 'motorbike'] # define 2 classses

num_classes = len(classes) # the number of classes

image_size = 224 # pixels of width or height

# load npy file

# need 'allow_pickle=True'

X_train, X_test, y_train, y_test = np.load('./image_files_224.npy', allow_pickle=True)

# one hot encording

y_train = np_utils.to_categorical(y_train, num_classes)

y_test = np_utils.to_categorical(y_test, num_classes)

# normalize

X_train = X_train.astype('float') / 255.0

X_test = X_train.astype('float') / 255.0

# define model

model = VGG16(

weights='imagenet', include_top=False,

input_shape=(image_size,image_size, 3)

)

print('Model loaded')

model.summary()

結果

input_1 (InputLayer) (None, 224, 224, 3)について

１枚の224 x 224 サイズの画像データにつき、 RGBの３枚に分解されたデータが入力データとし入ります。

まだ何枚のデータが入ってくるかわかりませんのでNoneになっています。

畳み込み２回　プーリング１回　畳み込み２回　プーリング１回　畳み込み３回・・・、で　１６層あるのだそうですがどういう数え方をしたら１６になるかは理解できていません・・・。

Model loaded
Model: "vgg16"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
input_1 (InputLayer)         (None, 224, 224, 3)       0
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 224, 224, 64)      1792
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 224, 224, 64)      36928
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 112, 112, 64)      0
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 112, 112, 128)     73856
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 112, 112, 128)     147584
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 56, 56, 128)       0
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 56, 56, 256)       295168
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 56, 56, 256)       590080
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 56, 56, 256)       590080
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 28, 28, 256)       0
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 28, 28, 512)       1180160
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 28, 28, 512)       2359808
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 28, 28, 512)       2359808
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 14, 14, 512)       0
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 14, 14, 512)       2359808
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 14, 14, 512)       2359808
_________________________________________________________________
block5_conv3 (Conv2D)        (None, 14, 14, 512)       2359808
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 7, 7, 512)         0
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0
_________________________________________________________________

Model loaded

Model: "vgg16"

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

input_1 (InputLayer) (None, 224, 224, 3) 0

_________________________________________________________________

block1_conv1 (Conv2D) (None, 224, 224, 64) 1792

_________________________________________________________________

block1_conv2 (Conv2D) (None, 224, 224, 64) 36928

_________________________________________________________________

block1_pool (MaxPooling2D) (None, 112, 112, 64) 0

_________________________________________________________________

block2_conv1 (Conv2D) (None, 112, 112, 128) 73856

_________________________________________________________________

block2_conv2 (Conv2D) (None, 112, 112, 128) 147584

_________________________________________________________________

block2_pool (MaxPooling2D) (None, 56, 56, 128) 0

_________________________________________________________________

block3_conv1 (Conv2D) (None, 56, 56, 256) 295168

_________________________________________________________________

block3_conv2 (Conv2D) (None, 56, 56, 256) 590080

_________________________________________________________________

block3_conv3 (Conv2D) (None, 56, 56, 256) 590080

_________________________________________________________________

block3_pool (MaxPooling2D) (None, 28, 28, 256) 0

_________________________________________________________________

block4_conv1 (Conv2D) (None, 28, 28, 512) 1180160

_________________________________________________________________

block4_conv2 (Conv2D) (None, 28, 28, 512) 2359808

_________________________________________________________________

block4_conv3 (Conv2D) (None, 28, 28, 512) 2359808

_________________________________________________________________

block4_pool (MaxPooling2D) (None, 14, 14, 512) 0

_________________________________________________________________

block5_conv1 (Conv2D) (None, 14, 14, 512) 2359808

_________________________________________________________________

block5_conv2 (Conv2D) (None, 14, 14, 512) 2359808

_________________________________________________________________

block5_conv3 (Conv2D) (None, 14, 14, 512) 2359808

_________________________________________________________________

block5_pool (MaxPooling2D) (None, 7, 7, 512) 0

=================================================================

Total params: 14,714,688

Trainable params: 14,714,688

Non-trainable params: 0

_________________________________________________________________

ひとまず、できていました。このあと、全結合層を後ろに追加していくことになります。それは次回！

次回

Web Application: 第７回　VGG16を使った転移学習

Web Application: 第５回　npyデータからCNN

こんにちは Keita_Nakamori(´・ω・`)です。

前回はTensor Flowにインプットするnpyデータを作成しました。

今回はTensorflowとKerasで畳み込みニューラルネットワークを構築して、訓練・評価してみます。

keras公式ページのSequential-model-guideを参考にします。

スクリプト：cnn_test.py

VGG-likeなconvnet　を参考に動作を確認していきます。

import numpy as np
import tensorflow as tf
import keras

from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D

from keras.optimizers import SGD, Adam
from keras.utils import np_utils


# initialize parameters
classes = ['car', 'motorbike'] # define 2 classses 
num_classes = len(classes)     # the number of classes
image_size = 150               # pixels of width or height

# load npy file
# need 'allow_pickle=True'
X_train, X_test, y_train, y_test = np.load('./image_files.npy', allow_pickle=True)

# one hot encording
y_train = np_utils.to_categorical(y_train, num_classes)
y_test = np_utils.to_categorical(y_test, num_classes)

# define model
model = Sequential()

model.add(Conv2D(32,(3, 3), activation='relu', input_shape=(image_size, image_size, 3)))
model.add(Conv2D(32,(3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Dropout(0.25)) # 25%のデータは捨てる

model.add(Flatten())
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))

# define optimiser
opt = SGD(lr=0.01) # rmsprop, adam etc...
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])

# trainning start
history = model.fit(X_train, y_train, batch_size=32, epochs=30)

# evaluation
score = model.evaluate(X_test, y_test, batch_size=32)
print()
print('Test loss:', score[0])
print('Test accuracy:', score[1])

import numpy as np

import tensorflow as tf

import keras

from keras.models import Sequential

from keras.layers import Dense, Dropout, Flatten

from keras.layers import Conv2D, MaxPooling2D

from keras.optimizers import SGD, Adam

from keras.utils import np_utils

# initialize parameters

classes = ['car', 'motorbike'] # define 2 classses

num_classes = len(classes) # the number of classes

image_size = 150 # pixels of width or height

# load npy file

# need 'allow_pickle=True'

X_train, X_test, y_train, y_test = np.load('./image_files.npy', allow_pickle=True)

# one hot encording

y_train = np_utils.to_categorical(y_train, num_classes)

y_test = np_utils.to_categorical(y_test, num_classes)

# define model

model = Sequential()

model.add(Conv2D(32,(3, 3), activation='relu', input_shape=(image_size, image_size, 3)))

model.add(Conv2D(32,(3, 3), activation='relu'))

model.add(MaxPooling2D(pool_size=(2,2)))

model.add(Dropout(0.25)) # 25%のデータは捨てる

model.add(Flatten())

model.add(Dense(256, activation='relu'))

model.add(Dropout(0.5))

model.add(Dense(num_classes, activation='softmax'))

# define optimiser

opt = SGD(lr=0.01) # rmsprop, adam etc...

model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])

# trainning start

history = model.fit(X_train, y_train, batch_size=32, epochs=30)

# evaluation

score = model.evaluate(X_test, y_test, batch_size=32)

print()

print('Test loss:', score[0])

print('Test accuracy:', score[1])

実行

では　Anacondaプロンプトから　$ conda info -e で仮想環境リストを確認して

(base) C:\Users\keita>conda info -e
# conda environments:
#
base * C:\Users\keita\Anaconda3
djangoai C:\Users\keita\Anaconda3\envs\djangoai

(base) C:\Users\keita>conda info -e

# conda environments:

base * C:\Users\keita\Anaconda3

djangoai C:\Users\keita\Anaconda3\envs\djangoai

djangoaiが存在することを確認したら、$ conda activate djangoai　して仮想環境に入ります。

(base) C:\Users\keita>conda activate djangoai

(djangoai) C:\Users\keita>

(base) C:\Users\keita>conda activate djangoai

(djangoai) C:\Users\keita>

カレントディレクトリをcnn_test.pyがある場所まで移動して、実行します。

(djangoai) C:\Users\keita>cd anaconda_projects
(djangoai) C:\Users\keita\anaconda_projects>cd djangoai
(djangoai) C:\Users\keita\anaconda_projects\djangoai>python cnn_test.py

(djangoai) C:\Users\keita>cd anaconda_projects

(djangoai) C:\Users\keita\anaconda_projects>cd djangoai

(djangoai) C:\Users\keita\anaconda_projects\djangoai>python cnn_test.py

結果

エポック数３０回やった結果、精度はacc: 0.99まで向上しましたが、検証データではTest accuracy:0.81までしか出ていません。おそらく過学習になっていると思われます。まだまだ改善の余地ありです。

Epoch 5/30
600/600 [==============================] - 13s 22ms/step - loss: 0.5880 - acc: 0.7017
Epoch 10/30
600/600 [==============================] - 13s 22ms/step - loss: 0.4863 - acc: 0.7850
Epoch 15/30
600/600 [==============================] - 14s 23ms/step - loss: 0.3483 - acc: 0.8617
Epoch 20/30
600/600 [==============================] - 13s 22ms/step - loss: 0.2374 - acc: 0.9150
Epoch 25/30
600/600 [==============================] - 13s 22ms/step - loss: 0.1508 - acc: 0.9633
600/600 [==============================] - 14s 23ms/step - loss: 0.0626 - acc: 0.9900
200/200 [==============================] - 2s 8ms/step

Test loss: 0.37387794256210327
Test accuracy: 0.81

Epoch 5/30

600/600 [==============================] - 13s 22ms/step - loss: 0.5880 - acc: 0.7017

Epoch 10/30

600/600 [==============================] - 13s 22ms/step - loss: 0.4863 - acc: 0.7850

Epoch 15/30

600/600 [==============================] - 14s 23ms/step - loss: 0.3483 - acc: 0.8617

Epoch 20/30

600/600 [==============================] - 13s 22ms/step - loss: 0.2374 - acc: 0.9150

Epoch 25/30

600/600 [==============================] - 13s 22ms/step - loss: 0.1508 - acc: 0.9633

600/600 [==============================] - 14s 23ms/step - loss: 0.0626 - acc: 0.9900

200/200 [==============================] - 2s 8ms/step

Test loss: 0.37387794256210327

Test accuracy: 0.81

オプティマイザーを変更

SGDからAdamに切り替えてみました。大きく悪化しました。（笑）

Epoch 10/30
600/600 [==============================] - 19s 32ms/step - loss: 8.0859 - acc: 0.4983

Epoch 20/30
600/600 [==============================] - 17s 28ms/step - loss: 8.0859 - acc: 0.4983

Epoch 30/30
600/600 [==============================] - 17s 29ms/step - loss: 8.0859 - acc: 0.4983
200/200 [==============================] - 1s 7ms/step

Test loss: 7.978457450866699
Test accuracy: 0.505

Epoch 10/30

600/600 [==============================] - 19s 32ms/step - loss: 8.0859 - acc: 0.4983

Epoch 20/30

600/600 [==============================] - 17s 28ms/step - loss: 8.0859 - acc: 0.4983

Epoch 30/30

600/600 [==============================] - 17s 29ms/step - loss: 8.0859 - acc: 0.4983

200/200 [==============================] - 1s 7ms/step

Test loss: 7.978457450866699

Test accuracy: 0.505

可視化にチャレンジ

djangoai仮想環境内で$pip install matplotlib　します。

matplotlib-3.1.1　が入りました。

下記コードを最後に追加しました。オプティマイザーはSGDに戻しました。

import&nbsp;matplotlib.pyplot&nbsp;as&nbsp;plt

# visualizer
loss     = history.history['loss']

nb_epoch = len(loss)
plt.plot(range(nb_epoch), loss, marker='.', label='loss')

plt.legend(loc='best', fontsize=10)
plt.grid()
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()

import matplotlib.pyplot as plt

# visualizer

loss = history.history['loss']

nb_epoch = len(loss)

plt.plot(range(nb_epoch), loss, marker='.', label='loss')

plt.legend(loc='best', fontsize=10)

plt.grid()

plt.xlabel('epoch')

plt.ylabel('loss')

plt.show()

横軸にエポック数　縦軸に損失です。　きれいな右肩下がりです。

次回

Web Application: 第６回　はじめてのwebアプリ

Web Application: 第４回　画像データをnpyデータに変換

こんにちはKeita_Nakamori(´・ω・`)です。

前回flickrから取得した画像データをTensor Flowが読めるように数値データに変換していきます。

必要なモジュールをインストール

pip install Pillow :
pip install scikit-learn :

Pillow-6.1.0とscikit-learn-0.21.3が入りました

スクリプト：generate_inputdata.py

from PIL import Image # image operation
import os             # get file list
import glob           # image treatment
import numpy as np    # culculation
from sklearn.model_selection import train_test_split # split data

# initialize parameters
classes = ['car', 'motorbike'] # define 2 classses 
num_classes = len(classes)     # the number of classes
image_size = 150               # pixels of width or height

# read images and cnvert numpy array
X = [] # ready to make list
Y = [] # ready to make list

# numbering and 
for index, class_label in enumerate(classes):
    
    # create class_label directories (car or motorbike)
    photos_dir ='./' + class_label

    # search jpeg fileb and create files object
    files = glob.glob(photos_dir + '/*.jpg')

    for i, file in enumerate(files):

        # open image files as instance
        image = Image.open(file)

        # convert image into RGB value data
        image = image.convert('RGB')

        # align to the same size (just in case)
        image = image.resize((image_size, image_size))

        # convert RGB value into numpy array 
        data = np.asarray(image)

        # normalize
        data = data / 255.0

        # append value to list
        X.append(data)
        Y.append(index)

# convert list into numpy array
X = np.array(X)
Y = np.array(Y)

# split data (training and test)
X_train, X_test, y_train, y_test = train_test_split(X, Y)

# replace 1 variable
xy =(X_train, X_test, y_train, y_test)

# save values as npy file 
np.save('./image_files.npy', xy)

from PIL import Image # image operation

import os # get file list

import glob # image treatment

import numpy as np # culculation

from sklearn.model_selection import train_test_split # split data

# initialize parameters

classes = ['car', 'motorbike'] # define 2 classses

num_classes = len(classes) # the number of classes

image_size = 150 # pixels of width or height

# read images and cnvert numpy array

X = [] # ready to make list

Y = [] # ready to make list

# numbering and

for index, class_label in enumerate(classes):

# create class_label directories (car or motorbike)

photos_dir ='./' + class_label

# search jpeg fileb and create files object

files = glob.glob(photos_dir + '/*.jpg')

for i, file in enumerate(files):

# open image files as instance

image = Image.open(file)

# convert image into RGB value data

image = image.convert('RGB')

# align to the same size (just in case)

image = image.resize((image_size, image_size))

# convert RGB value into numpy array

data = np.asarray(image)

# normalize

data = data / 255.0

# append value to list

X.append(data)

Y.append(index)

# convert list into numpy array

X = np.array(X)

Y = np.array(Y)

# split data (training and test)

X_train, X_test, y_train, y_test = train_test_split(X, Y)

# replace 1 variable

xy =(X_train, X_test, y_train, y_test)

# save values as npy file

np.save('./image_files.npy', xy)

実行しましょう

(djangoai) C:\Users\keita\anaconda_projects\djangoai>python generate_inputdata.py

フォルダを見ると新しく「image_files.npy」　　609,516,317 （６０９MB？）のデータが生成されました。

次回

Web Application: 第５回　はじめてのwebアプリ

2019年9月
月	火	水	木	金	土	日
						1
2	3	4	5	6	7	8
9	10	11	12	13	14	15
16	17	18	19	20	21	22
23	24	25	26	27	28	29
30