質問編集履歴
1
プログラムの修正
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -40,11 +40,11 @@
|
|
|
40
40
|
from icrawler.builtin import BingImageCrawler
|
|
41
41
|
from sklearn.model_selection import train_test_split
|
|
42
42
|
from keras.models import Sequential
|
|
43
|
-
from keras.layers import Conv2D, UpSampling2D, InputLayer, MaxPool2D, Dropout
|
|
43
|
+
from keras.layers import Conv2D, UpSampling2D, InputLayer, MaxPool2D, Dropout, BatchNormalization, Activation
|
|
44
44
|
from tensorflow.python.keras.optimizers import adam_v2
|
|
45
45
|
|
|
46
46
|
# カラー画像と白黒画像のペアを用意する関数
|
|
47
|
-
def prepare_data(img_paths, img_size=(
|
|
47
|
+
def prepare_data(img_paths, img_size=(96, 96)):
|
|
48
48
|
color_imgs = []
|
|
49
49
|
gray_imgs = []
|
|
50
50
|
|
|
@@ -55,12 +55,13 @@
|
|
|
55
55
|
color_img = cv2.resize(color_img, img_size)
|
|
56
56
|
|
|
57
57
|
# 白黒画像
|
|
58
|
-
|
|
58
|
+
|
|
59
59
|
gray_img = cv2.imread(img_path, cv2.IMREAD_GRAYSCALE)
|
|
60
60
|
gray_img = cv2.resize(gray_img, img_size)
|
|
61
61
|
gray_img = np.expand_dims(gray_img, axis=-1) # チャンネル次元を追加
|
|
62
|
+
|
|
63
|
+
|
|
62
64
|
'''
|
|
63
|
-
|
|
64
65
|
img = cv2.imread(img_path)
|
|
65
66
|
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
|
|
66
67
|
|
|
@@ -74,6 +75,7 @@
|
|
|
74
75
|
|
|
75
76
|
gray_img = cv2.resize(gray_img, img_size)
|
|
76
77
|
gray_img = np.expand_dims(gray_img, axis=-1) # チャンネル次元を追加
|
|
78
|
+
'''
|
|
77
79
|
|
|
78
80
|
color_imgs.append(color_img)
|
|
79
81
|
gray_imgs.append(gray_img)
|
|
@@ -85,39 +87,64 @@
|
|
|
85
87
|
return color_imgs, gray_imgs
|
|
86
88
|
|
|
87
89
|
# データセットのパス
|
|
88
|
-
dataset_paths = glob.glob('
|
|
90
|
+
dataset_paths = glob.glob('Python/ml/image3/slime2/*.png')
|
|
89
91
|
|
|
90
92
|
# カラー画像と白黒画像のペアを用意
|
|
91
93
|
color_images, gray_images = prepare_data(dataset_paths)
|
|
92
94
|
|
|
93
95
|
# データを訓練データとテストデータに分割
|
|
94
|
-
X_train, X_test, y_train, y_test = train_test_split(gray_images, color_images, test_size=0.
|
|
96
|
+
X_train, X_test, y_train, y_test = train_test_split(gray_images, color_images, test_size=0.2, random_state=42)
|
|
95
97
|
|
|
96
98
|
# モデルの定義
|
|
97
99
|
model = Sequential()
|
|
100
|
+
|
|
101
|
+
#モデルの初期化
|
|
98
|
-
model.add(InputLayer(
|
|
102
|
+
model.add(InputLayer(shape=(96, 96, 1)))
|
|
103
|
+
|
|
99
|
-
|
|
104
|
+
model.add(Conv2D(3, (3, 3), padding='same', strides=1))
|
|
105
|
+
model.add(BatchNormalization(momentum=0.9, epsilon=1e-5))
|
|
106
|
+
model.add(Activation('relu'))
|
|
107
|
+
model.add(Dropout(0.5))
|
|
108
|
+
|
|
100
|
-
model.add(Conv2D(64, (3, 3),
|
|
109
|
+
model.add(Conv2D(64, (3, 3), padding='same', strides=1))
|
|
110
|
+
model.add(BatchNormalization(momentum=0.9, epsilon=1e-5))
|
|
101
|
-
model.add(
|
|
111
|
+
model.add(Activation('relu'))
|
|
102
|
-
|
|
112
|
+
model.add(Dropout(0.5))
|
|
113
|
+
|
|
103
|
-
model.add(Conv2D(128, (3, 3),
|
|
114
|
+
model.add(Conv2D(128, (3, 3), padding='same', strides=1))
|
|
115
|
+
model.add(BatchNormalization(momentum=0.9, epsilon=1e-5))
|
|
116
|
+
model.add(Activation('relu'))
|
|
117
|
+
model.add(Dropout(0.5))
|
|
118
|
+
|
|
119
|
+
model.add(Conv2D(256, (3, 3), padding='same', strides=1))
|
|
120
|
+
model.add(BatchNormalization(momentum=0.9, epsilon=1e-5))
|
|
121
|
+
model.add(Activation('relu'))
|
|
122
|
+
model.add(Dropout(0.5))
|
|
123
|
+
|
|
104
|
-
model.add(Conv2D(128, (3, 3),
|
|
124
|
+
model.add(Conv2D(128, (3, 3), padding='same', strides=1))
|
|
105
|
-
|
|
125
|
+
model.add(BatchNormalization(momentum=0.9, epsilon=1e-5))
|
|
126
|
+
model.add(Activation('relu'))
|
|
127
|
+
model.add(Dropout(0.5))
|
|
128
|
+
|
|
106
|
-
model.add(Conv2D(
|
|
129
|
+
model.add(Conv2D(64, (3, 3), padding='same', strides=1))
|
|
130
|
+
model.add(BatchNormalization(momentum=0.9, epsilon=1e-5))
|
|
107
|
-
model.add(
|
|
131
|
+
model.add(Activation('relu'))
|
|
108
|
-
|
|
132
|
+
model.add(Dropout(0.5))
|
|
133
|
+
|
|
109
|
-
model.add(Conv2D(3, (3, 3),
|
|
134
|
+
model.add(Conv2D(3, (3, 3), padding='same', strides=1)) # Sigmoid から Linear に変更
|
|
110
|
-
|
|
135
|
+
model.add(BatchNormalization(momentum=0.9, epsilon=1e-5))
|
|
136
|
+
model.add(Activation('sigmoid'))
|
|
111
137
|
|
|
112
138
|
# モデルのコンパイル
|
|
139
|
+
#初期値：0.0002
|
|
113
|
-
optimizer = tf.
|
|
140
|
+
optimizer = tf.keras.optimizers.Adam(learning_rate=0.002)
|
|
114
141
|
model.compile(optimizer=optimizer, loss='mae', metrics=['accuracy'])
|
|
115
142
|
|
|
116
143
|
|
|
117
144
|
print(model.summary())
|
|
118
145
|
|
|
119
146
|
# モデルの学習
|
|
120
|
-
history = model.fit(X_train, y_train, epochs=20, batch_size=
|
|
147
|
+
history = model.fit(X_train, y_train, epochs=20, batch_size=150, validation_data=(X_test, y_test))
|
|
121
148
|
|
|
122
149
|
# 学習過程の損失と精度の情報を取得
|
|
123
150
|
train_loss = history.history['loss']
|
|
@@ -150,13 +177,13 @@
|
|
|
150
177
|
plt.show()
|
|
151
178
|
|
|
152
179
|
# ある白黒データに対する予測
|
|
153
|
-
input_gray_image = X_test[0].reshape(1,
|
|
180
|
+
input_gray_image = X_test[0].reshape(1, 96, 96, 1)
|
|
154
181
|
predicted_color_image = model.predict(input_gray_image)
|
|
155
182
|
|
|
156
183
|
# 結果の表示
|
|
157
184
|
plt.subplot(1, 3, 1)
|
|
158
185
|
plt.title('Input Grayscale Image')
|
|
159
|
-
plt.imshow(X_test[0].reshape(
|
|
186
|
+
plt.imshow(X_test[0].reshape(96, 96), cmap='gray') # グレースケール画像なので256x256に修正
|
|
160
187
|
|
|
161
188
|
plt.subplot(1, 3, 2)
|
|
162
189
|
plt.title('Input RGB Image')
|