拾ってきたモデルを自分のデータセットに使えるようにしたい
なぜか使えません.
データに合うように、入力層と出力層の大きさを変えました。
元はCIFAR-10用のモデルです。
使用コード
python
1import tensorflow as tf 2from tensorflow.keras.layers import Activation, Dense, Dropout, Conv2D, Flatten, MaxPool2D 3from tensorflow.keras.models import Sequential, load_model 4from tensorflow.keras.optimizers import Adam 5 6 7 8#モデルの作成 9model = Sequential() 10 11# Conv→Conv→Pool→Dropout 12model.add(Conv2D(128, (3, 3), activation='relu', padding='same', input_shape=(128, 128, 3))) 13model.add(Conv2D(128, (3, 3), activation='relu', padding='same')) 14model.add(MaxPool2D(pool_size=(2, 2))) 15model.add(Dropout(0.25)) 16 17# Conv→Conv→Pool→Dropout 18model.add(Conv2D(64, (3, 3), activation='relu', padding='same')) 19model.add(Conv2D(64, (3, 3), activation='relu', padding='same')) 20model.add(MaxPool2D(pool_size=(2, 2))) 21model.add(Dropout(0.25)) 22 23# Flatten→Dense→Dropout→Dense 24model.add(Flatten()) 25model.add(Dense(512, activation='relu')) 26model.add(Dropout(0.5)) 27model.add(Dense(7, activation='softmax'))
変更前
python
1import tensorflow as tf 2from tensorflow.keras.layers import Activation, Dense, Dropout, Conv2D, Flatten, MaxPool2D 3from tensorflow.keras.models import Sequential, load_model 4from tensorflow.keras.optimizers import Adam 5 6 7 8#モデルの作成 9model = Sequential() 10 11# Conv→Conv→Pool→Dropout 12model.add(Conv2D(32, (3, 3), activation='relu', padding='same', input_shape=(32, 32, 3))) 13model.add(Conv2D(32, (3, 3), activation='relu', padding='same')) 14model.add(MaxPool2D(pool_size=(2, 2))) 15model.add(Dropout(0.25)) 16 17# Conv→Conv→Pool→Dropout 18model.add(Conv2D(64, (3, 3), activation='relu', padding='same')) 19model.add(Conv2D(64, (3, 3), activation='relu', padding='same')) 20model.add(MaxPool2D(pool_size=(2, 2))) 21model.add(Dropout(0.25)) 22 23# Flatten→Dense→Dropout→Dense 24model.add(Flatten()) 25model.add(Dense(512, activation='relu')) 26model.add(Dropout(0.5)) 27model.add(Dense(10, activation='softmax'))
データ取得部分
python
1orig_image, orig_label, class_list = load_imgs(root_dir='kill_me_images/kill_me_baby_datasets/kill_me_baby_datasets/') 2#['agiri', 'botsu', 'others', 'sonya', 'yasuna', 'yasuna&agiri', 'yasuna&sonya']
関数部分(cv2.COLOR_BGR2RGBでBRGをRGBに調整)
python
1def load_imgs(root_dir): 2 3 print(class_list) 4 num_class = len(class_list) 5 img_paths = [] 6 labels = [] 7 images = [] 8 for cl_name in class_list: 9 img_names = os.listdir(os.path.join(root_dir, cl_name)) 10 for img_name in img_names: 11 img_paths.append(os.path.abspath(os.path.join(root_dir, cl_name, img_name))) 12 hot_cl_name = get_class_one_hot(cl_name, class_list) 13 labels.append(hot_cl_name) 14 15 for img_path in img_paths: 16 img = cv2.imread(img_path) 17 img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) 18 images.append(img) 19 20 images = np.array(images) 21 22 return np.array(images), np.array(labels), class_list
onehot
python
1def get_class_one_hot(class_str, class_list): 2 label = class_list.index(class_str) 3 label_hot = tf.one_hot(label, len(class_list)) 4 5 return label_hot
実行部分
python
1batch_size = 32 2 3model.compile(loss='categorical_crossentropy', optimizer=tf.keras.optimizers.Adam(), metrics=['accuracy']) 4tb_cb = tf.keras.callbacks.TensorBoard(log_dir="log_dir") 5ckps = [tb_cb] 6 7# 学習用データを用意する 8train_datagen = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1./255, horizontal_flip=True) 9train_generator = train_datagen.flow_from_directory('kill_me_images/kill_me_baby_datasets/', classes=class_list, target_size=(128, 128), batch_size=batch_size, class_mode='categorical') 10 11# 学習開始! 12model.fit_generator(train_generator, steps_per_epoch=train_generator.samples//batch_size, epochs=100, callbacks=ckps) 13model.save("models/killme_vgg16.h5") 14 15sess = tf.keras.backend.get_session() 16saver = tf.train.Saver() 17saver.save(sess, "models/killme_vgg16.ckpt")
あなたの回答
tips
プレビュー