make_model.py

from PIL import Image
import os, glob, numpy as np
from numpy import ndarray as nda
from sklearn.model_selection import train_test_split

class make_model():
    def __init__(self):
        #self.image_rotate()
        self.make_npy_file()
        self.make_model()

    def image_rotate(self):
        caltech_dir = "./multi_img_data/imgs_others/train"
        saving_dir = "./multi_img_data/imgs_others/train_rotated"
        categories = ["apple", "carrot", "melon", "strawberry", "tomato", "watermelon"]
        
        image_w = 128
        image_h = 128

        for idx, cat in enumerate(categories):
            image_dir = caltech_dir + "/" + cat
            files = glob.glob(image_dir + "/*.png")
            print(cat, " 파일 길이 : ", len(files))
            for i, f in enumerate(files):
                img = Image.open(f)
                img = img.convert("RGB")
                img = img.resize((image_w, image_h))
                white=(255, 255, 255)
                for j in range(-9, 9):
                    img_name=f.split('.')
                    img_name = img_name[1].split("\\")
                    img_ro=img.rotate(20*j, expand=1, fillcolor=white)
                    img_ro=img_ro.crop((img_ro.size[0]/2-image_w/2, img_ro.size[1]/2-image_h/2, img_ro.size[0]/2+image_w/2, img_ro.size[1]/2+image_h/2))
                    save_dir=saving_dir + "/" + cat + "/" + img_name[1]+"_"+str(j+10)+".png"
                    img_ro.save(save_dir)
                
                if i % 700 == 0:
                    print(cat, " : ", f)
        print("전처리 완료\n")

    def make_npy_file(self):
        caltech_dir = "./multi_img_data/imgs_others/train"

        categories = ["apple", "carrot", "melon", "strawberry", "tomato", "watermelon"]
        nb_classes = len(categories)

        image_w = 128
        image_h = 128

        pixels = image_h * image_w * 3

        X = []
        y = []

        for idx, cat in enumerate(categories):
            # one-hot 돌리기.
            label = [0 for i in range(nb_classes)]
            label[idx] = 1

            image_dir = caltech_dir + "/" + cat
            files = glob.glob(image_dir + "/*.png")
            print(cat, " 파일 길이 : ", len(files))
            for i, f in enumerate(files):
                img = Image.open(f)
                img = img.convert("RGB")
                img = img.resize((image_w, image_h))
                data=np.asarray(img)
               
                X.append(data)
                y.append(label)
                
                if i % 700 == 0:
                    print(cat, " : ", f)
        
        X = np.array(X)
        y = np.array(y)
        # 1 0 0 0 이면 airplanes
        # 0 1 0 0 이면 buddha 이런식

        X_train, X_test, y_train, y_test = train_test_split(X, y)
        xy = (X_train, X_test, y_train, y_test)
        print(len(y))
        np.save("./numpy_data/multi_image_data.npy", xy)

    def make_model(self):
        from keras.models import Sequential
        from keras.layers import Conv2D, MaxPooling2D, Dense, Flatten, Dropout
        from keras.callbacks import EarlyStopping, ModelCheckpoint
        import matplotlib.pyplot as plt
        import keras.backend.tensorflow_backend as K
        import tensorflow as tf

        config = tf.compat.v1.ConfigProto()
        config.gpu_options.allow_growth = True
        session = tf.compat.v1.Session(config=config)

        X_train, X_test, y_train, y_test = np.load("./numpy_data/multi_image_data.npy", allow_pickle=True)
        
        categories = ["apple", "carrot", "melon", "strawberry", "tomato", "watermelon"]
        nb_classes = len(categories)
        X_train = X_train.astype(float) / 255
        X_test = X_test.astype(float) / 255

        with K.tf_ops.device('/device:CPU:0'):
            model = Sequential()
            model.add(Conv2D(32, (3, 3), padding="same", input_shape=X_train.shape[1:],
                             activation='relu'))
            model.add(MaxPooling2D(pool_size=(2, 2)))
            model.add(Dropout(0.25))

            model.add(Conv2D(64, (3, 3), padding="same", activation='relu'))
            model.add(MaxPooling2D(pool_size=(2, 2)))
            model.add(Dropout(0.25))

            model.add(Flatten())
            model.add(Dense(256, activation='relu'))
            model.add(Dropout(0.5))
            model.add(Dense(nb_classes, activation='softmax'))
            model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
            model_dir = './model'

            if not os.path.exists(model_dir):
                os.mkdir(model_dir)

            model_path = model_dir + '/multi_img_classification.model'
            checkpoint = ModelCheckpoint(filepath=model_path, monitor='val_loss',
                                         verbose=1, save_best_only=True)
            early_stopping = EarlyStopping(monitor='val_loss', patience=6)

        model.summary()

        history = model.fit(X_train, y_train, batch_size=64, epochs=50, validation_data=(X_test, y_test), callbacks=[checkpoint, early_stopping])

        print("정확도 : %.4f" % (model.evaluate(X_test, y_test)[1]))


        y_vloss = history.history['val_loss']
        y_loss = history.history['loss']
        
        y_vacc = history.history['val_acc']
        y_acc = history.history['acc']
        # history에서 사용하는 key에 맞게 수정 (Keras 버전 == 2.2.5)

        x_len = np.arange(len(y_loss))

        plt.figure(1)
        plt.plot(x_len, y_vloss, marker='.', c='red', label='val_set_loss')
        plt.plot(x_len, y_loss, marker='.', c='blue', label='train_set_loss')
        plt.legend()
        plt.xlabel('epochs')
        plt.ylabel('loss')
        plt.grid()
        
        plt.figure(2)
        plt.plot(x_len, y_vacc, marker='.', c='red', label='val_set_acc')
        plt.plot(x_len, y_acc, marker='.', c='blue', label='train_set_acc')
        plt.legend()
        plt.xlabel('epochs')
        plt.ylabel('acc')
        plt.grid()
        
        plt.show()
        
if __name__ == '__main__':
    w = make_model()