Keras模型训练准确率近100%但predict对训练样本分类错误

我用Keras做游戏角色的多分类任务，每个类别约300张图片。目前的问题是：训练准确率接近100%、验证准确率约70%，但用model.predict()对刚训练过的图片进行预测时，分类结果完全错误。

我已经尝试了多种解决方法：更换数据加载方式、调整模型架构、修改model.predict()前的图像预处理函数，但都没有效果。

模型训练代码 #

class_names = ['ana', 'ashe','baby', 'ball','bap', 'bastion','brig', 'cass','doom', 'dva'
,'echo', 'genji','hanzo', 'hog','jq', 'junkrat','kiriko', 'lucio','mei', 'mercy','moira', 'orisa'
,'pharah', 'ram','reaper', 'rein','sigma', 'sojourn','soldier', 'sombra','sym', 'torb','tracer', 'widow'
,'winston', 'zarya','zen']

class_names_label = {class_name:i for i, class_name in enumerate(class_names)}
nb_classes = len(class_names)
print(class_names_label)
IMAGE_SIZE = (128,128)

def load_data():
    DIRECTORY = r'D:\crop-id\enemy-crop'
    CATEGORY = ['train', 'test']

    output = []

    for category in CATEGORY:
        path = os.path.join(DIRECTORY, category)
        print(path)
        images = []
        labels = []

        print('Loading {}'.format(category))

        for folder in os.listdir(path):
            label = class_names_label[folder]

            #iterate through each img in folder
            for file in os.listdir(os.path.join(path,folder)):

                #get path name of img
                img_path = os.path.join(os.path.join(path, folder), file)

                #open and resize img
                image = cv.imread(img_path)
                image = cv.cvtColor(image, cv.COLOR_BGR2RGB)
                image = cv.resize(image, IMAGE_SIZE)

                #append the image and its corresponding label to output
                images.append(image)
                labels.append(label)

        images = np.array(images, dtype = 'float32')
        labels = np.array(labels, dtype='int32')

        output.append((images, labels))

    return output

(train_images, train_labels), (test_images, test_labels) = load_data()

train_images, train_labels = shuffle(train_images, train_labels, random_state=25)

model = Sequential()
model.add(Conv2D(32, (3,3), activation='relu', input_shape=(128, 128, 3)))
model.add(MaxPooling2D())
model.add(BatchNormalization())

model.add(Conv2D(64, (3,3), activation='relu'))
model.add(MaxPooling2D())
model.add(BatchNormalization())

model.add(Conv2D(64, (3,3), activation='relu'))
model.add(MaxPooling2D()) 
model.add(BatchNormalization())

model.add(Flatten())
model.add(Dense(256, activation='relu'))
model.add(Dense(37, activation='softmax'))
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
model.summary()
model.fit(train_images, train_labels, batch_size=32, epochs=10, validation_split=.2)

model.save(os.path.join('models','plzwork14.h5'))

predictions = model.predict(test_images)
pred_labels = np.argmax(predictions, axis = 1)
print(classification_report(test_labels, pred_labels))

预测预处理代码 #

def prepare(img):
    img = cv.resize(img, (128,128))
    img = np.reshape(img, (128,128,3))
    img = np.expand_dims(img/255, 0)
    prediction = model.predict(img)
    prediction = prediction[0]
    print(prediction)
    print(class_names[np.argmax(prediction)])

img1 = cv.imread(r'C:\Users\andrew\Desktop\sombra.jpg')

prepare(img1)

内容的提问来源于stack exchange，提问作者Andrew