python - Keras CNN 准确率高，但预测不佳。如何改进

Question

我必须通过 CNN 识别视网膜疾病。我有 1400 张图像，每类 700 张。我的类(class)是(0 - 无 PDR)和(1 - PDR)。我正在尝试建立一个模型来识别输入视网膜是否患有 4 级疾病。

我正在对我的图像进行以下操作并将其全部调整为 256x256:

ImageCV[index] = cv2.addWeighted(ImageCV[index],4, cv2.GaussianBlur(ImageCV[index],(0,0), 256/30), -4, 128)

它用我的图片制作了以下内容: https://imgur.com/X1p9G1c

然后，当我训练模型时，我得到了非常高的准确度(例如 99....)，但是当我尝试预测一些测试图像时，它失败了..例如，我在测试中放入了 10 个 PDR 示例文件夹并尝试预测它们(全部必须为 1)..这就是结果:

[[0.]]
[[0.]]
[[1.]]
[[0.]]
[[0.]]
[[0.]]
[[1.]]
[[0.]]
[[0.]]
[[0.]]

这是我的模型:

visible = Input(shape=(256,256,3))
conv1 = Conv2D(16, kernel_size=(3,3), activation='relu', strides=(1, 1))(visible)
conv2 = Conv2D(16, kernel_size=(3,3), activation='relu', strides=(1, 1))(conv1)
bat1 = BatchNormalization()(conv2)
conv3 = ZeroPadding2D(padding=(1, 1))(bat1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv3)

conv4 = Conv2D(32, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(pool1)
conv5 = Conv2D(32, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(conv4)
bat2 = BatchNormalization()(conv5)
pool2 = MaxPooling2D(pool_size=(1, 1))(bat2)

conv6 = Conv2D(64, kernel_size=(3,3), activation='relu',strides=(1, 1), padding='valid')(pool2)
conv7 = Conv2D(64, kernel_size=(3,3), activation='relu',strides=(1, 1), padding='valid')(conv6)
bat3 = BatchNormalization()(conv7)
conv7 = ZeroPadding2D(padding=(1, 1))(bat3)
pool3 = MaxPooling2D(pool_size=(1, 1))(conv7)

conv8 = Conv2D(128, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(pool3)
conv9 = Conv2D(128, kernel_size=(2,2), activation='relu', strides=(1, 1), padding='valid')(conv8)
bat4 = BatchNormalization()(conv9)
pool4 = MaxPooling2D(pool_size=(1, 1))(bat4)

flat = Flatten()(pool4)

output = Dense(1, activation='sigmoid')(flat)
model = Model(inputs=visible, outputs=output)

opt = optimizers.adam(lr=0.001, decay=0.0)

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


data, labels = ReadImages(TRAIN_DIR)

test, lt = ReadImages(TEST_DIR)

data = np.array(data)
labels = np.array(labels)
test = np.array(test)
lt = np.array(lt)

np.random.permutation(len(data))
np.random.permutation(len(labels))
np.random.permutation(len(test))
np.random.permutation(len(lt))

model.fit(data, labels, epochs=7, validation_data = (test,lt))

model.save('model.h5')

这是预测.py

model = load_model('model.h5')

for filename in os.listdir(r'v/'):
    if filename.endswith(".jpg") or filename.endswith(".ppm") or filename.endswith(".jpeg"):
        ImageCV = cv2.resize(cv2.imread(os.path.join(TEST_DIR) + filename), (256,256))
        ImageCV = cv2.addWeighted(ImageCV,4, cv2.GaussianBlur(ImageCV,(0,0), 256/30), -4, 128)
        cv2.imshow('image', ImageCV)
        cv2.waitKey(0)
        cv2.destroyAllWindows()
        ImageCV = ImageCV.reshape(-1,256,256,3)
        print(model.predict(ImageCV))

我可以做些什么来改善我的预测？我非常感谢您的帮助

更新好吧，我尝试做答案中所说的所有事情，但仍然不起作用......这是我现在的代码:

visible = Input(shape=(256,256,3))
conv1 = Conv2D(16, kernel_size=(3,3), activation='relu', strides=(1, 1))(visible)
conv2 = Conv2D(32, kernel_size=(3,3), activation='relu', strides=(1, 1))(conv1)
bat1 = BatchNormalization()(conv2)
conv3 = ZeroPadding2D(padding=(1, 1))(bat1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv3)
drop1 = Dropout(0.30)(pool1)

conv4 = Conv2D(32, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(drop1)
conv5 = Conv2D(64, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(conv4)
bat2 = BatchNormalization()(conv5)
pool2 = MaxPooling2D(pool_size=(1, 1))(bat2)
drop1 = Dropout(0.30)(pool2)

conv6 = Conv2D(128, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(pool2)
conv7 = Conv2D(128, kernel_size=(2,2), activation='relu', strides=(1, 1), padding='valid')(conv6)
bat3 = BatchNormalization()(conv7)
pool3 = MaxPooling2D(pool_size=(1, 1))(bat3)
drop1 = Dropout(0.30)(pool3)

flat = Flatten()(pool3)
drop4 = Dropout(0.50)(flat)

output = Dense(1, activation='sigmoid')(drop4)
model = Model(inputs=visible, outputs=output)

opt = optimizers.adam(lr=0.001, decay=0.0)

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

data, labels = ReadImages(TRAIN_DIR)
test, lt = ReadImages(TEST_DIR)

data = np.array(data)
labels = np.array(labels)

perm = np.random.permutation(len(data))
data = data[perm]
labels = labels[perm]
#model.fit(data, labels, epochs=8, validation_data = (np.array(test), np.array(lt)))

aug = ImageDataGenerator(rotation_range=20, width_shift_range=0.2, height_shift_range=0.2, shear_range=0.15,
    horizontal_flip=True)

# train the network
model.fit_generator(aug.flow(data, labels, batch_size=32),
    validation_data=(np.array(test), np.array(lt)), steps_per_epoch=len(data) // 32,
    epochs=7)

这是返回:

Epoch 1/7
43/43 [==============================] - 1004s 23s/step - loss: 1.8090 - acc: 0.9724 - val_loss: 1.7871 - val_acc: 0.9861
Epoch 2/7
43/43 [==============================] - 1003s 23s/step - loss: 1.8449 - acc: 0.9801 - val_loss: 1.4828 - val_acc: 1.0000
Epoch 3/7
43/43 [==============================] - 1092s 25s/step - loss: 1.5704 - acc: 0.9920 - val_loss: 1.3985 - val_acc: 1.0000
Epoch 4/7
43/43 [==============================] - 1062s 25s/step - loss: 1.5219 - acc: 0.9898 - val_loss: 1.3167 - val_acc: 1.0000
Epoch 5/7
43/43 [==============================] - 990s 23s/step - loss: 2.5744 - acc: 0.9222 - val_loss: 2.9347 - val_acc: 0.9028
Epoch 6/7
43/43 [==============================] - 983s 23s/step - loss: 1.6053 - acc: 0.9840 - val_loss: 1.3299 - val_acc: 1.0000
Epoch 7/7
43/43 [==============================] - 974s 23s/step - loss: 1.6180 - acc: 0.9801 - val_loss: 1.5181 - val_acc: 0.9861

我添加了 dropout，减少了模型层数，添加了数据增强，但根本不起作用(所有预测都返回 0)...

请任何人都可以帮忙。

Answer 1

您似乎遇到了过度拟合的问题。我在这里遇到了一种个人困境，无论这是否偏离主题，因为人们可以给出的方法在某种程度上是基于意见的，但我在这里:首先，如果您需要规范过度拟合网络，则需要从 0.25 开始使用 dropout 并检查这是否可以改进模型。在处理过度拟合时，数据增强和批量归一化(您正在应用)是必须的。如果这仍然不能解决您的过度拟合问题，那么您应该尝试对网络架构进行调整，以使其更好地泛化。您是否检查了用于训练和测试的简单输入？

TLDR:尝试 dropout 和数据增强，如果它不起作用并且您的数据正确，您可能需要努力改进架构以建立更好的泛化模型。

编辑:处理此类模型的共识是首先使其过拟合，并具有良好的准确性，然后在可能的情况下努力在不损失准确性的情况下对其进行泛化。