python - Keras Predict_proba 中的神经网络始终返回等于 1 的概率

Question

我正在学习 ML、MNIST 集上的神经网络，但我对 Predict_proba 函数有疑问。我想接收模型做出的预测的概率，但是当我调用函数 Predict_proba 时，我总是收到像 [0, 0, 1., 0, 0, ...] 这样的数组，这意味着模型总是以 100% 的概率进行预测。

您能告诉我我的模型出了什么问题吗？为什么会发生这种情况以及如何解决？

我的模型看起来像:

# Load MNIST data set and split to train and test sets
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# Reshaping to format which CNN expects (batch, height, width, channels)
train_images = train_images.reshape(train_images.shape[0], train_images.shape[1], train_images.shape[2], 1).astype(
    "float32")
test_images = test_images.reshape(test_images.shape[0], test_images.shape[1], test_images.shape[2], 1).astype("float32")

# Normalize images from 0-255 to 0-1
train_images /= 255
test_images /= 255

# Use one hot encode to set classes
number_of_classes = 10

train_labels = keras.utils.to_categorical(train_labels, number_of_classes)
test_labels = keras.utils.to_categorical(test_labels, number_of_classes)

# Create model, add layers
model = Sequential()
model.add(Conv2D(32, (5, 5), input_shape=(train_images.shape[1], train_images.shape[2], 1), activation="relu"))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(32, (3, 3), activation="relu"))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(128, activation="relu"))
model.add(Dropout(0.5))
model.add(Dense(number_of_classes, activation="softmax"))

# Compile model
model.compile(loss="categorical_crossentropy", optimizer=Adam(), metrics=["accuracy"])

# Learn model
model.fit(train_images, train_labels, validation_data=(test_images, test_labels), epochs=7, batch_size=200)

# Test obtained model
score = model.evaluate(test_images, test_labels, verbose=0)
print("Model loss = {}".format(score[0]))
print("Model accuracy = {}".format(score[1]))

# Save model
model_filename = "cnn_model.h5"
model.save(model_filename)
print("CNN model saved in file: {}".format(model_filename))

为了加载图像，我使用 PIL 和 NP。 我使用 keras 中的 save 函数保存模型，并使用 keras.models 中的 load_model 将其加载到另一个脚本中，然后我只需调用

    def load_image_for_cnn(filename):
        img = Image.open(filename).convert("L")
        img = np.resize(img, (28, 28, 1))
        im2arr = np.array(img)
        return im2arr.reshape(1, 28, 28, 1)

    def load_cnn_model(self):
        return load_model("cnn_model.h5")

    def predict_probability(self, image):
        return self.model.predict_proba(image)[0]

使用它看起来像:

predictor.predict_probability(predictor.load_image_for_cnn(filename))

Answer 1

看看你的代码的这一部分:

# Normalize images from 0-255 to 0-1
train_images /= 255
test_images /= 255

加载新图像时您没有执行此操作:

def load_image_for_cnn(filename):
    img = Image.open(filename).convert("L")
    img = np.resize(img, (28, 28, 1))
    im2arr = np.array(img)
    return im2arr.reshape(1, 28, 28, 1)

测试任何新图像都需要应用与训练集相同的归一化，如果不这样做，就会得到奇怪的结果。您可以按如下方式标准化图像像素:

def load_image_for_cnn(filename):
    img = Image.open(filename).convert("L")
    img = np.resize(img, (28, 28, 1))
    im2arr = np.array(img)
    im2arr = im2arr / 255.0
    return im2arr.reshape(1, 28, 28, 1)