python - 如何使用 keras 和 tensorflow 后端将密集层的输出作为 numpy 数组？

Question

我是 Keras 和 Tensorflow 的新手。我正在使用深度学习进行人脸识别项目。我正在使用此代码(softmax 层 的输出)将主题的类标签作为输出，对于我的 100 个类的自定义数据集，准确度为 97.5%。

但现在我对特征向量表示感兴趣，所以我想通过网络传递测试图像，并在 softmax(最后一层)之前从激活的密集层提取输出。我引用了 Keras 文档，但似乎对我没有任何帮助。谁能帮我如何从密集层激活中提取输出并保存为 numpy 数组？提前致谢。

class Faces:
    @staticmethod
    def build(width, height, depth, classes, weightsPath=None):
        # initialize the model
        model = Sequential()
        model.add(Conv2D(100, (5, 5), padding="same",input_shape=(depth, height, width), data_format="channels_first"))
        model.add(Activation("relu"))
        model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2),data_format="channels_first"))

        model.add(Conv2D(100, (5, 5), padding="same"))
        model.add(Activation("relu"))
        model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), data_format="channels_first"))

        # 3 set of CONV => RELU => POOL
        model.add(Conv2D(100, (5, 5), padding="same"))
        model.add(Activation("relu"))
        model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2),data_format="channels_first"))

        # 4 set of CONV => RELU => POOL
        model.add(Conv2D(50, (5, 5), padding="same"))
        model.add(Activation("relu"))
        model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2),data_format="channels_first"))

        # 5 set of CONV => RELU => POOL
        model.add(Conv2D(50, (5, 5), padding="same"))
        model.add(Activation("relu"))
        model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), data_format="channels_first"))

        # 6 set of CONV => RELU => POOL
        model.add(Conv2D(50, (5, 5), padding="same"))
        model.add(Activation("relu"))
        model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), data_format="channels_first"))

        # set of FC => RELU layers
        model.add(Flatten())
        #model.add(Dense(classes))
        #model.add(Activation("relu"))

        # softmax classifier
        model.add(Dense(classes))
        model.add(Activation("softmax"))

        return model

ap = argparse.ArgumentParser()
ap.add_argument("-l", "--load-model", type=int, default=-1,
    help="(optional) whether or not pre-trained model should be loaded")
ap.add_argument("-w", "--weights", type=str,
    help="(optional) path to weights file")
args = vars(ap.parse_args())


path = 'C:\\Users\\Project\\FaceGallery'
image_paths = [os.path.join(path, f) for f in os.listdir(path)]
images = []
labels = []
name_map = {}
demo = {}
nbr = 0
j = 0
for image_path in image_paths:
    image_pil = Image.open(image_path).convert('L')
    image = np.array(image_pil, 'uint8')
    cv2.imshow("Image",image)
    cv2.waitKey(5)
    name = image_path.split("\\")[4][0:5]
    print(name)
    # Get the label of the image
    if name in demo.keys():
        pass
    else:
        demo[name] = j
        j = j+1
    nbr =demo[name]

    name_map[nbr] = name
    images.append(image)
    labels.append(nbr)
print(name_map)
# Training and testing data split ratio = 60:40
(trainData, testData, trainLabels, testLabels) = train_test_split(images, labels, test_size=0.4)

trainLabels = np_utils.to_categorical(trainLabels, 100)
testLabels = np_utils.to_categorical(testLabels, 100)

trainData = np.asarray(trainData)
testData = np.asarray(testData)

trainData = trainData[:, np.newaxis, :, :] / 255.0
testData = testData[:, np.newaxis, :, :] / 255.0

opt = SGD(lr=0.01)
model = Faces.build(width=200, height=200, depth=1, classes=100,
                    weightsPath=args["weights"] if args["load_model"] > 0 else None)

model.compile(loss="categorical_crossentropy", optimizer=opt, metrics=["accuracy"])
if args["load_model"] < 0:
    model.fit(trainData, trainLabels, batch_size=10, epochs=300)
(loss, accuracy) = model.evaluate(testData, testLabels, batch_size=100, verbose=1)
print("Accuracy: {:.2f}%".format(accuracy * 100))
if args["save_model"] > 0:
    model.save_weights(args["weights"], overwrite=True)

for i in np.arange(0, len(testLabels)):
    probs = model.predict(testData[np.newaxis, i])
    prediction = probs.argmax(axis=1)
    image = (testData[i][0] * 255).astype("uint8")
    name = "Subject " + str(prediction[0])
    if prediction[0] in name_map:
        name = name_map[prediction[0]]
    cv2.putText(image, name, (5, 20), cv2.FONT_HERSHEY_PLAIN, 1.3, (255, 255, 255), 2)
    print("Predicted: {}, Actual: {}".format(prediction[0], np.argmax(testLabels[i])))
    cv2.imshow("Testing Face", image)
    cv2.waitKey(1000)

Answer 1

参见 https://keras.io/getting-started/faq/ 如何获取中间层的输出？

您需要通过在定义中添加“名称”参数来命名您想要输出的图层。喜欢.. model.add(Dense(xx, name='my_dense'))
然后，您可以定义一个中间模型并通过执行诸如...之类的操作来运行它。

m2 = Model(inputs=model.input, outputs=model.get_layer('my_dense').output)
Y = m2.predict(X)