python - 尝试在烤宽面条基本示例(稍作修改)上使用自己的数据集来制作用于人脸识别的神经网络

Question

我正在做基本的烤宽面条示例: https://github.com/Lasagne/Lasagne/blob/master/examples/mnist.py

我通过将其与另一个类似的示例结合起来对其进行了轻微修改。

我正在尝试运行 CNN 模型，我在 CNN def 中添加了一些额外的输入，但它不应该有任何不同。还将示例中输入层的默认值 28 更改为 60(对于高度和宽度)，稍后在代码中使用类，但代码“卡在”最后一个网络行，这意味着代码是仍在运行，但没有任何反应。 Output when running the code. input_var 在主循环中定义如下:

input_var = T.tensor4('input_var')

其余代码:

def build_cnn(classes, height, width, input_var=None):

    print("Input layer, with height: {}, width: {} and input var: {}".format(height, width, input_var))

    network = lasagne.layers.InputLayer(shape = (None, 1, height, width),
                                    input_var=input_var)


    print("Convolutional layer with 32 kernels of size 5x5")
    network = lasagne.layers.Conv2DLayer(network,
                                         num_filters = 32,
                                         filter_size = (5, 5),
                                         nonlinearity = lasagne.nonlinearities.rectify,
                                         W = lasagne.init.HeNormal(gain = 'relu')) 

编辑:好吧，根据我迄今为止的尝试，问题似乎出在我自己的数据集上。我已经重新调整了我的数据集以匹配 MNIST 数据集(例如)。 X_train 的形状为 [图像、 channel 、高度、宽度]。其中 channel = 1，高度、宽度 = 60。检索这些的代码是:

def load_images():
    dataset_path = os.path.abspath("C:/Users/laende/Dropbox/Skole UiS/4. semester/Master/Master/data/test_database")
    [bilder, label, names] = read_images1(dataset_path, (28, 28))
    label = np.array(label)

    bilder = bilder / np.float32(256)
    bilder = bilder[:, newaxis, :, :]

    X_train1, X_test1, Y_train1, Y_test1 = train_test_split(bilder, label, test_size = 0.2)

    list_of_labels = list(xrange(max(label) + 1))
    classes = len(list_of_labels)

    return X_train1, X_test1, Y_train1, Y_test1, classes

其中 read_images1 是:

 def read_images1(path, sz = None, channel = None):
    c = 0
    X = []
    y = []
    folder_names = []
    for dirname, dirnames, filenames, in os.walk(path):
        for subdirname in dirnames:
            subject_path = os.path.join(dirname, subdirname)
            folder_names.append(subdirname)
            for filename in os.listdir(subject_path):
                try:
                    im = cv2.imread(os.path.join(subject_path, filename),     cv2.IMREAD_GRAYSCALE)

                    if (sz is not None):
                        im = cv2.resize(im, sz)

                    X.append(np.asarray(im, dtype = np.uint8))
                    y.append(c)

                except IOError, (errno, strerror):
                    print "I/O error ({0]): {1}".format(errno, strerror)
                except:
                    print "unexpected error:", sys.exc_info()[0]
                    raise
            c = c + 1
    return [X, y, folder_names]

main 中运行的代码:

def main(model='mlp', num_epochs=100):
    # Load the dataset

    print("Loading data...")
    mnist = 1
    if mnist == 1:
        classes = 10
        X_train, y_train, X_val, y_val, X_test, y_test = load_dataset()

        dataset = {
            'train': {'X': X_train, 'y': y_train},
            'test': {'X': X_test, 'y': y_test}}
        shape = dataset['train']['X'][0].shape

    else:
        X_train, X_test, y_train, y_test, classes = load_images()

        dataset = {
            'train': {'X': X_train, 'y': y_train},
            'test': {'X': X_test, 'y': y_test}}
        shape = dataset['train']['X'][0].shape

    input_var = T.tensor4('inputs')
    target_var = T.ivector('targets')

    print("Building model and compiling functions...")
    if model == 'mlp':
        network = build_mlp(height=int(shape[1]),
                            width=int(shape[2]),
                            channel=int(shape[0]),
                            classes=int(classes),
                            input_var=input_var)

如果 mnist = 1(在 main 中)，代码运行正常，如果我尝试使用自己的数据集，它会卡在 build_mlp 中(类似于 cnn 的原始问题):

def build_mlp(classes, channel, height, width, input_var=None):

    neurons = int(height * width)

    network = lasagne.layers.InputLayer(shape = (None, channel, height, width),
                                 input_var=input_var)

    network = lasagne.layers.DropoutLayer(network, p = 0.2)

   #Code gets stuck on this point, running forever, doing nothing.
   #No error messages received either.
    network = lasagne.layers.DenseLayer(
        network,
        num_units = neurons,
        nonlinearity = lasagne.nonlinearities.rectify,
        W = lasagne.init.GlorotUniform())

编辑2:经过一段时间的挣扎后，我发现在 read_images1() 中完成的图像大小调整导致了问题:

 def read_images1(path, sz = None, channel = None):
    c = 0
    X = []
    y = []
    folder_names = []
    for dirname, dirnames, filenames, in os.walk(path):
        for subdirname in dirnames:
            subject_path = os.path.join(dirname, subdirname)
            folder_names.append(subdirname)
            for filename in os.listdir(subject_path):
                try:
                    im = cv2.imread(os.path.join(subject_path, filename),     cv2.IMREAD_GRAYSCALE)
                    #This part caused the problems. 
                    if (sz is not None):
                        im = cv2.resize(im, sz)

                    X.append(np.asarray(im, dtype = np.uint8))
                    y.append(c)

                except IOError, (errno, strerror):
                    print "I/O error ({0]): {1}".format(errno, strerror)
                except:
                    print "unexpected error:", sys.exc_info()[0]
                    raise
            c = c + 1
    return [X, y, folder_names]

如果我没有通过任何大小调整并使用文件夹中的默认图像大小，神经网络就能够编译。有谁知道为什么吗？我将 read_images1() 更新为:

 def read_images1(path, sz = None, na = False):
    """

    :param path: sti til mappe med underliggende mapper tilhørende personer.
    :param sz: Størrelse på bildefilene
    :return: returnerer liste av bilder, labels og navn
    """
    c = 0
    X = []
    y = []
    folder_names = []
    for dirname, dirnames, filenames, in os.walk(path):
        for subdirname in dirnames:
            subject_path = os.path.join(dirname, subdirname)
            folder_names.append(subdirname)
            for filename in os.listdir(subject_path):
                try:
                    im = cv2.imread(os.path.join(subject_path, filename), cv2.IMREAD_GRAYSCALE)

                    if (sz is not None):
                        im = cv2.resize(im, dsize=sz, interpolation = cv2.INTER_LANCZOS4)

                    if (na == True):
                        im = im[newaxis, :, :]


                    X.append(np.asarray(im, dtype = np.uint8))
                    y.append(c)

                except IOError, (errno, strerror):
                    print "I/O error ({0]): {1}".format(errno, strerror)
                except:
                    print "unexpected error:", sys.exc_info()[0]
                    raise
            c = c + 1
    return [X, y, folder_names] 

如果我使用 sz = None 和 na = True 运行程序，则它可以工作。如果为 sz 参数指定任何大小，则代码在尝试再次编译神经网络时会被卡住。

Answer 1

好吧，我想我可以在这里看到一些问题，不确定您遇到的是哪一个......

1. 在 read_images1() 中，X 是 numpy 数组的 Python 列表。它在哪里转换为单个 numpy 数组？尝试添加 X = numpy.asarray(X) 。您还需要将其 reshape 为 (n_images, n_channels, width, height)，其中我假设 n_channels=1 表示灰度。网络需要 4D 输入而不是 3D。

2. 此代码list_of_labels = list(xrange(max(label) + 1)); classes = len(list_of_labels) 假设标签是从 0 到 N 的连续数字。是吗？

3. 您的 build_mlp(classes, height, width, input_var=None) 与原始示例 build_mlp(input_var=None) 非常不同。原始示例显然有效，因此无论出现什么问题都与差异有关。最大的区别之一是您不断分配给同一个变量，就像这样 network = lasagne.layers.DenseLayer(network, ...) ，其中原始的每一层都有不同的变量 l_hid1 = lasagne.layers.DenseLayer(l_in_drop, ...)

4. 此外，如果它在 build_mlp() 期间挂起，那么显然问题不在于您如何读取图像。尝试将原始版本的 build_mlp() 与您的图像一起使用。尝试单独运行它。跳过图像读取，只需使用常量参数调用 build_mlp() 即可。