预输入层到预训练模型

Question

我正在尝试将形状(1036800 )--最初的形状图像(480,720，3) --输入到经过预先训练的VGG16模型中，以预测连续值。

我尝试了以下几种代码的变体：

input = Input(shape=(1036800,), name='image_input')
initial_model = VGG16(weights='imagenet', include_top=False)
x = Flatten()(initial_model(input).output)
x = Dense(200, activation='relu')(x)
x = Dense(1)(x)
model = Model(inputs=input, outputs=x)

上述代码的先前变体产生了与输入相关的错误，即错误的维度，input_shape需要有3个通道(在VGG16初始化时使用(1036800 )该参数)，运行上述代码导致的最新错误如下：

Traceback (most recent call last):
  File "model_alex.py", line 57, in <module>
    model = initialize_model()
  File "model_alex.py", line 20, in initialize_model
    x = Flatten()(initial_model(input).output)
  File "/home/aicg2/.local/lib/python2.7/site-packages/keras/engine/topology.py", line 596, in __call__
    output = self.call(inputs, **kwargs)
  File "/home/aicg2/.local/lib/python2.7/site-packages/keras/engine/topology.py", line 2061, in call
    output_tensors, _, _ = self.run_internal_graph(inputs, masks)
  File "/home/aicg2/.local/lib/python2.7/site-packages/keras/engine/topology.py", line 2212, in run_internal_graph
    output_tensors = _to_list(layer.call(computed_tensor, **kwargs))
  File "/home/aicg2/.local/lib/python2.7/site-packages/keras/layers/convolutional.py", line 164, in call
    dilation_rate=self.dilation_rate)
  File "/home/aicg2/.local/lib/python2.7/site-packages/keras/backend/tensorflow_backend.py", line 3156, in conv2d
    data_format='NHWC')
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/nn_ops.py", line 639, in convolution
    input_channels_dim = input.get_shape()[num_spatial_dims + 1]
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/tensor_shape.py", line 500, in __getitem__
    return self._dims[key]
IndexError: list index out of range

这是全码。下面是脚本中使用的样本数据文件。

解决这一问题的可能方法之一是将原始图像文件的大小调整为224x224，并将它们转换为形状为numpy的数组(224、224、3)，以便将它们插入经过预先训练的模型的第一层。然而，我不想扭曲图像或浪费另一个晚上的预处理数据，而我应该已经在训练。

除此之外，我所能做的就是谷歌我的问题，试着去适应找到的解决方案，或者毫无目标地调整各种形状相关的参数和功能--在过去的4小时里，这两种方法都没有给我带来多大的帮助。

Answer 1

问题是，您不应该将图像压平成一维矢量，因为VGG16包含2D卷积层(例如，图像上的空间卷积)，这需要输入具有(number_of_images, image_height, image_width, image_channels)形状，因为keras.backend.image_data_format()返回'channels_last'。如果您的image_data_format是'channels_first'，则将输入数据形状更改为(number_of_images, image_channels, image_height, image_width)。

下面是您的固定代码(用Keras2.0.4测试)：

x_train = x_train.reshape((x_train.shape[0], 480, 720, 3))
x_test = x_test.reshape((x_test.shape[0], 480, 720, 3))

initial_model = VGG16(weights='imagenet', include_top=False)
input = Input(shape=(480, 720, 3), name='image_input')
x = Flatten()(initial_model(input))
x = Dense(200, activation='relu')(x)
x = BatchNormalization()(x)
x = Dropout(0.5)(x)
x = Dense(1)(x)
model = Model(inputs=input, outputs=x)
model.compile(loss='mse', optimizer='adam')

model.fit(x_train, y_train, epochs=20, batch_size=16)
score = model.evaluate(x_test, y_test, batch_size=16)