keras：ValueError：检查模型目标时出错：预期activation_1具有形状（None，60）但得到形状为（10、100）的数组

Question

我正在尝试移植RocAlphaGo 来玩亚马逊游戏，在尝试实施监督策略训练器时出现问题。

from keras.models import Sequential, Model
from keras.layers.core import Activation, Flatten
from keras.layers import convolutional

defaults = {
            "board": 10,
            "filters_per_layer": 128,
            "layers": 12,
            "filter_width_1": 5
}
# copy defaults, but override with anything in kwargs
params = defaults
network = Sequential()
# create first layer
network.add(convolutional.Convolution2D(
      input_shape=(6, 10, 10),
      nb_filter=128,
      nb_row=5,
      nb_col=5,
      init='uniform',
      activation='relu',
      border_mode='same'))

# create all other layers
for i in range(2, 13):
     # use filter_width_K if it is there, otherwise use 3
     filter_key = "filter_width_%d" % i
     filter_width = params.get(filter_key, 3)

     # use filters_per_layer_K if it is there, otherwise use default value
     filter_count_key = "filters_per_layer_%d" % i
     filter_nb = params.get(filter_count_key, 128)

     network.add(convolutional.Convolution2D(
         nb_filter=filter_nb,
         nb_row=filter_width,
         nb_col=filter_width,
         init='uniform',
         activation='relu',
         border_mode='same'))

# the last layer maps each <filters_per_layer> feature to a number
network.add(convolutional.Convolution2D(
    nb_filter=1,
    nb_row=1,
    nb_col=1,
    init='uniform',
    border_mode='same'))
# reshape output to be board x board
network.add(Flatten())
# softmax makes it into a probability distribution
network.add(Activation('softmax'))

• keras 1.2.0
• python 2.7

给出以下异常：

ValueError：检查模型目标时出错：预期activation_1 到 有形状 (None, 60) 但得到了形状 (10, 100) 的数组

训练数据集是一个 (10, 6, 10, 10) 数组，10 x 6 层，每层是一个 10x10 数组（棋盘），为什么模型需要 (None, 60) ？ 如果 chagne input_shape=(6, 10, 10) 到 input_shape=(10, 10, 10)，将得到：

ValueError：检查模型输入时出错：预期 convolution2d_input_1 具有形状 (None, 10, 10, 10) 但得到了数组 形状为 (10, 6, 10, 10)

所有代码都是here

Answer 1

正如 Matias 在 cmets 中所说，如果你添加

network.summary() 

您可能会注意到您的卷积应用于输入数据的前两个维度（即 (6,10,10)）。你在这里的特征是在第一维。默认情况下，当您使用 tensorflow 时，Keras 会假设您的特征所在的维度是第三个维度，而不是第一个维度。因此，当您在 (6,10,10) 数组上应用 128 个过滤器时，输出将是 (6,10,128)，如果我理解得很好，这不是您想要的。

因此，在最后一个卷积层的输出处，您会得到一个 (6,10,1) 数组，该数组平展为 (,60)，而不是您预期的 (,100)。

有两种方法可以修复您的网络。您可以将输入数据更改为 (10,10,6) 格式。或者你使用Convolution2D()层的data_format="channels_first"参数。

这是第二个选项的代码：

from keras.models import Sequential, Model
from keras.layers.core import Activation, Flatten
from keras.layers import convolutional

defaults = {
            "board": 10,
            "filters_per_layer": 128,
            "layers": 12,
            "filter_width_1": 5
}
# copy defaults, but override with anything in kwargs
params = defaults
network = Sequential()
# create first layer
network.add(convolutional.Convolution2D(
      input_shape=(6, 10, 10),
      nb_filter=128,
      nb_row=5,
      nb_col=5,
      init='uniform',
      activation='relu',
      border_mode='same',
      data_format='channels_first'
      ))

# create all other layers
for i in range(2, 13):
     # use filter_width_K if it is there, otherwise use 3
     filter_key = "filter_width_%d" % i
     filter_width = params.get(filter_key, 3)

     # use filters_per_layer_K if it is there, otherwise use default value
     filter_count_key = "filters_per_layer_%d" % i
     filter_nb = params.get(filter_count_key, 128)

     network.add(convolutional.Convolution2D(
         nb_filter=filter_nb,
         nb_row=filter_width,
         nb_col=filter_width,
         init='uniform',
         activation='relu',
         border_mode='same',
         data_format='channels_first'))

# the last layer maps each <filters_per_layer> feature to a number
network.add(convolutional.Convolution2D(
    nb_filter=1,
    nb_row=1,
    nb_col=1,
    init='uniform',
    border_mode='same',
    data_format='channels_first'))
# reshape output to be board x board
network.add(Flatten())
# softmax makes it into a probability distribution
network.add(Activation('softmax'))
# display your network summary
network.summary()

编辑

考虑到您的 keras 版本，您应该使用参数“dim_ordering”并将其设置为“th”。

我在keras documentation找到了这个信息

from keras.models import Sequential, Model
from keras.layers.core import Activation, Flatten
from keras.layers import convolutional

defaults = {
            "board": 10,
            "filters_per_layer": 128,
            "layers": 12,
            "filter_width_1": 5
}
# copy defaults, but override with anything in kwargs
params = defaults
network = Sequential()
# create first layer
network.add(convolutional.Convolution2D(
      input_shape=(6, 10, 10),
      nb_filter=128,
      nb_row=5,
      nb_col=5,
      init='uniform',
      activation='relu',
      border_mode='same',
      dim_ordering='th'
      ))

# create all other layers
for i in range(2, 13):
     # use filter_width_K if it is there, otherwise use 3
     filter_key = "filter_width_%d" % i
     filter_width = params.get(filter_key, 3)

     # use filters_per_layer_K if it is there, otherwise use default value
     filter_count_key = "filters_per_layer_%d" % i
     filter_nb = params.get(filter_count_key, 128)

     network.add(convolutional.Convolution2D(
         nb_filter=filter_nb,
         nb_row=filter_width,
         nb_col=filter_width,
         init='uniform',
         activation='relu',
         border_mode='same',
         dim_ordering='th'))

# the last layer maps each <filters_per_layer> feature to a number
network.add(convolutional.Convolution2D(
    nb_filter=1,
    nb_row=1,
    nb_col=1,
    init='uniform',
    border_mode='same',
    dim_ordering='th'))
# reshape output to be board x board
network.add(Flatten())
# softmax makes it into a probability distribution
network.add(Activation('softmax'))
# display your network summary
network.summary()