这个CNN对40x40x2的图像效果很好，但现在我想改为13x78x2，结果出现了以下错误。我应该如何修改我的CNN架构？

Negative dimension size caused by subtracting 3 from 2 for 'conv2d_13/convolution' (op: 'Conv2D') with input shapes: [?,2,35,64], [3,3,64,64].

我的代码如下：

  data_w = 40 #CHANGE TO 13  data_h = 40 #CHANGE TO 78  n_classes = 2  n_filters_1 = 32  n_filters_2 = 64  d_filter = 3  p_drop_1 = 0.25  p_drop_2 = 0.50     model = Sequential()  model.add(Convolution2D(n_filters_1, d_filter, d_filter, border_mode='valid', input_shape=(data_w, data_h,2)))  model.add(Activation('relu'))  model.add(Convolution2D(n_filters_1, d_filter, d_filter))  model.add(Activation('relu'))  model.add(MaxPooling2D(pool_size=(2, 2)))  model.add(Dropout(p_drop_1))  model.add(Convolution2D(n_filters_2, d_filter, d_filter, border_mode='valid'))  model.add(Activation('relu'))  model.add(Convolution2D(n_filters_2, d_filter, d_filter))  model.add(Activation('relu'))  model.add(MaxPooling2D(pool_size=(2, 2)))  model.add(Dropout(p_drop_1))  ## Used to flat the input (1, 10, 2, 2) -> (1, 40)  model.add(Flatten())  # Full Connected layer  model.add(Dense(256))  model.add(Activation('relu'))  # Drop layer  model.add(Dropout(p_drop_2))  # Output Full Connected layer  model.add(Dense(n_classes))  model.add(Activation('softmax'))

回答：

由于你选择了valid作为卷积的border_mode，这意味着在每个Convolution2D层中，使用3×3的滤波器大小会导致边界周围各去掉1个像素。此外，如果省略该参数，也会默认使用valid填充。如果你计算出每一层输出的尺寸减少情况，你会发现输出滤波器尺寸的一个维度（行）将会变为0，从而引发错误。假设d_filter = 3，我们来逐层查看输入图像尺寸为13×78时的输出滤波器尺寸。请注意，我省略了在Activation和Dropout层显示滤波器尺寸输出，因为我们已经知道它们保持输入的输出尺寸一致，以简化说明：

  model.add(Convolution2D(n_filters_1, d_filter, d_filter, border_mode='valid', input_shape=(data_w, data_h,2))) # 11 x 76  model.add(Activation('relu'))  model.add(Convolution2D(n_filters_1, d_filter, d_filter)) # 9 x 74  model.add(Activation('relu'))  model.add(MaxPooling2D(pool_size=(2, 2))) # 4 x 37  model.add(Dropout(p_drop_1))  model.add(Convolution2D(n_filters_2, d_filter, d_filter, border_mode='valid')) # 2 x 35  model.add(Activation('relu'))  model.add(Convolution2D(n_filters_2, d_filter, d_filter)) # 0 x 33 (!!!!)  model.add(Activation('relu'))  model.add(MaxPooling2D(pool_size=(2, 2)))  model.add(Dropout(p_drop_1))

我建议你立即更改border_mode为'same'。这样，每个Convolution2D层在到达Pooling层之前可以保持输出滤波器尺寸。我不确定你选择valid卷积的目的，但可以尝试这样做：

  model.add(Convolution2D(n_filters_1, d_filter, d_filter, border_mode='same', input_shape=(data_w, data_h,2))) # 13 x 78  model.add(Activation('relu'))  model.add(Convolution2D(n_filters_1, d_filter, d_filter), border_mode='same') # 13 x 78  model.add(Activation('relu'))  model.add(MaxPooling2D(pool_size=(2, 2))) # 6 x 39  model.add(Dropout(p_drop_1))  model.add(Convolution2D(n_filters_2, d_filter, d_filter, border_mode='same')) # 6 x 39  model.add(Activation('relu'))  model.add(Convolution2D(n_filters_2, d_filter, d_filter), border_mode='same') # 6 x 39  model.add(Activation('relu'))  model.add(MaxPooling2D(pool_size=(2, 2))) # 3 x 19  model.add(Dropout(p_drop_1))

如果不这样做，你需要删除一些Convolution2D和MaxPooling2D层，以便生成非零的滤波器输出。像我上面一样计算你需要删除多少层。我建议删除第一组使用n_filters_2滤波器的Convolution2D和Activation层之后的所有层：

  model.add(Convolution2D(n_filters_1, d_filter, d_filter, border_mode='valid', input_shape=(data_w, data_h,2))) # 11 x 76  model.add(Activation('relu'))  model.add(Convolution2D(n_filters_1, d_filter, d_filter)) # 9 x 74  model.add(Activation('relu'))  model.add(MaxPooling2D(pool_size=(2, 2))) # 4 x 37  model.add(Dropout(p_drop_1))  model.add(Convolution2D(n_filters_2, d_filter, d_filter, border_mode='valid')) # 2 x 35  model.add(Activation('relu'))#  model.add(Convolution2D(n_filters_2, d_filter, d_filter)) # 0 x 33 (!!!!)#  model.add(Activation('relu'))#  model.add(MaxPooling2D(pool_size=(2, 2)))#  model.add(Dropout(p_drop_1))