解決keras.backend.reshape中的錯誤ValueError: Tried to convert 'shape' to a tensor and failed. Error: Cannot convert a partially known TensorShape to a Tensor

許多CNN網絡都有Fusion layer作為融合層，比如：

 

 

 

 

 參考：https://arxiv.org/pdf/1712.03400.pdf

相關代碼：（https://github.com/baldassarreFe/deep-koalarization/blob/master/src/koalarization/fusion_layer.py）

class FusionLayer(Layer):
    def call(self, inputs, mask=None):
        imgs, embs = inputs
        reshaped_shape = imgs.shape[:3].concatenate(embs.shape[1])
        embs = K.repeat(embs, imgs.shape[1] * imgs.shape[2])
        embs = K.reshape(embs, reshaped_shape)
        return K.concatenate([imgs, embs], axis=3)

當我實際去做的時候， K.reshape 報錯：ValueError: Tried to convert 'shape' to a tensor and failed. Error: Cannot convert a partially known TensorShape to a Tensor

 reshaped_shape = enco_loco.shape[:3].concatenate(enco_glob.shape[1])
    fuse = K.repeat(enco_glob, enco_loco.shape[1]*enco_loco.shape[2])
    fuse = K.reshape(fuse, (reshaped_shape))
    fuse = K.concatenate([enco_loco, fuse], axis=3)

相關信息：

enco_loco:(None, 16, 16, 512)
enco_glob:(None, 512)
reshaped_shape:(None, 16, 16, 512)
enco_glob.shape:(None, 512)
fuse.shape:(None, 256, 512)

最后想把fuse從(None, 256, 512) 變成(None, 16, 16, 512) 就出現上述錯誤。

解決過程：

fuse = K.reshape(fuse, (-1, reshaped_shape[1], reshaped_shape[2], reshaped_shape[3]))

參考：https://github.com/matterport/Mask_RCNN/issues/1070

但又出現錯誤：AttributeError 'NoneType' object has no attribute '_inbound_nodes'

原來是因為：“只要使用Model，就必須保證該函數內全為layer而不能有其他函數，如果有其他函數必須用Lambda封裝為layer。”

參考：https://zhuanlan.zhihu.com/p/138075621

好吧，再改一下：

from keras.layers import  RepeatVector, Reshape
from keras.layers.merge import concatenate

reshaped_shape = enco_loco.shape[:3].concatenate(enco_glob.shape[1])
    fuse = RepeatVector(enco_loco.shape[1]*enco_loco.shape[2])(enco_glob)
    fuse = Reshape(( reshaped_shape[1], reshaped_shape[2], reshaped_shape[3]))(fuse)
    fuse = concatenate([enco_loco, fuse], axis=3)

注意這里的維度必須是( reshaped_shape[1], reshaped_shape[2], reshaped_shape[3]) 而不是 ( -1, reshaped_shape[1], reshaped_shape[2], reshaped_shape[3])

不然會出錯