许多CNN网络都有Fusion layer作为融合层,比如:

解决keras.backend.reshape中的错误ValueError: Tried to convert 'shape' to a tensor and failed. Error: Cannot convert a partially known TensorShape to a Tensor

 

 

 

 

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

相关代码:(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)

当我实际去做的时候, K.reshape 报错: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)

相关信息:

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从(None, 256, 512) 变成(None, 16, 16, 512) 就出现上述错误。

解决过程:

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

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

但又出现错误:AttributeError 'NoneType' object has no attribute '_inbound_nodes'

原来是因为:“只要使用Model,就必须保证该函数内全为layer而不能有其他函数,如果有其他函数必须用Lambda封装为layer。”

参考: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)

注意这里的维度必须是( reshaped_shape[1], reshaped_shape[2], reshaped_shape[3]) 而不是 ( -1, reshaped_shape[1], reshaped_shape[2], reshaped_shape[3])

不然会出错