许多CNN网络都有Fusion layer作为融合层,比如:
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。
解决方法:
fuse = K.reshape(fuse, (-1, reshaped_shape[1], reshaped_shape[2], reshaped_shape[3]))
参考:https://github.com/matterport/Mask_RCNN/issues/1070