Problem with the shape between ConvLSTM and MobileNet
stefanobini opened this issue · comments
The error is:
_
ValueError: Cannot reshape a tensor with 25690112 elements to shape [8,28,28,256] (1605632 elements) for '{{node clstm_reshape/Reshape}} = Reshape[T=DT_FLOAT, Tshape=DT_INT32](Placeholder, clstm_reshape/Reshape/shape)' with input shapes: [8,16,28,28,256], [4] and with input tensors computed as partial shapes: input[1] = [8,28,28,256].
_
The problem is that the output of the ConvLSTM, that is made up of 'depth' frames, it is given as input to a 2D-CNN (MobileNet) that can be work with only one frame. Obviously the problem can be easly solved by duplicate 'depth' time the number of 2D-CNN and then fuse the results. So I would like to know if it is just an oversight or the code is obsolete and therefore it is not possible to reproduce the results of the paper.