Can caffe take in a batch of inputs with different resolutions at once? If so how?

Question

I am thinking of building a larger network with more input layers which would consist of all the respective reshaped input resolutions from the batch. Please let me know if this is possible and if it is what would be the most efficient way?

For example: 3x160x160 + 3x48x48 + 3x128x128 | | | Rest of the network | | | Outputs

Answer 1

Caffe processes Blobs : these are N -dimensional arrays, therefore the dimensions of all elements in a single batch must be of the same shape (you can reshape between batches).
However, if you are going to use the same multiple shapes for all inputs in the batch, ie, you'll have inputs like N x3x160x160 + N x3x48x48 + N x3x128x128, then you can create three input layer (one for each shape) and feed the net with three types of blobs at each iteration.