Right here is a dinky PyTorch-essentially based equipment which enables for atmosphere pleasurable batched operations, e.g. for computing distances with out having to slowly loop over all instance pairs of a batch of records.
After having encountered mulitple conditions of torch modules/methods promising to handling batches whereas most attention-grabbing returning a vector of pairwise results (discover about instance beneath) as an alternate of the paunchy matrix, this equipment serves as a tool to wrap such methods in shriek to return paunchy matrices (e.g. distance matrices) the utilize of lickety-split, batched operations (with out loops).
First, let’s account for a customized distance aim that most attention-grabbing computes pair-wise distances for batches, so two batches of every 10 samples are
remodeled to a distance vector of shape (10,).
>>> def dummy_distance(x,y): """ Right here is a dummy distance d which enables for a batch dimension (bid with n conditions in a batch), however doesn't return the paunchy n x n distance matrix however most attention-grabbing a n-dimensional vector of the pair-wise distances d(x_i,y_i) for all i in (1,...,n). """ x_ = x.sum(axis=[1,2]) y_ = y.sum(axis=[1,2]) return x_ + y_ # batchdist wraps a torch module around this callable to compute # the paunchy n x n matrix with batched operations (no loops). >>> import batchdist as bd >>> batched = bd.BatchDistance(dummy_distance) # generate records (two batches of 256 samples of dimension [4,3]) >>> x1 = torch.rand(256,4,3) >>> x2 = torch.rand(256,4,3) >>> out1 = batched(x1, x2) # distance matrix of shape [256,256]
For more dinky print, consult the incorporated examples.
$ pip install batchdist