search

aesara-devs / aesara

Aesara is a Python library for defining, optimizing, and efficiently evaluating mathematical expressions involving multi-dimensional arrays.
https://aesara.readthedocs.io
Other
1.18k stars 154 forks source link

Replace broadcast_like with broadcast_to #159

Open brandonwillard opened 3 years ago

brandonwillard commented 3 years ago

The function broadcast_like can be replaced with the new broadcast_to function in most—if not all—cases.

The broadcast_to function uses view-based broadcasting instead of allocating filled arrays as broadcast_like does, so this change could introduce performance improvements in many general use situations (more specifically, when the basic algebraic optimizations that use broadcast_like are employed).

eigenfoo commented 3 years ago

@brandonwillard I assume this issue is still open, but with the following two files? The original link is broken. If so, I can submit a PR for this.

$ rg broadcast_like
tensor/math_opt.py
45:    broadcast_like,
1810:            new_out = broadcast_like(new_out, out, fgraph)
1835:            return [broadcast_like(1, node.outputs[0], fgraph)]
1837:            return [broadcast_like(node.inputs[0], node.outputs[0], fgraph)]
1881:            ret = broadcast_like(0, node.outputs[0], fgraph)
2033:                return [broadcast_like(0, node.outputs[0], fgraph)]
2058:                return [broadcast_like(rval, node.outputs[0], fgraph)]
2063:                    return [broadcast_like(-1, node.outputs[0], fgraph)]
2065:                    return [broadcast_like(1, node.outputs[0], fgraph)]

tensor/basic_opt.py
170:def broadcast_like(value, template, fgraph, dtype=None):
182:            "broadcast_like currently requires the "
1753:            o = broadcast_like(v, r, fgraph, dtype=v.dtype)
brandonwillard commented 3 years ago

Yes, theano.tensor.opt was split into theano.tensor.math_opt and theano.tensor.basic_opt, so those results were previously together in theano.tensor.opt.