Add examples benchmarking different contractor settings provided by cotengra #219
Conversation
| params = K.implicit_randn(shape=[nlayers, 2]) | ||
| | ||
| # run only once to trigger the compilation | ||
| K.jit( |
There was a problem hiding this comment.
why using jit and grad here?
There was a problem hiding this comment.
Actually, I find that jitting the function here speeds up the path-finding time. As you pointed out, the gradient calculation was unnecessary, so I removed it.
examples/cotengra_setting_bench.py Outdated
| | ||
| | ||
| def trigger_cotengra_optimization(n, nlayers, d): | ||
| g = nx.random_regular_graph(d, n) |
There was a problem hiding this comment.
please consider several different types of graph (say 1D, 2D lattice or all-to-all connection) for the circuit architectures for a better benchmark?
There was a problem hiding this comment.
I added graph_args to benchmark the performance for 1D, 2D lattice, and all-to-all connected graphs.
examples/cotengra_setting_bench.py Outdated
| minimize="flops", | ||
| parallel=True, | ||
| max_time=30, | ||
| max_repeats=30, |
There was a problem hiding this comment.
max_repeats is too small? maybe try 64 or 128. minimize can also be given as "combo" or "size" apart from "flops"
There was a problem hiding this comment.
I increased the max_repeats and added minimize_args to benchmark various minimize arguments that cotengra supports.
| thanks for the nice contribution. merged with some small tweaks:
|
this part still requires further investigation in the future as the path finding stage seems to be irrelevant from jitting things as they are implemented outside the ML backend. |
2 participants
Resolve #212
(I am contributing to this project as a unitaryHACK participant)