450 lines
15 KiB
Python
450 lines
15 KiB
Python
"""
|
|
Tests the accuracy of the opt_einsum paths in addition to unit tests for
|
|
the various path helper functions.
|
|
"""
|
|
|
|
import itertools
|
|
import sys
|
|
|
|
import numpy as np
|
|
import pytest
|
|
|
|
import opt_einsum as oe
|
|
|
|
explicit_path_tests = {
|
|
'GEMM1': ([set('abd'), set('ac'), set('bdc')], set(''), {
|
|
'a': 1,
|
|
'b': 2,
|
|
'c': 3,
|
|
'd': 4
|
|
}),
|
|
'Inner1': ([set('abcd'), set('abc'), set('bc')], set(''), {
|
|
'a': 5,
|
|
'b': 2,
|
|
'c': 3,
|
|
'd': 4
|
|
}),
|
|
}
|
|
|
|
# note that these tests have no unique solution due to the chosen dimensions
|
|
path_edge_tests = [
|
|
['greedy', 'eb,cb,fb->cef', ((0, 2), (0, 1))],
|
|
['branch-all', 'eb,cb,fb->cef', ((0, 2), (0, 1))],
|
|
['branch-2', 'eb,cb,fb->cef', ((0, 2), (0, 1))],
|
|
['optimal', 'eb,cb,fb->cef', ((0, 2), (0, 1))],
|
|
['dp', 'eb,cb,fb->cef', ((1, 2), (0, 1))],
|
|
['greedy', 'dd,fb,be,cdb->cef', ((0, 3), (0, 1), (0, 1))],
|
|
['branch-all', 'dd,fb,be,cdb->cef', ((0, 3), (0, 1), (0, 1))],
|
|
['branch-2', 'dd,fb,be,cdb->cef', ((0, 3), (0, 1), (0, 1))],
|
|
['optimal', 'dd,fb,be,cdb->cef', ((0, 3), (0, 1), (0, 1))],
|
|
['optimal', 'dd,fb,be,cdb->cef', ((0, 3), (0, 1), (0, 1))],
|
|
['dp', 'dd,fb,be,cdb->cef', ((0, 3), (0, 2), (0, 1))],
|
|
['greedy', 'bca,cdb,dbf,afc->', ((1, 2), (0, 2), (0, 1))],
|
|
['branch-all', 'bca,cdb,dbf,afc->', ((1, 2), (0, 2), (0, 1))],
|
|
['branch-2', 'bca,cdb,dbf,afc->', ((1, 2), (0, 2), (0, 1))],
|
|
['optimal', 'bca,cdb,dbf,afc->', ((1, 2), (0, 2), (0, 1))],
|
|
['dp', 'bca,cdb,dbf,afc->', ((1, 2), (1, 2), (0, 1))],
|
|
['greedy', 'dcc,fce,ea,dbf->ab', ((1, 2), (0, 1), (0, 1))],
|
|
['branch-all', 'dcc,fce,ea,dbf->ab', ((1, 2), (0, 2), (0, 1))],
|
|
['branch-2', 'dcc,fce,ea,dbf->ab', ((1, 2), (0, 2), (0, 1))],
|
|
['optimal', 'dcc,fce,ea,dbf->ab', ((1, 2), (0, 2), (0, 1))],
|
|
['dp', 'dcc,fce,ea,dbf->ab', ((1, 2), (0, 2), (0, 1))],
|
|
]
|
|
|
|
|
|
def check_path(test_output, benchmark, bypass=False):
|
|
if not isinstance(test_output, list):
|
|
return False
|
|
|
|
if len(test_output) != len(benchmark):
|
|
return False
|
|
|
|
ret = True
|
|
for pos in range(len(test_output)):
|
|
ret &= isinstance(test_output[pos], tuple)
|
|
ret &= test_output[pos] == benchmark[pos]
|
|
return ret
|
|
|
|
|
|
def assert_contract_order(func, test_data, max_size, benchmark):
|
|
|
|
test_output = func(test_data[0], test_data[1], test_data[2], max_size)
|
|
assert check_path(test_output, benchmark)
|
|
|
|
|
|
def test_size_by_dict():
|
|
|
|
sizes_dict = {}
|
|
for ind, val in zip('abcdez', [2, 5, 9, 11, 13, 0]):
|
|
sizes_dict[ind] = val
|
|
|
|
path_func = oe.helpers.compute_size_by_dict
|
|
|
|
assert 1 == path_func('', sizes_dict)
|
|
assert 2 == path_func('a', sizes_dict)
|
|
assert 5 == path_func('b', sizes_dict)
|
|
|
|
assert 0 == path_func('z', sizes_dict)
|
|
assert 0 == path_func('az', sizes_dict)
|
|
assert 0 == path_func('zbc', sizes_dict)
|
|
|
|
assert 104 == path_func('aaae', sizes_dict)
|
|
assert 12870 == path_func('abcde', sizes_dict)
|
|
|
|
|
|
def test_flop_cost():
|
|
|
|
size_dict = {v: 10 for v in "abcdef"}
|
|
|
|
# Loop over an array
|
|
assert 10 == oe.helpers.flop_count("a", False, 1, size_dict)
|
|
|
|
# Hadamard product (*)
|
|
assert 10 == oe.helpers.flop_count("a", False, 2, size_dict)
|
|
assert 100 == oe.helpers.flop_count("ab", False, 2, size_dict)
|
|
|
|
# Inner product (+, *)
|
|
assert 20 == oe.helpers.flop_count("a", True, 2, size_dict)
|
|
assert 200 == oe.helpers.flop_count("ab", True, 2, size_dict)
|
|
|
|
# Inner product x3 (+, *, *)
|
|
assert 30 == oe.helpers.flop_count("a", True, 3, size_dict)
|
|
|
|
# GEMM
|
|
assert 2000 == oe.helpers.flop_count("abc", True, 2, size_dict)
|
|
|
|
|
|
def test_bad_path_option():
|
|
with pytest.raises(KeyError):
|
|
oe.contract("a,b,c", [1], [2], [3], optimize='optimall')
|
|
|
|
|
|
def test_explicit_path():
|
|
x = oe.contract("a,b,c", [1], [2], [3], optimize=[(1, 2), (0, 1)])
|
|
assert x.item() == 6
|
|
|
|
|
|
def test_path_optimal():
|
|
|
|
test_func = oe.paths.optimal
|
|
|
|
test_data = explicit_path_tests['GEMM1']
|
|
assert_contract_order(test_func, test_data, 5000, [(0, 2), (0, 1)])
|
|
assert_contract_order(test_func, test_data, 0, [(0, 1, 2)])
|
|
|
|
|
|
def test_path_greedy():
|
|
|
|
test_func = oe.paths.greedy
|
|
|
|
test_data = explicit_path_tests['GEMM1']
|
|
assert_contract_order(test_func, test_data, 5000, [(0, 2), (0, 1)])
|
|
assert_contract_order(test_func, test_data, 0, [(0, 1, 2)])
|
|
|
|
|
|
def test_memory_paths():
|
|
|
|
expression = "abc,bdef,fghj,cem,mhk,ljk->adgl"
|
|
|
|
views = oe.helpers.build_views(expression)
|
|
|
|
# Test tiny memory limit
|
|
path_ret = oe.contract_path(expression, *views, optimize="optimal", memory_limit=5)
|
|
assert check_path(path_ret[0], [(0, 1, 2, 3, 4, 5)])
|
|
|
|
path_ret = oe.contract_path(expression, *views, optimize="greedy", memory_limit=5)
|
|
assert check_path(path_ret[0], [(0, 1, 2, 3, 4, 5)])
|
|
|
|
# Check the possibilities, greedy is capped
|
|
path_ret = oe.contract_path(expression, *views, optimize="optimal", memory_limit=-1)
|
|
assert check_path(path_ret[0], [(0, 3), (0, 4), (0, 2), (0, 2), (0, 1)])
|
|
|
|
path_ret = oe.contract_path(expression, *views, optimize="greedy", memory_limit=-1)
|
|
assert check_path(path_ret[0], [(0, 3), (0, 4), (0, 2), (0, 2), (0, 1)])
|
|
|
|
|
|
@pytest.mark.parametrize("alg,expression,order", path_edge_tests)
|
|
def test_path_edge_cases(alg, expression, order):
|
|
views = oe.helpers.build_views(expression)
|
|
|
|
# Test tiny memory limit
|
|
path_ret = oe.contract_path(expression, *views, optimize=alg)
|
|
assert check_path(path_ret[0], order)
|
|
|
|
|
|
def test_optimal_edge_cases():
|
|
|
|
# Edge test5
|
|
expression = 'a,ac,ab,ad,cd,bd,bc->'
|
|
edge_test4 = oe.helpers.build_views(expression, dimension_dict={"a": 20, "b": 20, "c": 20, "d": 20})
|
|
path, path_str = oe.contract_path(expression, *edge_test4, optimize='greedy', memory_limit='max_input')
|
|
assert check_path(path, [(0, 1), (0, 1, 2, 3, 4, 5)])
|
|
|
|
path, path_str = oe.contract_path(expression, *edge_test4, optimize='optimal', memory_limit='max_input')
|
|
assert check_path(path, [(0, 1), (0, 1, 2, 3, 4, 5)])
|
|
|
|
|
|
def test_greedy_edge_cases():
|
|
|
|
expression = "abc,cfd,dbe,efa"
|
|
dim_dict = {k: 20 for k in expression.replace(",", "")}
|
|
tensors = oe.helpers.build_views(expression, dimension_dict=dim_dict)
|
|
|
|
path, path_str = oe.contract_path(expression, *tensors, optimize='greedy', memory_limit='max_input')
|
|
assert check_path(path, [(0, 1, 2, 3)])
|
|
|
|
path, path_str = oe.contract_path(expression, *tensors, optimize='greedy', memory_limit=-1)
|
|
assert check_path(path, [(0, 1), (0, 2), (0, 1)])
|
|
|
|
|
|
def test_dp_edge_cases_dimension_1():
|
|
eq = 'nlp,nlq,pl->n'
|
|
shapes = [(1, 1, 1), (1, 1, 1), (1, 1)]
|
|
info = oe.contract_path(eq, *shapes, shapes=True, optimize='dp')[1]
|
|
assert max(info.scale_list) == 3
|
|
|
|
|
|
def test_dp_edge_cases_all_singlet_indices():
|
|
eq = 'a,bcd,efg->'
|
|
shapes = [(2, ), (2, 2, 2), (2, 2, 2)]
|
|
info = oe.contract_path(eq, *shapes, shapes=True, optimize='dp')[1]
|
|
assert max(info.scale_list) == 3
|
|
|
|
|
|
def test_custom_dp_can_optimize_for_outer_products():
|
|
eq = "a,b,abc->c"
|
|
|
|
da, db, dc = 2, 2, 3
|
|
shapes = [(da, ), (db, ), (da, db, dc)]
|
|
|
|
opt1 = oe.DynamicProgramming(search_outer=False)
|
|
opt2 = oe.DynamicProgramming(search_outer=True)
|
|
|
|
info1 = oe.contract_path(eq, *shapes, shapes=True, optimize=opt1)[1]
|
|
info2 = oe.contract_path(eq, *shapes, shapes=True, optimize=opt2)[1]
|
|
|
|
assert info2.opt_cost < info1.opt_cost
|
|
|
|
|
|
def test_custom_dp_can_optimize_for_size():
|
|
eq, shapes = oe.helpers.rand_equation(10, 4, seed=43)
|
|
|
|
opt1 = oe.DynamicProgramming(minimize='flops')
|
|
opt2 = oe.DynamicProgramming(minimize='size')
|
|
|
|
info1 = oe.contract_path(eq, *shapes, shapes=True, optimize=opt1)[1]
|
|
info2 = oe.contract_path(eq, *shapes, shapes=True, optimize=opt2)[1]
|
|
|
|
assert (info1.opt_cost < info2.opt_cost)
|
|
assert (info1.largest_intermediate > info2.largest_intermediate)
|
|
|
|
|
|
def test_custom_dp_can_set_cost_cap():
|
|
eq, shapes = oe.helpers.rand_equation(5, 3, seed=42)
|
|
opt1 = oe.DynamicProgramming(cost_cap=True)
|
|
opt2 = oe.DynamicProgramming(cost_cap=False)
|
|
opt3 = oe.DynamicProgramming(cost_cap=100)
|
|
info1 = oe.contract_path(eq, *shapes, shapes=True, optimize=opt1)[1]
|
|
info2 = oe.contract_path(eq, *shapes, shapes=True, optimize=opt2)[1]
|
|
info3 = oe.contract_path(eq, *shapes, shapes=True, optimize=opt3)[1]
|
|
assert info1.opt_cost == info2.opt_cost == info3.opt_cost
|
|
|
|
|
|
@pytest.mark.parametrize("optimize", ['greedy', 'branch-2', 'branch-all', 'optimal', 'dp'])
|
|
def test_can_optimize_outer_products(optimize):
|
|
a, b, c = [np.random.randn(10, 10) for _ in range(3)]
|
|
d = np.random.randn(10, 2)
|
|
assert oe.contract_path("ab,cd,ef,fg", a, b, c, d, optimize=optimize)[0] == [(2, 3), (0, 2), (0, 1)]
|
|
|
|
|
|
@pytest.mark.parametrize('num_symbols', [2, 3, 26, 26 + 26, 256 - 140, 300])
|
|
def test_large_path(num_symbols):
|
|
symbols = ''.join(oe.get_symbol(i) for i in range(num_symbols))
|
|
dimension_dict = dict(zip(symbols, itertools.cycle([2, 3, 4])))
|
|
expression = ','.join(symbols[t:t + 2] for t in range(num_symbols - 1))
|
|
tensors = oe.helpers.build_views(expression, dimension_dict=dimension_dict)
|
|
|
|
# Check that path construction does not crash
|
|
oe.contract_path(expression, *tensors, optimize='greedy')
|
|
|
|
|
|
def test_custom_random_greedy():
|
|
eq, shapes = oe.helpers.rand_equation(10, 4, seed=42)
|
|
views = list(map(np.ones, shapes))
|
|
|
|
with pytest.raises(ValueError):
|
|
oe.RandomGreedy(minimize='something')
|
|
|
|
optimizer = oe.RandomGreedy(max_repeats=10, minimize='flops')
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
assert len(optimizer.costs) == 10
|
|
assert len(optimizer.sizes) == 10
|
|
|
|
assert path == optimizer.path
|
|
assert optimizer.best['flops'] == min(optimizer.costs)
|
|
assert path_info.largest_intermediate == optimizer.best['size']
|
|
assert path_info.opt_cost == optimizer.best['flops']
|
|
|
|
# check can change settings and run again
|
|
optimizer.temperature = 0.0
|
|
optimizer.max_repeats = 6
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
assert len(optimizer.costs) == 16
|
|
assert len(optimizer.sizes) == 16
|
|
|
|
assert path == optimizer.path
|
|
assert optimizer.best['size'] == min(optimizer.sizes)
|
|
assert path_info.largest_intermediate == optimizer.best['size']
|
|
assert path_info.opt_cost == optimizer.best['flops']
|
|
|
|
# check error if we try and reuse the optimizer on a different expression
|
|
eq, shapes = oe.helpers.rand_equation(10, 4, seed=41)
|
|
views = list(map(np.ones, shapes))
|
|
with pytest.raises(ValueError):
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
|
|
def test_custom_branchbound():
|
|
eq, shapes = oe.helpers.rand_equation(8, 4, seed=42)
|
|
views = list(map(np.ones, shapes))
|
|
optimizer = oe.BranchBound(nbranch=2, cutoff_flops_factor=10, minimize='size')
|
|
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
assert path == optimizer.path
|
|
assert path_info.largest_intermediate == optimizer.best['size']
|
|
assert path_info.opt_cost == optimizer.best['flops']
|
|
|
|
# tweak settings and run again
|
|
optimizer.nbranch = 3
|
|
optimizer.cutoff_flops_factor = 4
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
assert path == optimizer.path
|
|
assert path_info.largest_intermediate == optimizer.best['size']
|
|
assert path_info.opt_cost == optimizer.best['flops']
|
|
|
|
# check error if we try and reuse the optimizer on a different expression
|
|
eq, shapes = oe.helpers.rand_equation(8, 4, seed=41)
|
|
views = list(map(np.ones, shapes))
|
|
with pytest.raises(ValueError):
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
|
|
@pytest.mark.skipif(sys.version_info < (3, 2), reason="requires python3.2 or higher")
|
|
def test_parallel_random_greedy():
|
|
from concurrent.futures import ProcessPoolExecutor
|
|
pool = ProcessPoolExecutor(2)
|
|
|
|
eq, shapes = oe.helpers.rand_equation(10, 4, seed=42)
|
|
views = list(map(np.ones, shapes))
|
|
|
|
optimizer = oe.RandomGreedy(max_repeats=10, parallel=pool)
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
assert len(optimizer.costs) == 10
|
|
assert len(optimizer.sizes) == 10
|
|
|
|
assert path == optimizer.path
|
|
assert optimizer.parallel is pool
|
|
assert optimizer._executor is pool
|
|
assert optimizer.best['flops'] == min(optimizer.costs)
|
|
assert path_info.largest_intermediate == optimizer.best['size']
|
|
assert path_info.opt_cost == optimizer.best['flops']
|
|
|
|
# now switch to max time algorithm
|
|
optimizer.max_repeats = int(1e6)
|
|
optimizer.max_time = 0.2
|
|
optimizer.parallel = 2
|
|
|
|
path, path_info = oe.contract_path(eq, *views, optimize=optimizer)
|
|
|
|
assert len(optimizer.costs) > 10
|
|
assert len(optimizer.sizes) > 10
|
|
|
|
assert path == optimizer.path
|
|
assert optimizer.best['flops'] == min(optimizer.costs)
|
|
assert path_info.largest_intermediate == optimizer.best['size']
|
|
assert path_info.opt_cost == optimizer.best['flops']
|
|
|
|
optimizer.parallel = True
|
|
assert optimizer._executor is not None
|
|
assert optimizer._executor is not pool
|
|
|
|
are_done = [f.running() or f.done() for f in optimizer._futures]
|
|
assert all(are_done)
|
|
|
|
|
|
def test_custom_path_optimizer():
|
|
class NaiveOptimizer(oe.paths.PathOptimizer):
|
|
def __call__(self, inputs, output, size_dict, memory_limit=None):
|
|
self.was_used = True
|
|
return [(0, 1)] * (len(inputs) - 1)
|
|
|
|
eq, shapes = oe.helpers.rand_equation(5, 3, seed=42, d_max=3)
|
|
views = list(map(np.ones, shapes))
|
|
|
|
exp = oe.contract(eq, *views, optimize=False)
|
|
|
|
optimizer = NaiveOptimizer()
|
|
out = oe.contract(eq, *views, optimize=optimizer)
|
|
assert exp == out
|
|
assert optimizer.was_used
|
|
|
|
|
|
def test_custom_random_optimizer():
|
|
class NaiveRandomOptimizer(oe.path_random.RandomOptimizer):
|
|
@staticmethod
|
|
def random_path(r, n, inputs, output, size_dict):
|
|
"""Picks a completely random contraction order.
|
|
"""
|
|
np.random.seed(r)
|
|
ssa_path = []
|
|
remaining = set(range(n))
|
|
while len(remaining) > 1:
|
|
i, j = np.random.choice(list(remaining), size=2, replace=False)
|
|
remaining.add(n + len(ssa_path))
|
|
remaining.remove(i)
|
|
remaining.remove(j)
|
|
ssa_path.append((i, j))
|
|
cost, size = oe.path_random.ssa_path_compute_cost(ssa_path, inputs, output, size_dict)
|
|
return ssa_path, cost, size
|
|
|
|
def setup(self, inputs, output, size_dict):
|
|
self.was_used = True
|
|
n = len(inputs)
|
|
trial_fn = self.random_path
|
|
trial_args = (n, inputs, output, size_dict)
|
|
return trial_fn, trial_args
|
|
|
|
eq, shapes = oe.helpers.rand_equation(5, 3, seed=42, d_max=3)
|
|
views = list(map(np.ones, shapes))
|
|
|
|
exp = oe.contract(eq, *views, optimize=False)
|
|
|
|
optimizer = NaiveRandomOptimizer(max_repeats=16)
|
|
out = oe.contract(eq, *views, optimize=optimizer)
|
|
assert exp == out
|
|
assert optimizer.was_used
|
|
|
|
assert len(optimizer.costs) == 16
|
|
|
|
|
|
def test_optimizer_registration():
|
|
def custom_optimizer(inputs, output, size_dict, memory_limit):
|
|
return [(0, 1)] * (len(inputs) - 1)
|
|
|
|
with pytest.raises(KeyError):
|
|
oe.paths.register_path_fn('optimal', custom_optimizer)
|
|
|
|
oe.paths.register_path_fn('custom', custom_optimizer)
|
|
assert 'custom' in oe.paths._PATH_OPTIONS
|
|
|
|
eq = 'ab,bc,cd'
|
|
shapes = [(2, 3), (3, 4), (4, 5)]
|
|
path, path_info = oe.contract_path(eq, *shapes, shapes=True, optimize='custom')
|
|
assert path == [(0, 1), (0, 1)]
|
|
del oe.paths._PATH_OPTIONS['custom']
|