def text_test(words_file=DEFAULT_TEST_DATA_FILE,
epsilon=DEFAULT_EPSILON,
delta=DEFAULT_DELTA,
n=count_min_sketch.DEFAULT_N,
n_factor=5):
words = read_words(words_file)
n *= n_factor
sketches = OrderedDict(
list=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=sketch_tables.ListBackedSketchTable),
array=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=sketch_tables.ArrayBackedSketchTable),
matrix=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=sketch_tables.NumpyMatrixBackedSketchTable),
bitarray=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=sketch_tables.BitarrayBackedSketchTable),
counter=Counter())
benchmark(words, sketches)
most_common_comparison(sketches, n / n_factor)
def test_update_strategies(words_file=DEFAULT_TEST_DATA_FILE,
epsilon=DEFAULT_EPSILON,
delta=DEFAULT_DELTA,
n=count_min_sketch.DEFAULT_N,
table_class=sketch_tables.ArrayBackedSketchTable,
n_factor=5):
words = read_words(words_file)
# Using a factor here to account for different values in the exact top n
n *= n_factor
sketches = OrderedDict(
array=count_min_sketch.TopNCountMinSketch(
delta, epsilon, n=n, table_class=table_class),
array_hash_pair=double_hashing.HashPairCMSketch(
delta, epsilon, n=n, table_class=table_class),
array_conservative=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=table_class,
update_strategy=update_strategy.ConservativeUpdateStrategy),
array_hash_pair_conservative=double_hashing.HashPairCMSketch(
delta,
epsilon,
n=n,
table_class=table_class,
update_strategy=update_strategy.ConservativeUpdateStrategy),
counter=Counter())
# This test does generate the hash-pair ones and the non-hash-pair ones
# with different sizes. However, forcing the non-hash-pair ones into the
# hash-pair scheme only generates worse results
benchmark(words, sketches)
most_common_comparison(sketches, n)
def basic_count_min_test():
count = 100000
numbers = [
random.randint(1, hash_strategy.ARBITRARY_LARGE_PRIME_NUMBER - 1)
for _ in range(count)
]
counts = [random.randint(10, 1000) for _ in range(count)]
cms = count_min_sketch.TopNCountMinSketch(10e-7, 0.005)
for (number, count) in izip(numbers, counts):
cms.insert(number, count)
print(cms.table.depth, cms.table.width, cms.table.depth * cms.table.width)
total_error = 0
percent_error = 0
for i in range(count):
error = float(cms.get(numbers[i]) - counts[i])
total_error += error
percent_error += error / counts[i]
print('Total error:', total_error / sum(counts))
print('Average percent error:', percent_error / count)
print(cms)
def test_lossy_strategy(words_file=DEFAULT_TEST_DATA_FILE,
epsilon=DEFAULT_EPSILON,
delta=DEFAULT_DELTA,
gamma=0.01,
bucket_size=100,
n=count_min_sketch.DEFAULT_N,
table_class=sketch_tables.ArrayBackedSketchTable):
words = read_words(words_file)
# To test lossy strategies, we process the files similarly as before
# And compare by buckets, akin to Goyal and Daumé (2010)
counter = Counter()
sketches = OrderedDict(
no_lossy=count_min_sketch.TopNCountMinSketch(
delta, epsilon, n=n, table_class=table_class),
lossy_no_threshold=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=table_class,
lossy_strategy=lossy_strategy.LossyUpdateStrategy(
gamma, lossy_strategy.no_threshold_func)),
lossy_1_threshold=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=table_class,
lossy_strategy=lossy_strategy.LossyUpdateStrategy(
gamma, lossy_strategy.one_threshold_func)),
lossy_window_size=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=table_class,
lossy_strategy=lossy_strategy.LossyUpdateStrategy(
gamma, lossy_strategy.window_size_threshold_func)),
lossy_sqrt_window=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=table_class,
lossy_strategy=lossy_strategy.LossyUpdateStrategy(
gamma, lossy_strategy.sqrt_window_size_threshold_func)),
counter=counter)
benchmark(words, sketches)
benchmark_by_buckets(bucket_size, sketches)
def test_sketch_hybrid(words_file=DEFAULT_TEST_DATA_FILE,
epsilon=DEFAULT_EPSILON,
delta=DEFAULT_DELTA,
n=count_min_sketch.DEFAULT_N,
table_class=sketch_tables.ArrayBackedSketchTable,
n_factor=5):
words = read_words(words_file)
# Using a factor here to account for different values in the exact top n
n *= n_factor
sketches = OrderedDict(
array_conservative=count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=table_class,
update_strategy=update_strategy.ConservativeUpdateStrategy),
array_hash_pair_conservative=double_hashing.HashPairCMSketch(
delta,
epsilon,
n=n,
table_class=table_class,
update_strategy=update_strategy.ConservativeUpdateStrategy),
hybrid_conservative=counter_sketch_hybrid.SketchCounterHybrid(
count_min_sketch.TopNCountMinSketch(
delta,
epsilon,
n=n,
table_class=table_class,
update_strategy=update_strategy.ConservativeUpdateStrategy)),
hybrid_hash_pair_conservative=counter_sketch_hybrid.SketchCounterHybrid(
double_hashing.HashPairCMSketch(
delta,
epsilon,
n=n,
table_class=table_class,
update_strategy=update_strategy.ConservativeUpdateStrategy)),
counter=Counter())
benchmark(words, sketches)
most_common_comparison(sketches, n)
def test_double_hashing(words_file=DEFAULT_TEST_DATA_FILE,
epsilon=DEFAULT_EPSILON,
delta=DEFAULT_DELTA,
n=count_min_sketch.DEFAULT_N,
table_class=sketch_tables.ArrayBackedSketchTable):
words = read_words(words_file)
sketches = OrderedDict(
array=count_min_sketch.TopNCountMinSketch(
delta, epsilon, table_class=table_class),
array_hash_pair=double_hashing.HashPairCMSketch(
delta=n, epsilon=epsilon, table_class=table_class),
array_hash_pair_multi=double_hashing.MultiHashPairTopNCMSketch(
delta, epsilon, table_class=table_class),
counter=Counter())
benchmark(words, sketches)
most_common_comparison(sketches, n)