def main():
print take(10, iter_abundant_numbers())
# print len(list(takewhile(lambda n: n < MAX, iter_abundant_numbers())))
table = {
sum(x)
for x in combinations_with_replacement(
takewhile(lambda x: x <= MAX / 2, iter_abundant_numbers()), 2)
}
print "Got table"
print sum(x for x in range(MAX) if x not in table)
def normalize_to_single_token_edits(edit):
def positions_it(region):
return itertools.chain(
(m2format.Region(pos, pos + 1)
for pos in range(region.beg, region.end)),
itertools.repeat(m2format.Region(region.end, region.end)),
)
def tokens_it(tokens):
return itertools.chain(
((token, ) for token in tokens),
itertools.repeat(tuple()),
)
region = edit.region
tokens = edit.tokens
if region.beg == region.end:
yield edit
return
n = max(region.end - region.beg, len(tokens))
for st_region, st_tokens in utils.take(
n, zip(positions_it(region), tokens_it(tokens))):
assert region.beg <= st_region.beg and st_region.end <= region.end
yield m2format.Edit(st_region, st_tokens, edit.type)
def first_child(elem, child_name):
'''
Get the first child element with a specific element name.
'''
first = take(1, elem.iter(child_name))
if (not first) or (len(first) == 0):
return None
return first[0]
def first_child(elem, child_name):
'''
Get the first child element with a specific element name.
'''
first = take(1, elem.iter(child_name))
if (not first) or (len(first) == 0):
return None
return first[0]
def forall_left(premise, premises, goal, tg):
ts = take(randnat(), guessterm(tg, premises, goal))
if KEEP_USED_PREMISES:
return [(ByForallLeft, [(premises | {premise.instantiate(t)}, goal)])
for t in ts]
else:
return [(ByForallLeft,
[(premises - {premise} | {premise.instantiate(t)}, goal)])
for t in ts]
def test3():
from chords import generateChords
allNotes = utils.take(100, generateChords(noteRange= (3,3)))
for notes in allNotes:
print "="*30, notes
chains = findBestChains(notes)
for chain in chains:
for c in chain:
print c
print
def test3():
from chords import generateChords
allNotes = utils.take(100, generateChords(noteRange=(3, 4)))
for notes in allNotes:
print "=" * 20, notes
chains = findBestChains(notes)
for chain in chains:
for chord in chain:
#print chord
print formatChord(chord)
#print formatChain(chain)
print
def test_generate():
from rationals import Rational
chords = utils.take(1000, generateChords((3, 3)))
chords = [(c, getHarmonicDistance(c)) for c in chords]
chords.sort(key=lambda (c, d): d)
for (i, (c, d)) in enumerate(chords[:50]):
chord = map(Rational, c)
for j in range(len(chord))[::-1]:
chord[j] /= chord[0] # normalize
# print
print "%03d." % (i + 1), c,
for r in chord:
print r.getFraction(0),
#print r.getSonant(),
print d
def frames_generator_rnn(dataset_dir,
split_key,
batch_size,
additional_data=None):
"""
Internally this method uses _sequential_infinite_iterator_rnn to iterate on the validation set.
If the batch_size is not divisible by the number of videos, the last videos will not be
returned, but they will be returned at the next iteration over the valid set, so we might
end with statistics that will be slightly different if computed multiple times on the
validation set returned by this generator.
"""
base_dir = join(dataset_dir, split_key)
dataset_structure = get_dataset_split_structure(base_dir)
all_classes = dataset_structure.keys()
d_iterator = _get_iterator(split_key)(base_dir, dataset_structure)
tot_videos = next(count_num_videos(dataset_dir, split_key))
num_calls = tot_videos // batch_size
yield num_calls
while True:
for _ in range(num_calls):
data = take(d_iterator, batch_size)
class_names, video_names, videos, labels = map(
np.array, zip(*data))
labels = to_categorical(labels, len(all_classes))
if additional_data is not None:
additional_data_batch = np.array([
additional_data[cl][v]
for cl, v in zip(class_names, video_names)
])
yield [additional_data_batch, videos], [labels]
else:
yield videos, labels
def read_batch(idx, paths, batch_size=32, img_size=(48, 48)):
'''
Read batch tuple (X, y) on index idx from paths. X are grayscale images, y are target RGB images
Args:
idx (number): Index of batch to take
paths (list): List of paths to read
batch_size (number): Batch size - number of tuples to read
img_size (tuple): Images size after squarification, in (W, H) format
'''
X, y = [], []
for img_rgb in take(batch_size,
paths[idx * batch_size:],
extract_fn=partial(read_image, size=img_size)):
img_gray = rgb2gray(img_rgb).reshape(img_size + (1, ))
X.append(np.repeat(img_gray, 3, axis=-1))
y.append(img_rgb)
return np.array(X), np.array(y)
def frames_generator_rnn(dataset_dir, split_key, batch_size):
"""
Internally this method uses _sequential_infinite_iterator_rnn to iterate on the validation set.
If the batch_size is not divisible by the number of videos, the last videos will not be
returned, but they will be returned at the next iteration over the valid set, so we might
end with statistics that will be slightly different if computed multiple times on the
validation set returned by this generator.
"""
base_dir = join(dataset_dir, split_key)
dataset_structure = get_dataset_split_structure(base_dir)
all_classes = dataset_structure.keys()
d_iterator = _get_iterator(split_key)(base_dir, dataset_structure)
tot_videos = next(count_num_videos(dataset_dir, split_key))
num_calls = tot_videos // batch_size
yield num_calls
while True:
for _ in range(num_calls):
data = take(d_iterator, batch_size)
videos, labels = map(np.array, zip(*data))
# if batch_size == 1:
# videos = np.array([videos])
# labels = np.array([labels])
# else:
# videos = np.array(videos)
# labels = np.array(labels)
videos = preprocess_images_tf(videos)
labels = to_categorical(labels, len(all_classes))
yield videos, labels
def pos_from_files(files, max_sents=INF, rem_id=True, shuffle=False, shuffle_seed=448):
sents = (sent for f in files for sent in pos_from_file(f, rem_id=rem_id))
return take(iter(shuffle_seq(sents, shuffle_seed)) if shuffle else sents, max_sents)
def max_happiness(happiness, attendees): h = lambda a, b, c: happiness[b][a] + happiness[b][c] seating = lambda xs: take(len(attendees), window(cycle(xs), 3)) total = lambda xs: sum(starmap(h, seating(xs))) optimal = max(permutations(attendees), key=total) return optimal, total(optimal)
def max_happiness(happiness, attendees):
h = lambda a, b, c: happiness[b][a] + happiness[b][c]
seating = lambda xs: take(len(attendees), window(cycle(xs), 3))
total = lambda xs: sum(starmap(h, seating(xs)))
optimal = max(permutations(attendees), key=total)
return optimal, total(optimal)
def main():
print take(10, iter_abundant_numbers())
# print len(list(takewhile(lambda n: n < MAX, iter_abundant_numbers())))
table = {sum(x) for x in combinations_with_replacement(takewhile(lambda x: x <= MAX / 2, iter_abundant_numbers()), 2)}
print "Got table"
print sum(x for x in range(MAX) if x not in table)
def exists_right(premises, goal, tg):
ts = take(randnat(), guessterm(tg, premises, goal))
return [(ByExistsRight, [(premises, goal.instantiate(t))]) for t in ts]
def random_positions(n):
return take(shuffle(positions()), n)
for _ in it.repeat(None):
neg_sampler_jit_pad_arr_(a)
yield a[pad:].copy()
# for i in a[pad:]:
# yield i
# ### Check distributions
#
# Just as a sanity check that the different implementations do the same thing, I randomly generate words according to how frequently they occur in the text with each of the samplers and scatter-plot them against the actual word frequency to check that they mostly lie on $y=x$.
# In[ ]:
from sklearn.preprocessing import LabelEncoder
from nltk.corpus import brown
some_text = take(brown.words(), int(1e6))
# In[ ]:
le = LabelEncoder()
smtok = le.fit_transform(some_text)
# In[ ]:
class NegSampler:
"""Container for the sampler generator functions.
This keeps track of the number of samples $K$ and the padding."""
def __init__(self, sampler, toks, K=None, ret_type=None, pad=None,
nxt=True, **kw):
@jit
def update(rng, i, opt_state, batch):
params = get_params(opt_state)
grads = grad(loss)(params, batch)
return opt_update(i, grads, opt_state)
if __name__ == '__main__':
key = random.PRNGKey(0)
# Create dataset
X_full = utils.get_datasets(dataset)
kfold = model_selection.KFold(pieces, shuffle=True, random_state=0)
for fold_iter, (idx_train, idx_test) in enumerate(
utils.take(pieces_to_run, kfold.split(X_full))):
X, X_test = X_full[idx_train], X_full[idx_test]
scaler = preprocessing.StandardScaler()
X = scaler.fit_transform(X)
X_test = scaler.transform(X_test)
delta = 1. / (X.shape[0]**1.1)
print('X: {}'.format(X.shape))
print('X test: {}'.format(X_test.shape))
print('Delta: {}'.format(delta))
# Create flow
modules = flow_utils.get_modules(flow, num_blocks, normalization,
num_hidden)
