def test_dist(self):
f = 2
i = AnnoyIndex(f, 'euclidean')
i.add_item(0, [0, 1])
i.add_item(1, [1, 1])
self.assertAlmostEqual(i.get_distance(0, 1), 1.0)
def test_metric_kwarg(self):
# Issue 211
i = AnnoyIndex(2, metric='euclidean')
i.add_item(0, [1, 0])
i.add_item(1, [9, 0])
self.assertAlmostEqual(i.get_distance(0, 1), 8)
self.assertEqual(i.f, 2)
class ImageSearchAnnoyCombo:
'''
load an Annoy index for approximate nearest neighbor computation
Annoy's angular distance uses dist(u,v) = 2(1-cos(u,v))
'''
def __init__(self,h5fname = 'X_ILSVRC2015.hdf5',annf='ILSVRC2015.ann',imageListPath = '/home/scratch/benediktb/RegionOfInterest/ILSVRC2015_filelist.txt',dset = 'fc6fc7'):
#load h5 data
h5f = h5py.File(h5fname,'r')
self.X = h5f[dset]
#load filenames
with open(imageListPath,'r') as f:
self.line_to_file = {i:line.rstrip() for i,line in enumerate(f)}
self.A = AnnoyIndex(self.X.shape[1],'angular')
self.A.load(annf)
def run_query_approx(self,query,n=100,accuracy_factor = 5):
nearest,scores = self.A.get_nns_by_vector(query, n, search_k=n*int(accuracy_factor)*128, include_distances=True)
return zip((self.line_to_file[i] for i in nearest),scores)
def run_query_exact(self,query,n=1000,nsmall=100):
#retrieve approximate nearest neighbors using Annoy, then do exact ranking by loading from h5 into memory
#use Annoy
if n < nsmall:
n = nsmall
indexes = self.A.get_nns_by_vector(query, n, search_k=-1, include_distances=False)
indexes_sorted = sorted(indexes)
#use scipy cdist (or normalize first and do dot product for faster computation)
#getting X by index from disc is very slow.
distance = (cdist(self.X[indexes_sorted], query.reshape((1,query.shape[0])), 'cosine'))[:,0]
ind = np.argpartition(distance, nsmall)[:nsmall]#partial sort, indices for top n,
s_ind = np.argsort(distance[ind])#sort
nearest = ind[s_ind]
scoresorted = distance[ind][s_ind]
return zip((self.line_to_file[indexes_sorted[i]] for i in nearest),scoresorted)
def test_tuple(self, n_points=1000, n_trees=10):
f = 10
i = AnnoyIndex(f, 'euclidean')
for j in xrange(n_points):
i.add_item(j, (random.gauss(0, 1) for x in xrange(f)))
i.build(n_trees)
def do(indextype):
a = AnnoyIndex(8, indextype[0])
a.load('points.%s.annoy' % indextype)
with open('points.%s.ann.txt' % indextype, 'w') as out:
for q_index in [1443, 1240, 818, 1725, 1290, 2031, 1117, 1211, 1902, 603]:
nns = a.get_nns_by_item(q_index, 10)
print >> out, '%s\t%s' % (q_index, ','.join([str(n) for n in nns]))
def test_dist(self):
f = 2
i = AnnoyIndex(f)
i.add_item(0, [0, 1])
i.add_item(1, [1, 1])
self.assertAlmostEqual(i.get_distance(0, 1), (2 * (1.0 - 2 ** -0.5))**0.5)
def _get_index(self, f, distance):
input = 'test/glove.twitter.27B.%dd.txt.gz' % f
output = 'test/glove.%d.%s.annoy' % (f, distance)
if not os.path.exists(output):
if not os.path.exists(input):
# Download GloVe pretrained vectors: http://nlp.stanford.edu/projects/glove/
# Hosting them on my own S3 bucket since the original files changed format
url = 'https://s3-us-west-1.amazonaws.com/annoy-vectors/glove.twitter.27B.%dd.txt.gz' % f
print('downloading', url, '->', input)
urlretrieve(url, input)
print('building index', distance, f)
annoy = AnnoyIndex(f, 12, "test_db", 10, 1000, 3048576000, 0)
v_v = []
items = []
for i, line in enumerate(gzip.open(input, 'rb')):
v = [float(x) for x in line.strip().split()[1:]]
v_v.append(v)
items.append(i)
if (i+1) % 10000 == 0:
print (i+1)
annoy.add_item_batch(items, v_v)
v_v = []
items = []
if v_v:
annoy.add_item_batch(items, v_v)
return annoy
def ANN(searchSpace):
dimension = searchSpace[0].shape[0]
t = AnnoyIndex(dimension, metric='euclidean')
for i in range(len(searchSpace)):
t.add_item(i, searchSpace[i])
t.build(10)
return t
def test_dist_degen(self):
f = 2
i = AnnoyIndex(f)
i.add_item(0, [1, 0])
i.add_item(1, [0, 0])
self.assertAlmostEqual(i.get_distance(0, 1), 2.0**0.5)
def retrieve(self):
print 'Loading necessary files..'
u = AnnoyIndex(self.dim, metric='angular')
u.load(index_file)
print 'ANN Retrieval..'
for n_neighbors in knns:
print 'Number of neighbors: ' + str(n_neighbors)
for mult in self.multipliers:
print 'Multiplier: ' + str(mult)
search_k = self.n_trees * n_neighbors * mult
filename = '.'.join((self.test_file.split('/')[-1].split('.')[:-1]))
with open(self.test_file, 'r') as data_file:
data = json.load(data_file)
qArray = []
for i in range(len(data["questions"])):
question_body = data["questions"][i]["body"]
question_id = data["questions"][i]["id"]
qcentroid = np.transpose(np.array(get_centroid_idf(question_body, self.emb, self.idf, self.stopwords, self.dim)))
anns = u.get_nns_by_vector(qcentroid, n_neighbors, search_k)
doc_anns = []
for n in anns:
doc_anns.append(self.idmap[n])
q = Question(question_body, question_id, doc_anns)
qArray.append(q)
directory = "system_results/"
if not os.path.exists(directory):
os.makedirs(directory)
with open(str(directory)+"/"+"CentIDF_annoy_"+str(n_trees)+"_"+str(n_neighbors)+"_"+str(mult)+".json", "w+") as outfile:
outfile.write(json.dumps({"questions":[ob.__dict__ for ob in qArray]}, indent=2))
def test_no_items(self):
idx = AnnoyIndex(100)
idx.build(n_trees=10)
idx.save('foo.idx')
idx = AnnoyIndex(100)
idx.load('foo.idx')
self.assertEquals(idx.get_n_items(), 0)
self.assertEquals(idx.get_nns_by_vector(vector=numpy.random.randn(100), n=50, include_distances=False), [])
def test_save_without_build(self):
# Issue #61
i = AnnoyIndex(10)
i.add_item(1000, [random.gauss(0, 1) for z in xrange(10)])
i.save('x.tree')
j = AnnoyIndex(10)
j.load('x.tree')
j.build(10)
def fit_annoy(data, n_trees=-1):
logger.info('Fitting Annoy Matcher...')
from annoy import AnnoyIndex
matcher = AnnoyIndex(data.shape[1], metric='euclidean')
for i, d in enumerate(data):
matcher.add_item(i, d)
matcher.build(n_trees)
return matcher
def test_get_item_vector(self):
f = 10
i = AnnoyIndex(f, 'euclidean')
i.add_item(0, [random.gauss(0, 1) for x in xrange(f)])
for j in xrange(100):
print(j, '...')
for k in xrange(1000 * 1000):
i.get_item_vector(0)
def build_index(counts,label_to_id,dimension):
index = AnnoyIndex(dimension,metric='angular')
for label,cnt_list in counts.items():
id = label_to_id[label]
index.add_item(id,cnt_list)
index.build(100)
return index
def test_wrong_length(self, n_points=1000, n_trees=10):
f = 10
i = AnnoyIndex(f, 'euclidean')
i.add_item(0, [random.gauss(0, 1) for x in xrange(f)])
self.assertRaises(IndexError, i.add_item, 1, [random.gauss(0, 1) for x in xrange(f+1000)])
self.assertRaises(IndexError, i.add_item, 2, [])
i.build(n_trees)
def __init__(self, fn_word, model_name, model_path):
self.model = QueryModel(fn_word, model_name, model_path)
self.queries = []
self.titles = []
self.query_index = 0
self.title_index = 0
self.query_ann = AnnoyIndex(self.model.dim, metric='euclidean')
self.title_ann = AnnoyIndex(self.model.dim, metric='euclidean')
def create_index_tree(clusters):
features = clusters.shape[1]
tree = AnnoyIndex(features, metric='euclidean')
for i, v in enumerate(clusters):
tree.add_item(i, v.tolist())
tree.build(features*2)
return tree
def _build_from_model(self, vectors, labels, num_features):
index = AnnoyIndex(num_features)
for vector_num, vector in enumerate(vectors):
index.add_item(vector_num, vector)
index.build(self.num_trees)
self.index = index
self.labels = labels
class FeatureNN:
tree = None
def __init__(self, features, tree_file):
self.tree = AnnoyIndex(features, metric='euclidean')
self.tree.load(str(tree_file))
def nn(self, x):
return self.tree.get_nns_by_vector(x.tolist(), 1)[0]
def test_numpy(self, n_points=1000, n_trees=10):
f = 10
i = AnnoyIndex(f, 'euclidean')
for j in xrange(n_points):
a = numpy.random.normal(size=f)
a = a.astype(random.choice([numpy.float64, numpy.float32, numpy.uint8, numpy.int16]))
i.add_item(j, a)
i.build(n_trees)
def test_dist_degen(self):
os.system("rm -rf test_db")
os.system("mkdir test_db")
f = 2
i = AnnoyIndex(f, 2, "test_db", 64, 1000, 3048576000, 0)
i.add_item(0, [1, 0])
i.add_item(1, [0, 0])
self.assertAlmostEqual(i.get_distance(0, 1), 2.0)
def _get_index(self, dataset):
url = 'http://vectors.erikbern.com/%s.hdf5' % dataset
vectors_fn = os.path.join('test', dataset + '.hdf5')
index_fn = os.path.join('test', dataset + '.annoy')
if not os.path.exists(vectors_fn):
print('downloading', url, '->', vectors_fn)
urlretrieve(url, vectors_fn)
dataset_f = h5py.File(vectors_fn)
distance = dataset_f.attrs['distance']
f = dataset_f['train'].shape[1]
annoy = AnnoyIndex(f, distance)
if not os.path.exists(index_fn):
print('adding items', distance, f)
for i, v in enumerate(dataset_f['train']):
annoy.add_item(i, v)
print('building index')
annoy.build(10)
annoy.save(index_fn)
else:
annoy.load(index_fn)
return annoy, dataset_f
def test_write_failed(self):
f = 40
# Build the initial index
t = AnnoyIndex(f)
for i in range(1000):
v = [random.gauss(0, 1) for z in range(f)]
t.add_item(i, v)
t.build(10)
if sys.platform == "linux" or sys.platform == "linux2":
# linux
try:
t.save("/dev/full")
self.fail("didn't get expected exception")
except Exception as e:
self.assertTrue(str(e).find("No space left on device") > 0)
elif sys.platform == "darwin":
volume = "FULLDISK"
device = os.popen('hdiutil attach -nomount ram://64').read()
os.popen('diskutil erasevolume MS-DOS %s %s' % (volume, device))
os.popen('touch "/Volumes/%s/full"' % volume)
try:
t.save('/Volumes/%s/annoy.tree' % volume)
self.fail("didn't get expected exception")
except Exception as e:
self.assertTrue(str(e).find("No space left on device") > 0)
finally:
os.popen("hdiutil detach %s" % device)
def precision(f=40, n=1000000):
t = AnnoyIndex(f)
for i in xrange(n):
v = []
for z in xrange(f):
v.append(random.gauss(0, 1))
t.add_item(i, v)
t.build(2 * f)
t.save('test.tree')
limits = [10, 100, 1000, 10000]
k = 10
prec_sum = {}
prec_n = 1000
time_sum = {}
for i in xrange(prec_n):
j = random.randrange(0, n)
print 'finding nbs for', j
closest = set(t.get_nns_by_item(j, n)[:k])
for limit in limits:
t0 = time.time()
toplist = t.get_nns_by_item(j, limit)
T = time.time() - t0
found = len(closest.intersection(toplist))
hitrate = 1.0 * found / k
prec_sum[limit] = prec_sum.get(limit, 0.0) + hitrate
time_sum[limit] = time_sum.get(limit, 0.0) + T
for limit in limits:
print 'limit: %-9d precision: %6.2f%% avg time: %.6fs' % (limit, 100.0 * prec_sum[limit] / (i + 1), time_sum[limit] / (i + 1))
def build_annoy_index(corpus, dimension, winlen, winstep):
print "Adding to Annoy index"
index = AnnoyIndex(dimension, "euclidean")
mfcc_list = []
i = 0
for filename, frames in corpus:
# print filename, frames.shape
for index_in_file, mfcc in enumerate(frames):
mfcc_list.append((filename, index_in_file))
index.add_item(i, mfcc.tolist())
assert mfcc_list[i] == (filename, index_in_file)
i += 1
opts = {"samplerate": desired_samplerate,
"winlen": winlen,
"winstep": winstep,
"numcep": 13,
"nfilt": 26,
"nfft": 512,
"ntrees": ANN_NTREES
}
cache_filename = "annoy_index_" + hashlib.md5(str([filename for filename, frames in corpus])).hexdigest() + "." + "_".join("%s=%s" % (k, v) for k, v in sorted(opts.items())) + ".tree"
if not os.path.exists(cache_filename):
print "Building Annoy index with %d trees" % ANN_NTREES
# index.build(-1)
index.build(ANN_NTREES)
index.save(cache_filename)
print "\tWrote cache to %s" % cache_filename
else:
print "\tReading cache from %s" % cache_filename
index.load(cache_filename)
return index, mfcc_list
def test_range_errors(self, n_points=1000, n_trees=10):
f = 10
i = AnnoyIndex(f, 'euclidean')
for j in xrange(n_points):
i.add_item(j, [random.gauss(0, 1) for x in xrange(f)])
self.assertRaises(IndexError, i.add_item, -1, [random.gauss(0, 1) for x in xrange(f)])
i.build(n_trees)
for bad_index in [-1000, -1, n_points, n_points + 1000]:
self.assertRaises(IndexError, i.get_distance, 0, bad_index)
self.assertRaises(IndexError, i.get_nns_by_item, bad_index, 1)
self.assertRaises(IndexError, i.get_item_vector, bad_index)
def test_threads(self):
n, f = 10000, 10
i = AnnoyIndex(f, 'euclidean')
for j in xrange(n):
i.add_item(j, numpy.random.normal(size=f))
i.build(10)
pool = multiprocessing.pool.ThreadPool()
def query_f(j):
i.get_nns_by_item(1, 1000)
pool.map(query_f, range(n))
def test_dist(self):
os.system("rm -rf test_db")
os.system("mkdir test_db")
f = 2
i = AnnoyIndex(f, 2, "test_db", 64, 1000, 3048576000, 0)
# i.verbose(True)
i.add_item(0, [0, 1])
i.add_item(1, [1, 1])
self.assertAlmostEqual(i.get_distance(0, 1), 2 * (1.0 - 2 ** -0.5))
def test_get_nns_by_item_batch(self):
print "test_get_nns_by_item_batch "
os.system("rm -rf test_db")
os.system("mkdir test_db")
f = 3
i = AnnoyIndex(f, 3, "test_db", 10, 1000, 3048576000, 0)
i.add_item_batch([0,1,2], [[2, 1, 0], [1, 2, 0], [0, 0, 1]])
self.assertEqual(i.get_nns_by_item(0, 3), [0, 1, 2])
self.assertEqual(i.get_nns_by_item(1, 3), [1, 0, 2])
self.assertTrue(i.get_nns_by_item(2, 3) in [[2, 0, 1], [2, 1, 0]]) # could be either
# coding: utf-8
from annoy import AnnoyIndex
import json
import random
import redis
redis_client = redis.StrictRedis(host='localhost', port=6379, db=0)
illust_vector_list = []
dim = 512
t = AnnoyIndex(dim) # Length of item vector that will be indexed
path = '../keras/pixiv-ranking-features.txt'
with open(path, 'r') as f:
i = 0
line = f.readline()
while line:
js = json.loads(line)
print(js['illust_id'])
illust_vector_list.append(js)
if i > 1e3:
break
i = i + 1
line = f.readline()
for i, illust_vector in enumerate(illust_vector_list):
illust_id = illust_vector['illust_id']
line2illustId = f"line2illustId_{i}"
illustId2line = f"illustId2line_{illust_id}"
print(f"{line2illustId} -> {illustId2line}")
def main():
# load face features
u = AnnoyIndex(512, metric="euclidean")
u.load(config.FACE_FEATURES)
video = args.video_path
f = open(args.output_path, "w")
fvs = FileVideoStream(video).start()
time.sleep(1.0)
fps = FPS().start()
count = 0
frame_count = 0
while fvs.more():
img = fvs.read()
f.write("frame_{}".format(frame_count))
img = cv2.resize(img,
None,
fx=config.RESIZE_IMAGE,
fy=config.RESIZE_IMAGE)
bboxlist = detector.detect(img)
for b in bboxlist:
x1, y1, x2, y2, s = b
if (s > config.DETECTION_THRESHOLD):
x1 = int(x1)
x2 = int(x2)
y1 = int(y1)
y2 = int(y2)
width = x2 - x1
height = y2 - y1
if width >= config.MIN_SIZE and height >= config.MIN_SIZE:
ret = mtcnn_detector.detect_face(img[y1:y2, x1:x2],
det_type=1)
if ret is None:
continue
face_image = "frame_" + str(frame_count) + "_face_" + str(
count) + ".jpg"
count += 1
f.write(",object_{},position,{},{},{},{}".format(
count, x1, x2, y1, y2))
bbox, landmarks = ret
if landmarks is None:
continue
pointx = landmarks[0][:5]
pointy = landmarks[0][5:]
pointx_img_space = map(lambda x: x + x1, pointx)
pointy_img_space = map(lambda y: y + y1, pointy)
landmarks_img_space = list(pointx_img_space) + list(
pointy_img_space)
bbox_process = np.array([x1, y1, x2, y2])
landmarks_process = np.array(landmarks_img_space).reshape(
(2, 5)).T
nimg = preprocess(img,
bbox_process,
landmarks_process,
image_size="112,112")
e = embedding.get_feature(nimg)
match = u.get_nns_by_vector(e,
config.TOP_ACC_NUM,
include_distances=True)
pose = headpose.get_pose(nimg)
cv2.imwrite(
os.path.join(FACE_IMAGE_PATH, "aligned_" + face_image),
nimg)
f.write(",pose,{},{},{}".format(pose[0], pose[1], pose[2]))
identified_face_top_1 = face_db_pd[(match[0][0])]
score_top_1 = match[1][0]
if score_top_1 < config.MATCH_THRESHOLD:
cv2.putText(img, identified_face_top_1, (x1, y1),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0))
for i in range(config.TOP_ACC_NUM):
identified_face = face_db_pd[(match[0][i])]
if (abs(pose[0]) > config.POSE_THRESHOLD
or abs(pose[1]) > config.POSE_THRESHOLD
or abs(pose[2]) > config.POSE_THRESHOLD):
continue
if (match[1][i] < config.MATCH_THRESHOLD):
f.write(",{},{}".format(
identified_face, match[1][i]))
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 1)
frame_count += 1
f.write("\n")
cv2.imshow("img_window", img)
fps.update()
if cv2.waitKey(1) & 0xFF == ord('q'):
break
fps.stop()
cv2.destroyAllWindows()
fvs.stop()
f.close()
def initialiseAnnoy(part_of_speech):
# Loads Annoy ADJ vectors
if part_of_speech == "adjective":
adj_annoy = AnnoyIndex(50, metric='angular')
adjectives = list()
adj_lookup = dict()
print("\n\tLoading adjective vectors...")
for i, line in enumerate(open("./vectors/adjVectors", "r")):
line = line.strip()
word, vec_s = line.split(" ")
vec = [float(n) for n in vec_s.split()]
adj_annoy.add_item(i, vec)
adj_lookup[word] = vec
adjectives.append(word)
adj_annoy.build(50)
return adj_annoy, adjectives, adj_lookup
# Loads Annoy VERB vectors
elif part_of_speech == "verb":
verb_annoy = AnnoyIndex(50, metric='angular')
verbs = list()
verb_lookup = dict()
print("\n\tLoading verb vectors...")
for i, line in enumerate(open("./vectors/verbVectors", "r")):
line = line.strip()
word, vec_s = line.split(" ")
vec = [float(n) for n in vec_s.split()]
verb_annoy.add_item(i, vec)
verb_lookup[word] = vec
verbs.append(word)
verb_annoy.build(50)
return verb_annoy, verbs, verb_lookup
# Loads Annoy NOUN vectors
elif part_of_speech == "noun":
noun_annoy = AnnoyIndex(50, metric='angular')
nouns = list()
noun_lookup = dict()
print("\n\tLoading noun vectors...")
for i, line in enumerate(open("./vectors/nounVectors", "r")):
line = line.strip()
word, vec_s = line.split(" ")
vec = [float(n) for n in vec_s.split()]
noun_annoy.add_item(i, vec)
noun_lookup[word] = vec
nouns.append(word)
noun_annoy.build(50)
return noun_annoy, nouns, noun_lookup
else:
raise Exception("Part of speech must either be 'adjective', 'verb' or 'noun'.")
def main(project_name):
tic = time.time()
logger = Logger('_05_make_submission_1000_{}'.format(project_name))
logger.info('=' * 50)
model_path = '_model/embedding_model_{}.pt'.format(project_name)
logger.info('load model from {}'.format(model_path))
model = torch.load(model_path)
model.eval()
dir_target = '../../input/test'
embedder = ImgEmbedder(model, dir_target)
sample_submission = pd.read_csv('../../dataset/sample_submission.csv')
images = list()
with open(
os.path.join('_embed_index',
'index_names_{}.json'.format(project_name)),
'r') as f:
index_names = json.load(f)
test_id_list = sample_submission.id
f = 512
u = AnnoyIndex(f, metric='euclidean')
u.load(
os.path.join('_embed_index',
'index_features_{}.ann'.format(project_name)))
logger.info('===> embed test images and get nearest neighbors')
search_k = 1_000_000
for test_id in tqdm(test_id_list):
target_file = '{}.jpg'.format(test_id)
try:
img_feature = embedder.get_vector(target_file)
indeces = u.get_nns_by_vector(img_feature.tolist(),
n=1000,
search_k=search_k)
except:
indeces = list(range(1000))
names = [index_names[index] for index in indeces]
images.append(' '.join(names))
submission = pd.DataFrame(test_id_list, columns=['id'])
submission['images'] = images
output_path = '../../submission/submission_1000_{}.csv'.format(
project_name)
submission.to_csv(output_path, index=False)
toc = time.time() - tic
logger.info('Elapsed time: {:.1f} [min]'.format(toc / 60.0))
def test_item_vector_after_save(self):
# Issue #279
a = AnnoyIndex(3)
a.verbose(True)
a.add_item(1, [1, 0, 0])
a.add_item(2, [0, 1, 0])
a.add_item(3, [0, 0, 1])
a.build(-1)
self.assertEquals(a.get_n_items(), 4)
self.assertEquals(a.get_item_vector(3), [0, 0, 1])
self.assertEquals(set(a.get_nns_by_item(1, 999)), set([1, 2, 3]))
a.save('something.annoy')
self.assertEquals(a.get_n_items(), 4)
self.assertEquals(a.get_item_vector(3), [0, 0, 1])
self.assertEquals(set(a.get_nns_by_item(1, 999)), set([1, 2, 3]))
def test_seed(self):
i = AnnoyIndex(10)
i.load('test/test.tree')
i.set_seed(42)
def test_metric_f_kwargs(self):
i = AnnoyIndex(f=3, metric='euclidean')
def get_subsampling_index2(data_process, standard_scale = True, cutoff_sig = 0.02, rate = 0.3, \
method = "pykdtree", verbose = 1, image_index = []):
"""
Using Nearest-Neighbor search based algorithm, find the list of indices of the subsampled dataset
Parameters
-------------
data_process: List. the list of datapoints, with selected features
standard_scale [True]: Boolean. Whether to apply standard scaler to the dataset prior to subsampling
cutoff_sig [0.02]: Float. cutoff significance. the cutoff distance equals to the Euclidean
norm of the standard deviations in all dimensions of the data points
rate [0.3]: Float. possibility of deletion
method ["pykdtree"]: String. which backend nearest neighbour model to use.
possible choices: ["pykdtree", "nmslib", "sklearn", "scipy", "annoy", "flann"]
verbose [1]: integer. level of verbosity
Return
-------------
overall_keep_list: The list of indices of the final subsampled entries
"""
if verbose >= 1:
print("Started NN-subsampling, original length: {}".format(
len(data_process)))
method = method.lower()
start = time.time()
if method == "flann":
if verbose >= 1:
print("use flann backend")
elif method == "pykdtree":
if verbose >= 1:
print("use pykdtree backend")
elif method == "sklearn":
if verbose >= 1:
print("use slearn nearest neighbors backend")
elif method == "scipy":
if verbose >= 1:
print("use scipy cKDTree backend")
elif method == "annoy":
if verbose >= 1:
print("use annoy backend")
elif method == "nmslib":
if verbose >= 1:
print("use nmslib backend")
else:
print("method {} not impletemented".format(method))
raise NotImplemented
# apply standard scaling
if standard_scale:
if verbose >= 2:
print("Subample with standard scaled data")
data_process = StandardScaler().fit_transform(
np.asarray(data_process).copy())
else:
if verbose >= 2:
print("Subample with original data")
data_process = np.asarray(data_process).copy()
#set cutoff distance
list_of_descs = zip(*data_process)
sum_std2 = 0.
for descs in list_of_descs:
temp_std = np.std(descs)
sum_std2 += temp_std**2
cutoff = cutoff_sig * np.sqrt(sum_std2)
#initialize the index
overall_keep_list = np.arange(len(data_process)).tolist()
keep_going = True
iter_count = 1
while keep_going:
if verbose >= 2:
print('start iteration {}, total length: {}'.format(
iter_count, len(overall_keep_list)))
start_cycle = time.time()
temp_data_process = get_array_based_on_index(data_process.copy(),
overall_keep_list)
temp_image_index = get_array_based_on_index(image_index,
overall_keep_list)
#build and query nearest neighbour model
if method == "flann":
flann = FLANN()
indices, distances = flann.nn(temp_data_process,
temp_data_process,
2,
algorithm="kmeans")
elif method == "scipy":
kd_tree = cKDTree(temp_data_process)
distances, indices = kd_tree.query(temp_data_process, k=2)
elif method == "pykdtree":
kd_tree = KDTree(temp_data_process, leafsize=6)
distances, indices = kd_tree.query(temp_data_process, k=2)
elif method == "sklearn":
nbrs = NearestNeighbors(n_neighbors=2,
algorithm='kd_tree',
n_jobs=-1).fit(temp_data_process)
distances, indices = nbrs.kneighbors(temp_data_process)
elif method == "annoy":
annoy = AnnoyIndex(len(temp_data_process[0]), metric='euclidean')
for i in range(len(temp_data_process)):
annoy.add_item(i, temp_data_process[i])
annoy.build(1)
distances = []
indices = []
for i in range(len(temp_data_process)):
temp_index, temp_dist = annoy.get_nns_by_vector(
temp_data_process[i], 2, include_distances=True)
indices.append([i, temp_index[1]])
distances.append([0.0, temp_dist[1]])
elif method == "nmslib":
index = nmslib.init(method='hnsw', space='l2')
index.addDataPointBatch(temp_data_process)
index.createIndex(print_progress=False)
neighbours = index.knnQueryBatch(temp_data_process, k=2)
distances = []
indices = []
for item in neighbours:
indices.append(item[0])
distances.append(item[1])
else:
raise NotImplemented
# if distance between each point and its nearest neighbor is below cutoff distance,
# add the nearest neighbout to the candidate removal list
remove_index_li = []
index_li = []
for index, distance in zip(indices, distances):
index_li.append(index[0])
if distance[1] <= cutoff:
remove_index_li.append(index[1])
# randomly select datapoints in the candidate removal list (based on rate)
# and form the final removal list of this iteration
# stop the cycle if the final removal list is empty
temp_num = int(ceil(float(len(remove_index_li)) * rate))
if temp_num == 0:
keep_going = False
#remove_index_li = random_subsampling(remove_index_li,temp_num)
remove_index_li = rank_subsampling(remove_index_li, temp_num,
temp_image_index)
temp_keep_list = remove_list_from_list(index_li, remove_index_li)
overall_keep_list = [overall_keep_list[i] for i in temp_keep_list]
try:
if len(overall_keep_list) == old_overall_keep_list_len:
keep_going = False
print("stopped because length didn't change")
except:
pass
if verbose >= 2:
print('end iteration {}. length: {}\t time:{}'.format(
iter_count, len(overall_keep_list),
time.time() - start_cycle))
iter_count += 1
old_overall_keep_list_len = len(overall_keep_list)
if verbose >= 1:
print('end NN-subsampling. length: {}\t time:{}'.format(
len(overall_keep_list),
time.time() - start))
return overall_keep_list
def test_save_twice(self):
# Issue #100
t = AnnoyIndex(10)
t.save("t.ann")
t.save("t.ann")
def test_construct_destruct(self):
for x in range(100000):
i = AnnoyIndex(10)
i.add_item(1000, [random.gauss(0, 1) for z in range(10)])
def test_construct_load_destruct(self):
for x in range(100000):
i = AnnoyIndex(10)
i.load('test/test.tree')
def test_load_unload(self):
# Issue #108
i = AnnoyIndex(10)
for x in range(100000):
i.load('test/test.tree')
i.unload()
from nltk import ngrams
import random, json, glob, os, codecs, random
import numpy as np
# data structures
file_index_to_file_name = {}
file_index_to_file_vector = {}
# config
dims = 2048
n_nearest_neighbors = 200
trees = 100
infiles = glob.glob('feature_vectors/*.npz')
# build ann index
t = AnnoyIndex(dims)
for file_index, i in enumerate(infiles):
file_vector = np.loadtxt(i)
file_name = os.path.basename(i).split('.')[0]
print("file_name = %s" % file_name)
file_index_to_file_name[file_index] = file_name
file_index_to_file_vector[file_index] = file_vector
t.add_item(file_index, file_vector)
t.build(trees)
# create a nearest neighbors json file for each input
if not os.path.exists('bran_neighbors'):
os.makedirs('bran_neighbors')
# do the similarity search for the feature vectors
# this is a <filename*.npz> file
def generate_triplets_from_ANN(model, sequences, entity2unique, entity2same, unique_text, test):
predictions = model.predict(sequences)
t = AnnoyIndex(len(predictions[0]), metric='euclidean') # Length of item vector that will be indexed
t.set_seed(123)
for i in range(len(predictions)):
# print(predictions[i])
v = predictions[i]
t.add_item(i, v)
t.build(100) # 100 trees
match = 0
no_match = 0
ann_accuracy = 0
total = 0
precise = 0
triplets = {}
closest_positive_counts = []
pos_distances = []
neg_distances = []
all_pos_distances = []
all_neg_distances = []
triplets['anchor'] = []
triplets['positive'] = []
triplets['negative'] = []
if test:
NNlen = TEST_NEIGHBOR_LEN
else:
NNlen = TRAIN_NEIGHBOR_LEN
for key in entity2same:
index = entity2unique[key]
nearest = t.get_nns_by_vector(predictions[index], NNlen)
nearest_text = set([unique_text[i] for i in nearest])
expected_text = set(entity2same[key])
# annoy has this annoying habit of returning the queried item back as a nearest neighbor. Remove it.
if key in nearest_text:
nearest_text.remove(key)
# print("query={} names = {} true_match = {}".format(unique_text[index], nearest_text, expected_text))
overlap = expected_text.intersection(nearest_text)
# collect up some statistics on how well we did on the match
m = len(overlap)
match += m
# since we asked for only x nearest neighbors, and we get at most x-1 neighbors that are not the same as key (!)
# make sure we adjust our estimate of no match appropriately
no_match += min(len(expected_text), NNlen - 1) - m
# sample only the negatives that are true negatives
# that is, they are not in the expected set - sampling only 'semi-hard negatives is not defined here'
# positives = expected_text - nearest_text
positives = overlap
negatives = nearest_text - expected_text
# print(key + str(expected_text) + str(nearest_text))
for i in negatives:
for j in positives:
dist_pos = t.get_distance(index, entity2unique[j])
pos_distances.append(dist_pos)
dist_neg = t.get_distance(index, entity2unique[i])
neg_distances.append(dist_neg)
if dist_pos < dist_neg:
ann_accuracy += 1
total += 1
# print(key + "|" + j + "|" + i)
# print(dist_pos)
# print(dist_neg)
min_neg_distance = 1000000
for i in negatives:
dist_neg = t.get_distance(index, entity2unique[i])
all_neg_distances.append(dist_neg)
if dist_neg < min_neg_distance:
min_neg_distance = dist_neg
for j in expected_text:
dist_pos = t.get_distance(index, entity2unique[j])
all_pos_distances.append(dist_pos)
closest_pos_count = 0
for p in overlap:
dist_pos = t.get_distance(index, entity2unique[p])
if dist_pos < min_neg_distance:
closest_pos_count+=1
if closest_pos_count > 0:
precise+=1
closest_positive_counts.append(closest_pos_count / min(len(expected_text), NNlen - 1))
for i in negatives:
for j in expected_text:
triplets['anchor'].append(key)
triplets['positive'].append(j)
triplets['negative'].append(i)
print("mean closest positive count:" + str(statistics.mean(closest_positive_counts)))
print("mean positive distance:" + str(statistics.mean(pos_distances)))
print("stdev positive distance:" + str(statistics.stdev(pos_distances)))
print("max positive distance:" + str(max(pos_distances)))
print("mean neg distance:" + str(statistics.mean(neg_distances)))
print("stdev neg distance:" + str(statistics.stdev(neg_distances)))
print("max neg distance:" + str(max(neg_distances)))
print("mean all positive distance:" + str(statistics.mean(all_pos_distances)))
print("stdev all positive distance:" + str(statistics.stdev(all_pos_distances)))
print("max all positive distance:" + str(max(all_pos_distances)))
print("mean all neg distance:" + str(statistics.mean(all_neg_distances)))
print("stdev all neg distance:" + str(statistics.stdev(all_neg_distances)))
print("max all neg distance:" + str(max(all_neg_distances)))
print("Accuracy in the ANN for triplets that obey the distance func:" + str(ann_accuracy / total))
print("Precision at 1: " + str(precise / len(entity2same)))
obj = {}
obj['accuracy'] = ann_accuracy / total
obj['steps'] = 1
with open(output_file_name_for_hpo, 'w', encoding='utf8') as out:
json.dump(obj, out)
if test:
return match/(match + no_match)
else:
return triplets, match/(match + no_match)
for p in classif['body']['predictions']:
uri = p['uri']
rois = p['rois']
sys.stdout.write('\rIndexing image '+str(d)+'/'+str(len(onlyfiles)) + ' : ' + str(len(rois)) + ' rois total:' + str(c) + ' ')
sys.stdout.flush()
for roi in rois:
bbox = roi['bbox']
cat = roi['cat']
prob = roi['prob']
vals = roi['vals']
if c == 0:
layer_size = len(vals)
s['layer_size'] = layer_size
t = AnnoyIndex(layer_size,metric) # prepare index
t.add_item(c,vals)
s[str(c)] = {'uri':uri, 'bbox' : bbox, 'cat' : cat, 'prob' : prob}
c = c + 1
d = d + 1
#if c >= 10000:
# break
print 'building index...\n'
print 'layer_size=',layer_size
t.build(ntrees)
t.save('index.ann')
s.close()
if args.search:
s = shelve.open('data.bin')
u = AnnoyIndex(s['layer_size'],metric)
#!/bin/env python3
import sys, argparse
from annoy import AnnoyIndex
import re
import statistics
parser = argparse.ArgumentParser()
parser.add_argument('--file', help='Input file')
parser.add_argument('--out', help='Outfile base')
parser.add_argument('--L', help='Fingerprint length')
parser.add_argument('--norm', help='Normalize')
args = parser.parse_args()
a = AnnoyIndex(int(args.L))
i = 0
names = []
with open(args.file, 'r') as f:
for line in f:
id, statements, *v = line.split("\t")
id = re.sub('.json.gz', '', id)
id = re.sub('\.', '|', id)
names.append(id)
v = [float(j) for j in v]
if args.norm:
avg = statistics.mean(v)
std = statistics.stdev(v)
v = [(j - avg) / std for j in v]
a.add_item(i, v)
i = i + 1
class FakeDetector:
"""
#TODO: Complete the description
"""
def __init__(self):
"""
"""
self.tree = None
self.cos = nn.CosineSimilarity(dim=1, eps=1e-6)
# self.embedder = EmbedSentence()
def build(self,
brand_names: str,
n_tree: int = 100,
embedding_size: int = 100):
"""
#TODO: fill details of param
:param brand_names:
:param n_tree:
:param embedding_size:
:return:
"""
brand_names_sentence = sentence_char2vec(brand_names)
self.tree = AnnoyIndex(embedding_size, 'angular')
for value, (_, _token) in enumerate(brand_names_sentence.items()):
self.tree.add_item(value, _token)
self.tree.build(n_tree)
def fake_detector(
self,
text: str,
embedding_size: int = 100,
detection_range: Tuple = (0.97, 0.99)) -> bool:
"""
#TODO: fill details of params
:param text:
:param embedding_size:
:param detection_range:
:return:
"""
found_match = False
text_sentence = sentence_char2vec(text)
for _, (_, _token) in enumerate(text_sentence.items()):
match = self.tree.get_nns_by_vector(_token, 1)[0]
sim_score = round(
float(
self.cos(
_token.view(-1, embedding_size),
torch.tensor(self.tree.get_item_vector(match)).view(
-1, embedding_size))), 2)
if detection_range[0] <= sim_score <= detection_range[1]:
found_match = True
if found_match:
break
return found_match
# In[27]:
from annoy import AnnoyIndex
# In[28]:
# Choose a random image to experiment
random_image_index = random.randint(0, num_items)
# Note: the results may change if the image is changed
# First, we build a search index with two hyperparameters - the number of dimensions of the dataset, and the number of trees.
# In[29]:
annoy_index = AnnoyIndex(
num_dimensions) # Length of item vector that will be indexed
for i in range(num_items):
annoy_index.add_item(i, dataset[i])
annoy_index.build(40) #40 trees
# Now let’s find out the time it takes to search the 5 nearest neighbors of one image.
# In[30]:
#u = AnnoyIndex(num_dimensions)
#Time the search for one image for Annoy
get_ipython().run_line_magic(
'timeit',
'annoy_index.get_nns_by_vector(query, 5, include_distances=True )')
# Now THAT is blazing fast! To put this in perspective, for such a modestly sized dataset, this can serve almost 15000 requests on a single CPU core. Considering most CPUs have multiple cores, it should be able to handle 100K+ requests on a single system. The best part is that it lets you share the same index in memory between multiple processes. Hence, the biggest index can be equivalent to the size of your overall RAM, making it possible to serve multiple requests on a single system.
def test_load_save(self):
# Issue #61
i = AnnoyIndex(10)
i.load('test/test.tree')
u = i.get_item_vector(99)
i.save('i.tree')
v = i.get_item_vector(99)
self.assertEqual(u, v)
j = AnnoyIndex(10)
j.load('test/test.tree')
w = i.get_item_vector(99)
self.assertEqual(u, w)
# Ensure specifying if prefault is allowed does not impact result
j.save('j.tree', True)
k = AnnoyIndex(10)
k.load('j.tree', True)
x = k.get_item_vector(99)
self.assertEqual(u, x)
k.save('k.tree', False)
l = AnnoyIndex(10)
l.load('k.tree', False)
y = l.get_item_vector(99)
self.assertEqual(u, y)
vectorArr.append(vector)
except Exception as e:
print(per)
print(e)
np.save('vectors', vectorArr)
np.save('persons', nameArr)
###########################################
## ANNOY SIDE ##
###########################################
file = np.load('vectors.npy')
f = 512
t = AnnoyIndex(
f, metric="euclidean") # Length of item vector that will be indexed
for i in range(len(file)):
v = file[i]
t.add_item(i, v)
t.build(10) # 10 trees
t.save('test.ann')
u = AnnoyIndex(f, metric="euclidean")
u.load('test.ann') # super fast, will just mmap the file
av = np.load('vectors.npy')
pr = np.load('persons.npy')
arr = u.get_nns_by_vector(av[0], 3, include_distances=True)
def test_load_save(self):
# Issue #61
i = AnnoyIndex(10)
i.load('test/test.tree')
u = i.get_item_vector(99)
i.save('x.tree')
v = i.get_item_vector(99)
self.assertEqual(u, v)
j = AnnoyIndex(10)
j.load('test/test.tree')
w = i.get_item_vector(99)
self.assertEqual(u, w)
class EntityType(object):
"""Convenience wrapper around Annoy.
More generally a way to collect vectors within the same entity type and
quickly find similar vectors.
* Helps deal with non-contiguous ids through an id map.
* Checks for 0 vectors before returning matches.
"""
def __init__(self, nfactor, ntrees, metric='angular',
entity_type_id=None, entity_type=None):
"""Initialize EntityType."""
# metadata
self._nfactor = nfactor
self._metric = metric
# object is accessed using this id. e.g. 'user'
self._entity_type = entity_type
# data is loaded in using this id. This can be more compact than the
# entity_type, depending on the data source
self._entity_type_id = entity_type_id
self._ntrees = ntrees
# data
self._ann_obj = AnnoyIndex(nfactor, metric)
# maps entity id to internal representation of id
self._ann_map = {}
# maps internal representation of id to entity id
self._ann_map_inv = {}
self._nitems = 0
def add_item(self, entity_id, factors):
"""Add item, populating id map."""
if entity_id in self._ann_map:
raise ValueError('Duplicate entity: type = {0}, id = {1}'.format(
self._entity_type, entity_id))
self._ann_obj.add_item(self._nitems, factors)
self._ann_map[entity_id] = self._nitems
self._nitems = self._nitems + 1
def build(self, verbose=False):
"""Build annoy model, create invert dictionary for future lookups."""
self._ann_obj.verbose(verbose)
self._ann_obj.build(self._ntrees)
# this is only necessary after build, so we'll create it here
self._ann_map_inv = {v: k for k, v in self._ann_map.items()}
def get_nns_by_vector(self, vec, n, search_k):
"""Get nearest neighbors from an input vector."""
nns = self._ann_obj.get_nns_by_vector(vec, n, search_k)
return [self._ann_map_inv[x] for x in nns]
def get_item_vector(self, entity_id):
"""Get a vector for an entity."""
if entity_id in self._ann_map:
return self._ann_obj.get_item_vector(self._ann_map[entity_id])
else:
return []
def __iter__(self):
"""Iterate over object, return (entity_id, vector) tuples."""
return (EntityVector(
entity_id=entity_id,
vector=self.get_item_vector(entity_id)
) for entity_id in self._ann_map.keys())
def get_nfactor(self):
return self._nfactor
def load(self, pkl, filepath):
entity_type = pkl.get_entity_type(self._entity_type_id)
self.__dict__ = entity_type.__dict__
# initialize index
self._ann_obj = AnnoyIndex(pkl.get_nfactor(), entity_type._metric)
# mmap the file
self._ann_obj.load(filepath)
words_to_wvs = {}
import numpy as np
with open("wvs.txt") as rf:
for line in rf:
split_line = line.strip().split("\t")
if len(split_line) == 2:
word = split_line[0]
#print(word)
vec = [float(i) for i in split_line[1].split(" ")]
words.append(word)
words_to_wvs[word] = vec
w1 = "pikachu"
w2 = 'kanto'
w3 = "sinnoh"
w4 = "pachirisu"
words.append("vec({}) - vec({}) + vec({}) =".format(w1, w2, w3))
words_to_wvs["vec({}) - vec({}) + vec({}) =".format(w1, w2, w3)] = list(
np.asarray(words_to_wvs[w1]) - np.asarray(words_to_wvs[w2]) +
np.asarray(words_to_wvs[w3]))
t = AnnoyIndex(100, 'angular') # Length of item vector that will be indexed
for ct, i in enumerate(words):
t.add_item(ct, words_to_wvs[i])
t.build(100) # 10 trees
wwca = [w1, w2, w3, w4, "vec({}) - vec({}) + vec({}) =".format(w1, w2, w3)]
for word in wwca:
print("10 NEAREST NEIGHBORS OF {}".format(word))
nearest_neighbors = t.get_nns_by_item(words.index(word), 10)
for nn in nearest_neighbors:
print(words[nn])
def test_save_without_build(self):
# Issue #61
i = AnnoyIndex(10)
i.add_item(1000, [random.gauss(0, 1) for z in range(10)])
i.save('x.tree')
j = AnnoyIndex(10)
j.load('x.tree')
j.build(10)
def main(_):
parser = argparse.ArgumentParser(description='TransE.')
parser.add_argument('--data',
dest='data_dir',
type=str,
help="Data folder",
default='./data/FB15k/')
parser.add_argument('--lr',
dest='lr',
type=float,
help="Learning rate",
default=1e-2)
parser.add_argument("--dim",
dest='dim',
type=int,
help="Embedding dimension",
default=256)
parser.add_argument("--batch",
dest='batch',
type=int,
help="Batch size",
default=32)
parser.add_argument("--worker",
dest='n_worker',
type=int,
help="Evaluation worker",
default=3)
parser.add_argument("--generator",
dest='n_generator',
type=int,
help="Data generator",
default=10)
parser.add_argument("--eval_batch",
dest="eval_batch",
type=int,
help="Evaluation batch size",
default=32)
parser.add_argument("--save_dir",
dest='save_dir',
type=str,
help="Model path",
default='./transE')
parser.add_argument("--load_model",
dest='load_model',
type=str,
help="Model file",
default="")
parser.add_argument("--save_per",
dest='save_per',
type=int,
help="Save per x iteration",
default=1)
parser.add_argument("--eval_per",
dest='eval_per',
type=int,
help="Evaluate every x iteration",
default=3)
parser.add_argument("--max_iter",
dest='max_iter',
type=int,
help="Max iteration",
default=30)
parser.add_argument("--summary_dir",
dest='summary_dir',
type=str,
help="summary directory",
default='./transE_summary/')
parser.add_argument("--keep",
dest='drop_out',
type=float,
help="Keep prob (1.0 keep all, 0. drop all)",
default=0.5)
parser.add_argument("--optimizer",
dest='optimizer',
type=str,
help="Optimizer",
default='gradient')
parser.add_argument("--prefix",
dest='prefix',
type=str,
help="model_prefix",
default='DEFAULT')
parser.add_argument("--loss_weight",
dest='loss_weight',
type=float,
help="Weight on parameter loss",
default=1e-2)
parser.add_argument("--neg_weight",
dest='neg_weight',
type=float,
help="Sampling weight on negative examples",
default=0.5)
parser.add_argument("--save_per_batch",
dest='save_per_batch',
type=int,
help='evaluate and save after every x batches',
default=1000)
parser.add_argument(
"--outfile_prefix",
dest='outfile_prefix',
type=str,
help='The filename of output file is outfile_prefix.txt',
default='test_output')
parser.add_argument("--neg_sample",
dest='neg_sample',
type=int,
help='No. of neg. samples per (h,r) or (t,r) pair',
default=5)
parser.add_argument(
"--fanout_thresh",
dest='fanout_thresh',
type=int,
help='threshold on fanout of entities to be considered',
default=2)
parser.add_argument('--annoy_n_trees',
dest='annoy_n_trees',
type=int,
help='builds a forest of n_trees trees',
default=10)
parser.add_argument(
'--annoy_search_k',
dest='annoy_search_k',
type=int,
help='During the query it will inspect up to search_k nodes',
default=-1)
parser.add_argument('--eval_after',
dest='eval_after',
type=int,
help='Evaluate after this many no. of epochs',
default=4)
args = parser.parse_args()
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
print(args)
model = TransE(args.data_dir,
embed_dim=args.dim,
fanout_thresh=args.fanout_thresh,
eval_batch=args.eval_batch)
train_pos_neg_list, \
train_loss, train_op = train_ops(model, learning_rate=args.lr,
optimizer_str=args.optimizer,
regularizer_weight=args.loss_weight)
get_embedding_op = embedding_ops(model)
# test_input, test_head, test_tail = test_ops(model)
f1 = open('%s/%s.txt' % (args.save_dir, args.outfile_prefix), 'w')
with tf.Session() as session:
tf.global_variables_initializer().run()
all_var = tf.all_variables()
print 'printing all', len(all_var), ' TF variables:'
for var in all_var:
print var.name, var.get_shape()
saver = tf.train.Saver(restore_sequentially=True)
iter_offset = 0
if args.load_model is not None and os.path.exists(args.load_model):
saver.restore(session, args.load_model)
iter_offset = int(
args.load_model.split('.')[-2].split('_')[-1]) + 1
f1.write("Load model from %s, iteration %d restored.\n" %
(args.load_model, iter_offset))
total_inst = model.n_train
best_filtered_mean_rank = float("inf")
f1.write("preparing training data...\n")
nbatches_count = 0
# training_data_list = []
training_data_pos_neg_list = []
for dat in model.raw_training_data(batch_size=args.batch):
# raw_training_data_queue.put(dat)
# training_data_list.append(dat)
ps_list = data_generator_func(dat, model.tr_h, model.hr_t,
model.n_entity, args.neg_sample,
model.n_relation)
assert ps_list is not None
training_data_pos_neg_list.append(ps_list)
nbatches_count += 1
f1.write("training data prepared.\n")
f1.write("No. of batches : %d\n" % nbatches_count)
f1.close()
start_time = timeit.default_timer()
for n_iter in range(iter_offset, args.max_iter):
accu_loss = 0.
ninst = 0
# f1.close()
for batch_id in range(nbatches_count):
f1 = open('%s/%s.txt' % (args.save_dir, args.outfile_prefix),
'a')
pos_neg_list = training_data_pos_neg_list[batch_id]
#print data_e
l, _ = session.run([train_loss, train_op],
{train_pos_neg_list: pos_neg_list})
accu_loss += l
ninst += len(pos_neg_list)
# print('len(pos_neg_list) = %d\n' % len(pos_neg_list))
if ninst % (5000) is not None:
f1.write('[%d sec](%d/%d) : %.2f -- loss : %.5f \n' %
(timeit.default_timer() - start_time, ninst,
total_inst, float(ninst) / total_inst, l))
f1.close()
f1 = open('%s/%s.txt' % (args.save_dir, args.outfile_prefix), 'a')
f1.write("")
f1.write("iter %d avg loss %.5f, time %.3f\n" %
(n_iter, accu_loss / ninst,
timeit.default_timer() - start_time))
# if n_iter == args.max_iter - 1:
# save_path = saver.save(session,
# os.path.join(args.save_dir,
# "TransE_" + str(args.prefix) + "_" + str(n_iter) + ".ckpt"))
# f1.write("Model saved at %s\n" % save_path)
with tf.device('/cpu'):
if n_iter > args.eval_after and (n_iter % args.eval_per == 0 or
n_iter == args.max_iter - 1):
t = AnnoyIndex(model.embed_dim, metric='euclidean')
ent_embedding, rel_embedding = session.run(
get_embedding_op, {train_pos_neg_list: pos_neg_list})
# sess = tf.InteractiveSession()
# with sess.as_default():
# ent_embedding = model.ent_embeddings.eval()
print np.asarray(ent_embedding).shape
print np.asarray(rel_embedding).shape
# print ent_embedding[10,:]
# print rel_embedding[10,:]
print 'Index creation started'
for i in xrange(model.n_entity):
v = ent_embedding[i, :]
t.add_item(i, v)
t.build(args.annoy_n_trees)
print 'Index creation completed'
# n = int(0.0005 * model.n_entity)
n = 1000
# search_k = int(n * args.annoy_n_trees/100.0)
search_k = 1000
print 'No. of items = %d' % t.get_n_items()
print sum(t.get_item_vector(0))
print sum(ent_embedding[0, :])
assert sum(t.get_item_vector(0)) == sum(
ent_embedding[0, :])
if n_iter == args.max_iter - 1:
eval_dict = zip(
[model.validation_data, model.testing_data],
['VALID', 'TEST'])
else:
eval_dict = zip([model.validation_data], ['VALID'])
for data_func, test_type in eval_dict:
accu_mean_rank_h = list()
accu_mean_rank_t = list()
accu_filtered_mean_rank_h = list()
accu_filtered_mean_rank_t = list()
evaluation_count = 0
evaluation_batch = []
batch_id = 0
for testing_data in data_func(
batch_size=args.eval_batch):
batch_id += 1
print 'test_type: %s, batch id: %d' % (test_type,
batch_id)
head_ids = list()
tail_ids = list()
for i in xrange(testing_data.shape[0]):
# try:
# print (ent_embedding[testing_data[i,0],:] + rel_embedding[testing_data[i,2],:])
tail_ids.append(
t.get_nns_by_vector(
(ent_embedding[testing_data[i, 0], :] +
rel_embedding[testing_data[i, 2], :]),
n, search_k))
head_ids.append(
t.get_nns_by_vector(
(ent_embedding[testing_data[i, 1], :] -
rel_embedding[testing_data[i, 2], :]),
n, search_k))
# except:
# print 'i = %d' % i
# print 'testing_data[i,0] = %d' % testing_data[i,0]
# print 'testing_data[i,1] = %d' % testing_data[i,1]
# print 'testing_data[i,2] = %d' % testing_data[i,2]
# print head_ids
# print tail_ids
evaluation_batch.append(
(testing_data, head_ids, tail_ids))
evaluation_count += 1
while evaluation_count > 0:
evaluation_count -= 1
# (mrh, fmrh), (mrt, fmrt) = result_queue.get()
(mrh, fmrh), (mrt, fmrt) = worker_func(
evaluation_batch[evaluation_count - 1],
model.hr_t, model.tr_h)
accu_mean_rank_h += mrh
accu_mean_rank_t += mrt
accu_filtered_mean_rank_h += fmrh
accu_filtered_mean_rank_t += fmrt
f1.write(
"[%s] ITER %d [HEAD PREDICTION] MEAN RANK: %.1f FILTERED MEAN RANK %.1f HIT@10 %.3f FILTERED HIT@10 %.3f\n"
% (test_type, n_iter, np.mean(accu_mean_rank_h),
np.mean(accu_filtered_mean_rank_h),
np.mean(
np.asarray(accu_mean_rank_h, dtype=np.int32)
< 10),
np.mean(
np.asarray(accu_filtered_mean_rank_h,
dtype=np.int32) < 10)))
f1.write(
"[%s] ITER %d [TAIL PREDICTION] MEAN RANK: %.1f FILTERED MEAN RANK %.1f HIT@10 %.3f FILTERED HIT@10 %.3f\n"
% (test_type, n_iter, np.mean(accu_mean_rank_t),
np.mean(accu_filtered_mean_rank_t),
np.mean(
np.asarray(accu_mean_rank_t, dtype=np.int32)
< 10),
np.mean(
np.asarray(accu_filtered_mean_rank_t,
dtype=np.int32) < 10)))
if test_type == 'VALID':
filtered_mean_rank = (
np.mean(accu_filtered_mean_rank_t) +
np.mean(accu_mean_rank_h)) / 2.0
if filtered_mean_rank < best_filtered_mean_rank:
save_path = saver.save(
session,
os.path.join(
args.save_dir,
"TransE_" + str(args.prefix) + "_" +
str(n_iter) + ".ckpt"))
f1.write("Model saved at %s\n" % save_path)
best_filtered_mean_rank = filtered_mean_rank
f1.close()
def test_unbuild_with_loaded_tree(self):
i = AnnoyIndex(10)
i.load('test/test.tree')
i.unbuild()
from keras.preprocessing import image
from keras.applications.vgg16 import preprocess_input, VGG16
from keras.models import Model
from annoy import AnnoyIndex
# img_dir_path = 'dataset/All/'
img_dir_path = 'dataDrivenArt/bin/data/images/'
annoy_model_path = 'model/x-fresh-flatten.ann'
# annoy_dim = 4096 # fc2
annoy_dim = 25088
base_model = VGG16(weights='imagenet')
model = Model(inputs=base_model.input,
outputs=base_model.get_layer('flatten').output)
annoy_model = AnnoyIndex(annoy_dim)
for i in range(1, 3988):
# img_path = img_dir_path + str(i) + '.jpg'
img_path = "C:\\Users\\santa\\Desktop\\Python\\x-fresh\\dataDrivenArt\\bin\\data\\images\\{0:04d}.jpg".format(
i + 0)
img = image.load_img(img_path, target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
flatten_features = model.predict(x)
annoy_model.add_item(i, flatten_features[0])
print(img_path, 'saved')
definitions[definitions_index]['language'],
}
for resource in [
data_pipeline.TREE_LABEL, data_pipeline.GRAPH_LABEL,
data_pipeline.CODE_TOKENS_LABEL
]:
sample[resource] = loaded_sample[resource]
samples.append(sample)
definitions_index += 1
data_file_code_representations = model.get_code_representations(
samples)
code_representations_all.extend(data_file_code_representations)
print('len(code_representations_all)', len(code_representations_all))
indices = AnnoyIndex(code_representations_all[0].shape[0], 'angular')
for index, vector in tqdm(enumerate(code_representations_all)):
if vector is not None:
indices.add_item(index, vector)
indices.build(200)
print('Index is built')
for query in queries:
for idx, _ in zip(*query_model(query, model, indices, language)):
predictions.append(
(query, language, definitions[idx]['identifier'],
definitions[idx]['url']))
df = pd.DataFrame(predictions,
columns=['query', 'language', 'identifier', 'url'])
df.to_csv(predictions_csv, index=False)
def test_not_found_tree(self):
i = AnnoyIndex(10)
self.assertRaises(IOError, i.load, 'nonexists.tree')
def load_annoy_tree(model_file_name, vector_dims):
tree = AnnoyIndex(vector_dims)
tree.load(model_file_name)
return tree
