def __init__(self, hash_size, input_dim, random_sampling = True, dict_type = 'int32', cuda_client_type = 'local', cuda_server = 'locahost', random_dims = 32, num_hashtables=1, storage_config=None, matrices_filename=None, overwrite=False):
self.hash_size = hash_size
self.input_dim = input_dim
self.num_hashtables = num_hashtables
if storage_config is None:
storage_config = {'dict': None}
self.storage_config = {storage_config: {}}
if storage_config == 'random':
self.storage_config = {'random': {'r': random_dims, 'dim': hash_size, 'random': random_sampling, 't': dict_type}}
if matrices_filename and not matrices_filename.endswith('.npz'):
raise ValueError("The specified file name must end with .npz")
self.matrices_filename = matrices_filename
self.overwrite = overwrite
self._init_uniform_planes()
self._init_hashtables()
self.loaded_keys = None
#self.cuda_hamming = CudaHamming()
cudaclient_options = {'host': cuda_server, 'port': 8080}
self.cuda_hamming = cudaclient(cuda_client_type, cudaclient_options)
def main():
args = parse_parameters()
(lsh, np_feature_vecs) = init_lsh(args)
lsh = run(args, lsh)
ground_truth = None
if args.gt != None:
(ground_truth, ground_truth_num) = load_ground_truth(args.gt, args.gtf)
if args.c != 'y' and args.i != 'y' and args.e != None and args.s == 'random':
client = cudaclient('net', {'host': args.host, 'port': 8080})
b_begin = int(args.b)
b_end = int(args.b) + 1
# when given a range of expanding levels
if args.b_begin != -1 and args.b_end != -1:
b_begin = int(args.b_begin)
b_end = int(args.b_end)
for cur_expand_level in range(b_begin, b_end):
client.send_query(['reset'])
client.send_query([args.title])
client.send_query(['expand_level: ' + str(cur_expand_level)])
total_found = {'10': 0, '100': 0}
for feature_idx in range(0, np_feature_vecs.shape[0]):
feature = np_feature_vecs[feature_idx]
if args.p != 'y':
retrived = lsh.query(feature, num_results = int(args.k), expand_level = cur_expand_level, distance_func = 'hamming')
else:
retrived = lsh.query_in_compressed_domain(feature, num_results = int(args.k), expand_level = cur_expand_level, distance_func = 'hamming', gpu_mode = args.g, vlq_mode = args.l)
total_found['10'] += cal_recall(retrived, ground_truth, feature_idx, int(args.k), topGT = 10)
total_found['100'] += cal_recall(retrived, ground_truth, feature_idx, int(args.k), topGT = 100)
recall_r = total_found['10'] / float(np_feature_vecs.shape[0])
recall_r_str = "recall@" + args.k + " GT@10: " + str(recall_r)
print recall_r_str
client.send_query([recall_r_str])
recall_r = total_found['100'] / float(np_feature_vecs.shape[0])
recall_r_str = "recall@" + args.k + " GT@100: " + str(recall_r)
print recall_r_str
client.send_query([recall_r_str])
def __init__(self,
hash_size,
input_dim,
random_sampling=True,
dict_type='int32',
cuda_client_type='local',
cuda_server='locahost',
random_dims=32,
num_hashtables=1,
storage_config=None,
matrices_filename=None,
overwrite=False):
self.hash_size = hash_size
self.input_dim = input_dim
self.num_hashtables = num_hashtables
if storage_config is None:
storage_config = {'dict': None}
self.storage_config = {storage_config: {}}
if storage_config == 'random':
self.storage_config = {
'random': {
'r': random_dims,
'dim': hash_size,
'random': random_sampling,
't': dict_type
}
}
if matrices_filename and not matrices_filename.endswith('.npz'):
raise ValueError("The specified file name must end with .npz")
self.matrices_filename = matrices_filename
self.overwrite = overwrite
self._init_uniform_planes()
self._init_hashtables()
self.loaded_keys = None
#self.cuda_hamming = CudaHamming()
cudaclient_options = {'host': cuda_server, 'port': 8080}
self.cuda_hamming = cudaclient(cuda_client_type, cudaclient_options)
def main():
args = parse_parameters()
(lsh, np_feature_vecs) = init_lsh(args)
lsh = run(args, lsh)
ground_truth = None
if args.gt != None:
(ground_truth, ground_truth_num) = load_ground_truth(args.gt, args.gtf)
if args.c != 'y' and args.i != 'y' and args.e != None and args.s == 'random':
client = cudaclient('net', {'host': args.host, 'port': 8080})
b_begin = int(args.b)
b_end = int(args.b) + 1
# when given a range of expanding levels
if args.b_begin != -1 and args.b_end != -1:
b_begin = int(args.b_begin)
b_end = int(args.b_end)
for cur_expand_level in range(b_begin, b_end):
client.send_query(['reset'])
client.send_query([args.title])
client.send_query(['expand_level: ' + str(cur_expand_level)])
total_found = {'10': 0, '100': 0}
for feature_idx in range(0, np_feature_vecs.shape[0]):
feature = np_feature_vecs[feature_idx]
if args.p != 'y':
retrived = lsh.query(feature,
num_results=int(args.k),
expand_level=cur_expand_level,
distance_func='hamming')
else:
retrived = lsh.query_in_compressed_domain(
feature,
num_results=int(args.k),
expand_level=cur_expand_level,
distance_func='hamming',
gpu_mode=args.g,
vlq_mode=args.l)
total_found['10'] += cal_recall(retrived,
ground_truth,
feature_idx,
int(args.k),
topGT=10)
total_found['100'] += cal_recall(retrived,
ground_truth,
feature_idx,
int(args.k),
topGT=100)
recall_r = total_found['10'] / float(np_feature_vecs.shape[0])
recall_r_str = "recall@" + args.k + " GT@10: " + str(recall_r)
print recall_r_str
client.send_query([recall_r_str])
recall_r = total_found['100'] / float(np_feature_vecs.shape[0])
recall_r_str = "recall@" + args.k + " GT@100: " + str(recall_r)
print recall_r_str
client.send_query([recall_r_str])