def find_edges(input, test, K):
print(f"building kNN classifier ... ", end=" ")
st_time = time.time()
if kNN_type <= 3:
input, test = input.todense(), test.todense()
if kNN_type == 1:
from sklearn.neighbors import NearestNeighbors
tree = NearestNeighbors(n_neighbors=K + 1, algorithm='ball_tree').fit(input)
elif kNN_type == 2:
from scipy import spatial
tree = spatial.KDTree(input)
elif kNN_type == 3:
from n2 import HnswIndex
tree = HnswIndex(input.shape[1], distance_type) # distance_type in ['angular', 'L2']
for index in tqdm(range(input.shape[0])):
tree.add_data(input[index, :])
tree.build(n_threads=10)
elif kNN_type == 4:
import pysparnn.cluster_index as ci
input_num = input.shape[0]
tree = ci.MultiClusterIndex(input, range(input_num))
else:
raise NotImplementedError
print(f"time={time.time()-st_time:.3f}s")
print("finding indices ... ", end=" ")
if kNN_type == 1:
_, indices = tree.kneighbors(test)
elif kNN_type == 2:
_, indices = tree.query(test, k=K + 1)
elif kNN_type == 3:
indices = []
for i in tqdm(range(test.shape[0])):
indices.append(tree.search_by_vector(test[i, :], k=K + 1))
else:
indices = tree.search(test, k=K+1, k_clusters=100, return_distance=False)
print(f"time={time.time()-st_time:.3f}s")
edge_list = []
for index1, per in enumerate(indices):
for index2 in per:
index2 = int(index2)
if index1 != index2:
edge_list.append((index1, index2))
print(f"done! .... time={time.time()-st_time:.3f}s")
return edge_list
class NNF():
def __init__(self,
image,
target_mask,
source_mask,
patch_size=(11, 11),
patch_weight=None,
num_neighbors=1):
im_h, im_w, im_ch = image.shape
if patch_weight is None:
self.patch_weight = np.ones(patch_size, dtype=_im_dtype)
self.patch_size = patch_size
self.num_neighb = num_neighbors
print("Build NNF index: ", end=" ")
start = time.time()
if _NN_algorithm != "PatchMatch":
self.source_ind = op.masked_indices(source_mask)
self.target_ind = op.masked_indices(target_mask)
# convert array indices to patch indices
pad = patch_size[0] // 2
ind_y, ind_x = np.divmod(self.source_ind, im_w)
self.source_ind = (ind_x - pad) + (ind_y - pad) * (im_w - 2 * pad)
source_point_cloud = extract_patches_2d( image, patch_size=patch_size )[self.source_ind].reshape((self.source_ind.size,-1)) \
* np.repeat(np.sqrt(self.patch_weight),im_ch)
# need this because of FLANN bug (?) with memory release
self.target_point_cloud = np.zeros(
(self.target_ind.size, source_point_cloud.shape[-1]),
dtype=_im_dtype)
if _NN_algorithm == "FLANN":
self.nn = flann.FLANN()
self.nn.build_index(source_point_cloud,
algorithm="kdtree",
trees=1) #, log_level = "info")
elif _NN_algorithm == "Sklearn":
self.nn = NearestNeighbors(
n_neighbors=num_neighbors,
algorithm='kd_tree',
metric='minkowski',
n_jobs=-1) #,metric_params={'w':self.patch_weight})
self.nn.fit(X=source_point_cloud)
elif _NN_algorithm == "FAISS":
self.nn = faiss.IndexHNSWFlat(source_point_cloud.shape[1], 50)
self.nn.add(source_point_cloud)
if _NN_algorithm == "PatchMatch":
self.nn = pm.PatchMatch(target_mask,
source_mask,
patch_size=patch_size,
lambdas=np.ones_like(image,
dtype=_im_dtype))
print('%f sec' % (time.time() - start))
def calculate_nnf(self, image, init_guess=None):
im_h, im_w, im_ch = image.shape
print("Query NNF index: ", end=" ")
start = time.time()
if _NN_algorithm != "PatchMatch":
ind_nnf = np.zeros((im_h * im_w, self.num_neighb), dtype='int32')
dist_nnf = np.zeros((im_h * im_w, self.num_neighb))
# convert array indices to patch indices
pad = self.patch_size[0] // 2
ind_y, ind_x = np.divmod(self.target_ind, im_w)
ind = (ind_x - pad) + (ind_y - pad) * (im_w - 2 * pad)
# need this because of FLANN bug (?) with memory release
np.copyto(self.target_point_cloud,
extract_patches_2d(
image, patch_size=self.patch_size)[ind].reshape(
(self.target_ind.size, -1)),
casting='same_kind',
where=True)
self.target_point_cloud *= np.repeat(np.sqrt(self.patch_weight),
im_ch)
# note that "ind" are patch indices, not array indices
if _NN_algorithm == "FLANN":
ind, dist = self.nn.nn_index(self.target_point_cloud,
self.num_neighb)
elif _NN_algorithm == "Sklearn":
dist, ind = self.nn.kneighbors(X=self.target_point_cloud,
return_distance=True)
elif _NN_algorithm == "FAISS":
dist, ind = self.nn.search(self.target_point_cloud,
self.num_neighb)
if _NN_algorithm != "PatchMatch":
ind = ind.reshape((ind.shape[0], self.num_neighb))
dist = dist.reshape((dist.shape[0], self.num_neighb))
# convert patch indices to array indices
ind = self.source_ind[ind.ravel()]
ind_y, ind_x = np.divmod(ind, im_w - 2 * pad)
ind = (ind_x + pad) + (ind_y + pad) * im_w
ind = np.reshape(ind, (-1, self.num_neighb))
for n in range(self.num_neighb):
ind_nnf[self.target_ind, :] = ind #[:,n]
dist_nnf[self.target_ind, :] = dist #[:,n]
elif _NN_algorithm == "PatchMatch":
ind_nnf, dist_nnf = self.nn.find_nnf(image, init_guess=init_guess)
print('%f sec' % (time.time() - start))
ind_nnf = ind_nnf.reshape((-1, self.num_neighb))
dist_nnf = dist_nnf.reshape((-1, self.num_neighb))
return ind_nnf, dist_nnf
def find_edges(input, test, K):
print(f"\tbuilding kNN classifier ... ", end=" ")
st_time = time.time()
if kNN_type in [1, 2]:
input, test = input.todense(), test.todense()
if kNN_type == 1:
from sklearn.neighbors import NearestNeighbors
tree = NearestNeighbors(n_neighbors=K + 1, algorithm='ball_tree').fit(input)
elif kNN_type == 2:
from scipy import spatial
tree = spatial.KDTree(input)
elif kNN_type == 3:
from n2 import HnswIndex
tree = HnswIndex(input.shape[1], distance_type) # distance_type in ['angular', 'L2']
for index in tqdm(range(input.shape[0])):
tree.add_data(input[index, :])
tree.build(n_threads=20)
elif kNN_type == 4:
import pysparnn.cluster_index as ci
input_num = input.shape[0]
tree = ci.MultiClusterIndex(input, range(input_num))
elif kNN_type == 5:
import nmslib
M, efC = 30, 100
index_time_params = {'M': M, 'indexThreadQty': num_threads, 'efConstruction': efC, 'post': 0}
space_names = ['l2_sparse', 'cosinesimil_sparse'] # https://github.com/nmslib/nmslib/blob/master/manual/spaces.md
space_name = space_names[0]
data_type = nmslib.DataType.SPARSE_VECTOR
tree = nmslib.init(method='hnsw', space=space_name, data_type=data_type)
'''
def calc_zero_rows(i):
if input[i, :].getnnz() == 0:
return 1
else:
return 0
pool = Pool(num_threads)
zero_row_num = sum(pool.map(calc_zero_rows, range(input.shape[0])))
print(f"# zero rows in input = {zero_row_num}", end=" ")
'''
tree.addDataPointBatch(input)
tree.createIndex(index_time_params, print_progress=True)
# Setting query-time parameters
efS = 100
query_time_params = {'efSearch': efS}
print('Setting query-time parameters', query_time_params, end=" ")
tree.setQueryTimeParams(query_time_params)
else:
raise NotImplementedError
print(f"time={time.time()-st_time:.3f}s")
print("\tfinding indices ... ", end=" ")
if kNN_type == 1:
_, indices = tree.kneighbors(test)
elif kNN_type == 2:
_, indices = tree.query(test, k=K + 1)
elif kNN_type == 3:
indices = []
for i in tqdm(range(test.shape[0])):
indices.append(tree.search_by_vector(test[i, :], k=K + 1))
elif kNN_type == 4:
indices = tree.search(test, k=K+1, k_clusters=100, return_distance=False)
elif kNN_type == 5:
'''
def calc_zero_rows2(i):
if test[i, :].getnnz() == 0:
return 1
else:
return 0
pool = Pool(num_threads)
zero_row_num = sum(pool.map(calc_zero_rows2, range(test.shape[0])))
print(f"# zero rows in test = {zero_row_num}")
'''
indices_ = tree.knnQueryBatch(test, k=K+1, num_threads=num_threads)
indices = [i[0] for i in indices_]
del indices_
else:
raise NotImplementedError
print(f"time={time.time()-st_time:.3f}s")
edge_list = []
for index1, per in enumerate(indices):
assert len(per) == K+1, f"index1={index1} len(per)={len(per)} != K={K}"
for index2 in per:
index2 = int(index2)
if index1 != index2:
edge_list.append((index1, index2))
print(f"\tget edges done! .... time={time.time()-st_time:.3f}s")
return edge_list
class CustomKMeans(object):
def __init__(
self,
n_centers: int = 2,
method: str = 'kmeans',
max_iter: int = 30,
max_iter_no_progress: int = 10,
tol_progress: np.float = 1e-3,
random_state: Optional[int] = 17,
nn_km_branching: int = 32, # branching for k-means tree
nn_km_iter: int = 20, # number of iterations per k-means step
nn_kd_trees: int = 32, # number of randomized trees to use
nn_checks: int = 75, # number of leaves to check in the search
nn_autotune: np.float = -1, # auto-tuning of nn parameters
apply_fix: bool = False,
save_log: bool = True,
gpu_idx: int = 0,
verbose: bool = False):
assert method in {'kmeans', 'kdtree', 'exact', 'exact-gpu'}
# FLANN hyperparameters
# The whole list of FLANN parameters is available here:
# https://github.com/mariusmuja/flann/blob/master/src/cpp/flann/flann.h
self.params = {
'kdtree': {
'algorithm': 'kdtree',
'num_neighbors': 1,
'trees': nn_kd_trees,
'checks': nn_checks,
'target_precision': nn_autotune
},
'kmeans': {
'algorithm': 'kmeans',
'num_neighbors': 1,
'branching': nn_km_branching,
'iterations': nn_km_iter,
'checks': nn_checks,
'target_precision': nn_autotune
},
'exact': {
'algorithm': 'exact'
},
'exact-gpu': {
'algorithm': 'exact-gpu'
}
}
# GPU parameters
self.gpu = None
self.gpu_idx = gpu_idx
# Nearest neighbors search method set up
if method in {'kmeans', 'kdtree'}:
self.nn_search = pyflann.FLANN()
elif method == 'exact':
self.nn_search = NearestNeighbors(n_neighbors=1,
algorithm='kd_tree',
leaf_size=nn_checks,
metric='minkowski',
p=2,
n_jobs=1)
else:
self.nn_search = None
self.gpu = faiss.StandardGpuResources()
self.n_centers = n_centers
self.method = method
self.max_iter = max_iter
self.max_iter_no_progress = max_iter_no_progress
self.tol_progress = tol_progress
self.random_state = random_state
self.verbose = verbose
self._dim = None
self._n_samples = None
self.centers_ = None
self.labels_ = None
self.sqdist_ = None
self.stats_ = None
self.session_id = f'n_centers_{self.n_centers}-' + \
'-'.join(
[f'{param}_{val}' for param, val
in self.params[self.method].items()]
)
self.apply_fix = apply_fix
self.log_ = []
self.save_log = save_log
def __reset_stats(self):
self.stats_ = {
'measure': [], # the cost function
'evaluation': [], # time to evaluate labels
'assignment': [], # time to re-assign labels
'n_centers': self.n_centers,
'apply_fix': self.apply_fix,
**self.params[self.method]
}
def __update_stats(self, measure: np.float, time_eval: np.float,
time_assign: np.float):
self.stats_['measure'].append(measure)
self.stats_['evaluation'].append(time_eval)
self.stats_['assignment'].append(time_assign)
def fit(self, data: np.ndarray) -> None:
data = data.astype(np.float32)
self._n_samples, self._dim = data.shape
np.random.seed(self.random_state)
# Initialize centers (we use uniform random initialization)
self.centers_ = np.random.uniform(np.min(data),
np.max(data),
size=self.n_centers *
self._dim).reshape(
self.n_centers,
self._dim).astype(data.dtype)
# Start fitting
self.__reset_stats()
self.log_.append(
f'Session id: {self.session_id} '
f'at {datetime.utcnow().strftime("%Y-%m-%d %H:%M:%S")}')
progress = self.max_iter_no_progress
for it in range(self.max_iter):
tic_it = time()
# Evaluate labels and squared distances
if self.method in {'kmeans', 'kdtree'}:
labels_, sqdist = self.nn_search.nn(self.centers_, data,
**self.params[self.method])
if self.apply_fix and it > 0:
reassigned = np.where(labels_ != self.labels_)[0]
if len(reassigned):
sqdist_checked = np.linalg.norm(
data[reassigned, :] -
self.centers_[self.labels_[reassigned], :],
axis=1)**2
correct = sqdist[reassigned] <= sqdist_checked.ravel()
incorrect = ~correct
to_assign = reassigned[correct]
to_leave = reassigned[incorrect]
if len(to_assign):
self.labels_[to_assign] = labels_[to_assign]
self.sqdist_ = sqdist
if len(to_leave):
self.sqdist_[to_leave] = sqdist_checked[incorrect]
else:
self.labels_, self.sqdist_ = labels_, sqdist
elif self.method == 'exact':
self.nn_search.fit(self.centers_)
self.sqdist_, self.labels_ = self.nn_search.kneighbors(
X=data, return_distance=True)
self.sqdist_, self.labels_ = \
self.sqdist_.ravel()**2, self.labels_.ravel()
else:
index_flat = faiss.IndexFlatL2(self._dim)
self.nn_search = faiss.index_cpu_to_gpu(
self.gpu, self.gpu_idx, index_flat)
self.nn_search.add(self.centers_)
self.sqdist_, self.labels_ = self.nn_search.search(data, 1)
self.sqdist_, self.labels_ = \
self.sqdist_.ravel(), self.labels_.ravel()
toc = time()
t1 = toc - tic_it
# Update centers
tic = time()
for label in range(self.n_centers):
idx = np.where(self.labels_ == label)[0]
if len(idx):
self.centers_[label] = data[idx].mean(axis=0)
else:
# If the cluster is empty, move its center anyway
self.centers_[label] = self.centers_.mean(axis=0)
toc_it = time()
t2 = toc_it - tic
# Print progress
self.log_.append(f'\t--> iteration {it} '
f'has been finished -- {toc_it - tic_it:.3e}s')
if self.verbose:
print(self.log_[-1])
# Check convergence
p = np.bincount(self.labels_, weights=self.sqdist_).sum()
p /= self._n_samples
self.__update_stats(p, t1, t2)
if it >= 1 and self.stats_['measure'][-2] - \
self.stats_['measure'][-1] < self.tol_progress:
progress -= 1
if not progress:
print(f'\nIteration {it}: '
f'no progress during '
f'last {self.max_iter_no_progress} iterations')
break
else:
progress = self.max_iter_no_progress
# Save the log
self.log_.append(self.time_report())
if self.save_log:
path = Path(f'../output/logs/'
f'{self.__class__.__name__}/').resolve()
path.mkdir(parents=True, exist_ok=True)
with open(f'{str(path)}/{self.session_id}.txt', 'w') as file:
for line in self.log_:
file.write(f'{line}\n')
def fit_predict(self, data: np.ndarray) -> np.ndarray:
self.fit(data)
return self.labels_
def time_report(self) -> str:
if self.stats_ is None:
rep = 'No statistics are available.'
else:
t_eval = np.asarray(self.stats_['evaluation'])
t_assign = np.asarray(self.stats_['assignment'])
rep = f'\nEvaluation time per iteration:\n' \
f'\tAVG. = {t_eval.mean()}s\n' \
f'\tSTD. = {t_eval.std()}s\n' \
f'Assignment time per iteration:\n' \
f'\tAVG. = {t_assign.mean()}s\n\tSTD. = {t_assign.std()}s'
return rep
def find_edges(input, test, K, cluster_ids, query_ids):
print(f"\tbuilding kNN classifier ... ", end=" ")
st_time = time.time()
if kNN_type in [1, 2]:
input, test = input.todense(), test.todense()
if kNN_type == 1:
from sklearn.neighbors import NearestNeighbors
tree = NearestNeighbors(n_neighbors=K + 1, algorithm='ball_tree').fit(input)
elif kNN_type == 2:
from scipy import spatial
tree = spatial.KDTree(input)
elif kNN_type == 3:
from n2 import HnswIndex
tree = HnswIndex(input.shape[1], distance_type) # distance_type in ['angular', 'L2']
for index in tqdm(range(input.shape[0])):
tree.add_data(input[index, :])
tree.build(n_threads=20)
elif kNN_type == 4:
import pysparnn.cluster_index as ci
input_num = input.shape[0]
tree = ci.MultiClusterIndex(input, range(input_num))
elif kNN_type == 5:
import nmslib
M, efC, num_threads = 30, 100, 10
index_time_params = {'M': M, 'indexThreadQty': num_threads, 'efConstruction': efC, 'post': 0}
space_name = 'cosinesimil_sparse'
data_type = nmslib.DataType.SPARSE_VECTOR
tree = nmslib.init(method='hnsw', space=space_name, data_type=data_type)
print(f"type(input) = {type(input)} type(test)={type(test)}", end=" ")
tree.addDataPointBatch(input)
tree.createIndex(index_time_params)
# Setting query-time parameters
efS = 100
query_time_params = {'efSearch': efS}
print('Setting query-time parameters', query_time_params)
tree.setQueryTimeParams(query_time_params)
else:
raise NotImplementedError
print(f"time={time.time()-st_time:.3f}s")
print("\tfinding indices ... ", end=" ")
if kNN_type == 1:
_, indices = tree.kneighbors(test)
elif kNN_type == 2:
_, indices = tree.query(test, k=K + 1)
elif kNN_type == 3:
indices = []
for i in tqdm(range(test.shape[0])):
indices.append(tree.search_by_vector(test[i, :], k=K + 1))
elif kNN_type == 4:
indices = tree.search(test, k=K+1, k_clusters=100, return_distance=False)
elif kNN_type == 5:
indices_ = tree.knnQueryBatch(test, k=K, num_threads=num_threads)
indices = [i[0] for i in indices_]
del indices_
else:
raise NotImplementedError
print(f"time={time.time()-st_time:.3f}s")
edge_list = []
for index1, per in enumerate(indices):
for index2 in per:
index2 = int(index2)
if index1 != index2:
edge_list.append((query_ids[index1], center_ids[index2]))
print(f"\tdone! .... time={time.time()-st_time:.3f}s")
return edge_list