def __init__(self):
self.loader = EmbeddingLoader()
self.svdK = 0
class KNN():
def __init__(self):
self.loader = EmbeddingLoader()
self.svdK = 0
def setEmbeddingDir(self, dirname):
self.dirname = dirname
def setLimitArray(self, limit):
self.limit = limit
def setK(self, k):
self.k = k
def setSVDk(self, k):
self.svdK = k
def setNoCache(self, no_cache):
self.no_cache = no_cache
def initialize(self):
self.emb = self._loadEmbedding(self.dirname)
# Normalize vectors for cosine similarity calculation
from sklearn.preprocessing import Normalizer
self.emb.x = Normalizer().fit_transform(self.emb.x)
def _loadEmbedding(self, dirname):
return self.loader.load(dirname, svd_k=self.svdK)
def getNeighborsFor(self, funcId):
"""
Retrieve k nearest neighbors of funcId, possibly
limited to data points in self.limit
"""
# This is VERY inefficient: We calculate entire
# distance matrices although all we need is the
# distance from one data point to all others
if self.limit:
validNeighborIds = [funcId] + [x for x in self.limit if x != funcId]
validNeighbors = [self.emb.rTOC[str(x)] for x in validNeighborIds]
X = self.emb.x[validNeighbors, :]
D = 1.0 - (X * self.emb.x[0, :].T).todense()
NNI = list(D[:,0].argsort(axis=0))[:self.k]
return [validNeighborIds[x] for x in NNI]
else:
dataPointIndex = self.emb.rTOC[funcId]
X = self.emb.x
D = 1.0 - (X * self.emb.x[dataPointIndex, :].T).todense()
NNI = list(D[:,0].argsort(axis=0))[:self.k]
return [self.emb.TOC[x] for x in NNI]
def calculateDistances(self):
self.emb.D = self._calculateDistanceMatrix()
self._calculateNearestNeighbors()
def _calculateNearestNeighbors(self):
self.emb.NNI = self.emb.D.argsort(axis=0)
def _calculateDistanceMatrix(self):
return pairwise_distances(self.emb.x, metric='cosine')
