def __getitem__(self, index):
h = None
if self.cache is None or index not in self.cache:
h = gh.djikstra_wrapper((self.graph, [index]))
if self.cache is not None:
self.cache[index] = h
#logging.info(f"info {index}")
else:
h = self.cache[index]
#logging.info(f"hit {index}")
idx = torch.tensor([(index, j) for j in range(self.n) if j != index],
dtype=torch.long)
v = torch.DoubleTensor(h).view(-1)[idx[:, 1]]
return (idx, v)
def __getitem__(self, index):
if index not in self.cache:
if self.verbose: logging.info(f"Cache miss for {index}")
h = gh.djikstra_wrapper(
(self.graph,
[index]))[0, :] if self.Z is None else self.Z[index, :]
# add in all the edges
cur = 0
self.idx_cache[index, :, 0] = index
neighbors = scipy.sparse.find(self.graph[index, :])[1]
for e in neighbors:
self.idx_cache[index, cur, 1] = int(e)
self.val_cache[index, cur] = self.scale * h[e]
self.w_cache[index, cur] = self.weight_fn(1.0)
cur += 1
if cur >= self.nbr_frac * self.subsample: break
scratch = np.array(range(self.n))
np.random.shuffle(scratch)
i = 0
while cur < self.subsample and i < self.n:
v = scratch[i]
if v != index and v not in neighbors:
self.idx_cache[index, cur, 1] = int(v)
self.val_cache[index, cur] = self.scale * h[v]
# self.val_cache[index,cur] = 0
self.w_cache[index, cur] = self.weight_fn(h[v])
cur += 1
i += 1
if self.verbose:
logging.info(
f"\t neighbors={neighbors} {self.idx_cache[index,:,1].numpy().T}"
)
self.cache.add(index)
self.n_cached += 1
# if self.n_cached % (max(self.n//20,1)) == 0: logging.info(f"\t Cached {self.n_cached} of {self.n}")
# print("GraphRowSubSampler: idx shape ", self.idx_cache[index,:].size())
return (self.idx_cache[index, :], self.val_cache[index, :],
self.w_cache[index, :])