def testTopKSum(self):
d = 10
c = 10000
r = 20
h = d
a = TopHCS(h, d, c, r, **self.csvecArgs)
b = TopHCS(h, d, c, r, **self.csvecArgs)
zerosHalf = torch.zeros(d//2, dtype=torch.float, device=self.device)
vec = torch.cat((torch.randn(d//2, device=self.device), zerosHalf), 0)
vec2 = torch.cat((zerosHalf, torch.randn(d//2, device=self.device)), 0)
a.store(vec)
b.store(vec2)
result = TopHCS.topKSum([a, b], d)
expected = vec + vec2
self.assertTrue(torch.equal(expected, result))
def testTopKSum2(self):
d = 10
c = 10000
r = 20
h = d
a = TopHCS(h, d, c, r, **self.csvecArgs)
b = TopHCS(h, d, c, r, **self.csvecArgs)
c = TopHCS(h, d, c, r, **self.csvecArgs)
vec = torch.randn(d, device=self.device)
vec2 = torch.randn(d, device=self.device)
vec3 = torch.randn(d, device=self.device)
a.store(vec)
b.store(vec2)
c.store(vec3)
result = TopHCS.topKSum([a, b, c], d)
expected = vec + vec2 + vec3
self.assertTrue(torch.equal(expected, result))
w_2 = TopHCS(d=d,
c=c,
r=r,
h=h_curr,
numBlocks=numBlocks,
device=device)
w_2.store(vecs[2])
print(w_0.topH)
print("")
w_3 = TopHCS(d=d,
c=c,
r=r,
h=h_curr,
numBlocks=numBlocks,
device=device)
w_3.store(vecs[3])
print(w_0.topH)
print("")
workers = [w_0, w_1, w_2, w_3]
result = TopHCS.topKSum(workers, k)
#topHCS_accuracy = (expected == result).nonzero().numel() / k
#print("num of non 0 elements in topH result : ", result[expectedIndices].nonzero().numel())
topHCS_accuracy[
h_i, k_i] = (expected[expectedIndices] *
result[expectedIndices]).nonzero().numel() / k
print('result for h =', h_curr, " :", result)
print('topHCS accuracy :', '\n', topHCS_accuracy)
for h_i, hVal in enumerate(hVals):
h = int(k * hVal)
workers = [] #reusing
for vec in vecs:
w = TopHCS(h=h,
d=d,
c=c,
r=r,
numBlocks=numBlocks,
device=device)
w.store(vec)
workers.append(w)
assert (len(workers) == len(vecs))
recovered_2 = TopHCS.topKSum(workers, k)
indexAcc_2[
h_i,
k_i] = (expected[expectedIndices] *
recovered_2[expectedIndices]).nonzero().numel() / k
diff = recovered_2[recovered_2 != 0] - expected[recovered_2 != 0]
L1_2[h_i, k_i] = torch.median(torch.abs(diff))
L2_2[h_i, k_i] = torch.median(diff**2)
#indexAcc = (expected == recovered).nonzero().numel() / k
# Use above only when k == d
#print("\n", "k = %r; index accuracy = %r" % (k, indexAcc))
colors = ["#A30CE8", "#FF0000", "#E8710C", "#FFD20D"]