def testUnsketch(self):
# make sure heavy hitter recovery works correctly
# use a gigantic sketch so there's no chance of collision
d = 5
c = 10000
r = 20
a = CSVec(d, c, r, **self.csvecArgs)
vec = torch.rand(d).to(self.device)
a += vec
with self.subTest(method="topk"):
recovered = a.unSketch(k=d)
self.assertTrue(torch.allclose(recovered, vec))
with self.subTest(method="epsilon"):
thr = vec.abs().min() * 0.9
recovered = a.unSketch(epsilon=thr / vec.norm())
self.assertTrue(torch.allclose(recovered, vec))
def testSketchSum(self):
d = 5
c = 10000
r = 20
summed = CSVec(d, c, r, **self.csvecArgs)
for i in range(d):
vec = torch.zeros(d).to(self.device)
vec[i] = 1
sketch = CSVec(d, c, r, **self.csvecArgs)
sketch += vec
summed += sketch
recovered = summed.unSketch(k=d)
trueSum = torch.ones(d).to(self.device)
self.assertTrue(torch.allclose(recovered, trueSum))
y = comm.recv(source=n, tag=11)
table_rx = np.array(table_rx) + np.array(y)
# Ramp up alpha
if epoch <= 5:
alpha_t = alpha + epoch * (0.3 - alpha) / 5
elif epoch > 5 and epoch <= 10:
alpha_t = 0.3 - (epoch - 5) * (0.3 - alpha) / 5
else:
alpha_t = alpha
table_rx = alpha * (1.0 / N) * table_rx
S.table = torch.tensor(table_rx)
S_e = S + S_e
unsketched = S_e.unSketch(k=k)
np_unsketched = unsketched.numpy()
S.zero()
S.accumulateVec(unsketched)
S_e.table = S_e.table - S.table
# Rehape unsketched
print(unsketched.shape)
shapes = [
model.trainable_weights[i].shape
for i in range(len(model.trainable_weights))
]
grad_tx = []
topHCSAcc = np.zeros((len(hVals), len(kVals)))
csvecL2 = np.zeros(len(kVals))
topHCSL2 = np.zeros((len(hVals), len(kVals)))
for k_i, k in enumerate(kVals):
d, c, r, numBlocks = len(summed), cols, 15, 30
#ipdb.set_trace()
expected = torch.zeros(len(summed), device=device)
expected[expectedIndices[-k:].to(device)] = summed[
expectedIndices[-k:].to(device)]
assert (summed.size() == vecs[0].size())
w_0 = CSVec(d=d, c=c, r=r, numBlocks=numBlocks, device=device)
w_0 += vecs[0]
print("worker added")
result = w_0.unSketch(k)
csvecAcc[k_i] = (expected[expectedIndices] *
result[expectedIndices]).nonzero().numel() / k
print("k = {}".format(k))
csvecL2[k_i] = (torch.sum((result - expected)**2))**0.5
for h_i, h in enumerate(hVals):
result = torch.zeros(len(summed), device=device)
w_0 = TopHCS(d=d,
c=c,
r=r,
h=h,
numBlocks=numBlocks,
device=device)
w_0.store(vecs[0])
print("worker added")
workers = [w_0]
