Problem: precision of faiss output

[2012fang]

vector1 = array([[-8.44551474e-02, 4.56927270e-02, -1.23174498e-02,
-3.28429081e-02, -7.21267983e-02, 1.78922843e-02,
5.72474487e-02, 1.16973273e-01, -2.67414916e-02,
-2.48716399e-02, 7.50321448e-02, -8.46290886e-02,
7.03940690e-02, 1.73218511e-02, -9.68408026e-03,
5.39224297e-02, -2.21001990e-02, -6.22086227e-02,
-5.71195409e-03, -4.89818901e-02, -2.62292381e-02,
-7.56375724e-04, -3.10133863e-03, 2.50932369e-02,
-2.45238114e-02, 2.69615464e-03, 1.92056168e-02,
-3.85982543e-02, 3.91353038e-04, 2.78586075e-02,
3.11407298e-02, 3.98177765e-02, 1.26895290e-02,
4.59183939e-02, -1.79051757e-02, -5.52790565e-03,
6.54223328e-03, 8.18098924e-05, -3.61517854e-02,
3.96922370e-03, 1.84885133e-02, 7.61720911e-02,
-1.71365440e-02, 2.89176721e-02, 1.41874431e-02,
2.79633328e-02, -9.05210152e-03, -5.40017057e-03,
1.52942166e-02, -4.57295403e-02, -1.27692312e-01,
2.91015562e-02, 3.44425137e-03, -1.32864729e-01,
1.61509942e-02, -4.71395580e-03, -3.31216329e-03,
-5.12617938e-02, -1.18143307e-02, 4.09921259e-03,
-1.69061478e-02, -6.64441613e-03, -1.79887109e-04,
3.39703285e-03, 2.31174359e-04, -6.80163456e-03,
-4.33840603e-02, 1.65534336e-02, -1.14529179e-02,
-5.13268374e-02, 8.57468881e-03, -4.03255783e-02,
1.27376884e-01, 2.95607466e-02, -7.07077840e-03,
6.40317500e-02, -4.19284543e-03, -1.87955312e-02,
3.14377132e-03, 3.27055436e-03, -2.84685977e-02,
3.05155981e-02, -1.38661405e-02, 2.08585709e-02,
-8.18277225e-02, 7.66183436e-03, -8.52554291e-03,
-4.12235968e-03, -2.16198359e-02, 8.11005197e-03,
5.65368915e-03, 7.68653117e-03, 1.53549192e-02,
-2.66802274e-02, 6.90017221e-03, -1.03461863e-02,
-3.84186208e-02, 7.67239253e-04, -7.69756362e-02,
-2.08837376e-03, 1.70545429e-01, -3.28036658e-02,
-1.23308171e-02, 2.23817527e-02, 2.46881787e-02,
-9.55710560e-03, -1.12988511e-02, 4.77782590e-03,
2.15060432e-02, -8.62445123e-03, -4.57427930e-03,
-9.77519713e-03, -8.32691789e-03, -1.17511367e-02,
8.34971201e-03, -8.32115579e-03, -1.06792047e-03,
4.26274911e-03, 2.91057304e-03, -2.71207988e-02,
1.73099246e-02, 5.73569834e-02, -1.06890816e-02,
5.14503904e-02, 5.71261998e-03, -6.38413336e-03,
2.71979533e-03, -2.98482575e-03, -2.09113173e-02,
-3.34538966e-02, 1.03973553e-01, 1.61792163e-03,
-2.10588938e-03, 1.16627095e-02, -4.08083200e-02,
-2.15507522e-02, -3.02650742e-02, -2.24047992e-02,
-5.13026938e-02, 3.36436518e-02, -1.44781824e-02,
1.46059990e-02, 1.78493012e-03, 1.40046664e-02,
1.03832912e-02, 5.89301228e-04, -8.61266702e-02,
2.70281229e-02, 1.32350363e-02, 2.37457342e-02,
6.16388721e-03, 1.34771280e-02, -3.71912145e-03,
5.50323725e-03, 1.83100328e-02, -1.15995295e-02,
4.94935736e-03, -3.10285240e-02, 2.88795643e-02,
2.23691966e-02, -5.46695804e-03, -1.25807934e-02,
1.06846122e-03, -9.27604176e-03, 2.31033235e-04,
-1.27426907e-02, -1.49435410e-02, -3.63811618e-04,
4.62528929e-04, 1.41673416e-01, -2.24193204e-02,
2.65473872e-03, 1.73668750e-02, -8.37226305e-03,
-7.68812199e-04, -5.59632443e-02, 1.76156759e-02,
-1.90522298e-02, 1.30176125e-02, -2.61259405e-03,
2.12291051e-02, -5.47764264e-03, 1.14389755e-01,
-1.40980510e-02, -4.03918093e-03, 1.17943855e-02,
2.88138464e-02, -8.01990330e-02, 5.73012047e-03,
1.15558198e-02, 8.47641262e-04, -1.48011073e-02,
-5.33317542e-03, 5.24302188e-04, 2.74261590e-02,
1.25564018e-03, 3.57892131e-04, 7.67934648e-03,
3.72179062e-03, -3.13465036e-02, -2.22872272e-02,
1.36735262e-02, -5.66293187e-02, -4.33784463e-02,
-1.10101821e-02, 1.83216650e-02, -1.62522821e-03,
5.90732263e-04, 1.58377308e-02, -3.08983680e-03,
-1.30449077e-02, 5.49028115e-03, 1.46150414e-03,
-3.28909699e-03, -2.34518554e-02, -2.49394365e-02,
3.35218087e-02, -3.72543884e-03, 8.32281914e-03,
1.74827911e-02, 1.36323106e-02, 1.96427316e-03,
-1.54801579e-02, 2.18579508e-02, 1.53032644e-02,
-1.24354541e-01, 4.61954763e-03, 6.42319210e-03,
-1.99424778e-03, 2.53712293e-02, -2.28092237e-03,
8.56701657e-03, -8.70297104e-03, 4.91529889e-03,
2.96094920e-02, -1.40514761e-01, -1.64863258e-03,
1.46121383e-02, -1.70896668e-02, -1.08292485e-02,
-1.93554610e-02, -9.06635891e-04, -5.14678983e-03,
8.32778960e-03, -3.03074601e-03, 5.09683415e-03,
2.29578353e-02, -2.38610152e-03, -3.95761512e-04,
-1.23001286e-03, -6.09071506e-03, -6.79963641e-03,
-1.66030917e-02, 5.24643157e-03, -2.21087039e-02,
-1.45360304e-03, 7.90787116e-03, -1.70765687e-02,
3.24698351e-03, 3.14700068e-03, 6.82703545e-03,
6.86773192e-03, -3.06519661e-02, -1.18442383e-02,
-9.98975523e-03, -1.33655027e-01, -3.16358381e-03,
1.30502349e-02, -3.98504548e-03, 8.38416163e-03,
-1.52457866e-03, -1.47919785e-02, -1.19871600e-02,
2.25242693e-02, -7.72700063e-04, 1.61835514e-02,
-3.06999348e-02, 4.59993165e-03, -1.12461997e-02,
-6.39913836e-03, -9.07016173e-03, 1.04761925e-02,
2.33298331e-03, 1.48088823e-03, 3.24343471e-03,
-7.87130930e-03, -1.27785711e-03, 2.93481238e-02,
1.79333263e-03, -4.18773573e-03, -2.99077109e-03,
-4.42271261e-03, -9.84390359e-03, -1.00352280e-02,
-4.04074416e-03, -9.34695639e-03, 1.35453782e-04,
1.59593262e-02, 6.48296159e-03, -1.59481931e-02,
1.58800000e+03]], dtype=float32)

array([-3.52100320e-02, 3.23681682e-02, 4.95685935e-02, 2.25345828e-02,
-2.19355747e-02, 3.47894281e-02, 1.88222472e-02, -2.67079379e-02,
2.93516293e-02, 2.51668319e-02, -2.38688551e-02, 4.35155258e-02,
1.52722495e-02, 1.03030205e-02, 6.67062914e-03, -1.76866446e-02,
-1.97095834e-02, -1.20737469e-02, 1.72977033e-03, 1.84225533e-02,
8.42243433e-03, -9.13981907e-03, -4.81031649e-02, 4.02168371e-03,
1.59920063e-02, -3.62270512e-02, -3.90865281e-02, 1.65622961e-02,
6.56311307e-03, -4.22318000e-03, 5.76240476e-03, -9.56683885e-03,
-1.00857817e-01, -2.68988730e-03, -1.93930697e-02, 7.16141844e-03,
2.22620480e-02, 3.36622037e-02, -1.15636700e-04, -1.84167828e-02,
-4.98541743e-02, -1.08579248e-02, 5.92702953e-03, 1.49204656e-02,
-2.58545415e-03, 4.85718600e-04, 1.40515016e-02, 1.60455033e-02,
-5.20697981e-03, -6.94154855e-03, -5.77796390e-03, 6.35442790e-03,
-3.31752072e-03, -9.27312002e-02, 4.30252589e-03, -5.70000196e-03,
-1.40949450e-02, -9.79238469e-03, -4.61103162e-03, 5.95919183e-03,
-1.21632749e-02, -2.81034783e-03, -1.30199082e-02, -1.80452187e-02,
8.07777699e-03, 3.62024433e-03, -9.82603058e-03, 2.10663397e-02,
-2.22302061e-02, 4.63656615e-03, -6.54944265e-03, -6.16694689e-02,
5.04274257e-02, -2.02144403e-03, 6.51451899e-03, -6.59902347e-04,
8.13715812e-03, 5.53054269e-04, 1.91264637e-02, 8.31208937e-03,
-5.69219403e-02, -1.88740622e-02, 1.34680541e-02, -1.88715551e-02,
-8.75640661e-02, -9.99674015e-03, -7.21022440e-03, -9.92303900e-03,
8.10161047e-03, 3.18781212e-02, -1.60904031e-03, 1.19031547e-02,
-1.93211790e-02, -5.63503662e-03, 1.11618862e-02, 1.78706311e-02,
-1.40397782e-02, -6.33524777e-03, 2.45600231e-02, -9.09450185e-03,
8.00114498e-02, -1.33760469e-02, 3.82367289e-03, 1.21191954e-02,
2.95079383e-03, 1.11511629e-02, -2.26596161e-03, -2.27466263e-02,
1.96284917e-03, 9.57483612e-03, -5.43116173e-03, -5.90029582e-02,
4.02573869e-03, 1.08914718e-03, -9.73853189e-03, -1.54579254e-02,
2.85442569e-03, -1.38227968e-02, 1.03292111e-02, -1.40649490e-02,
-1.08934157e-02, 7.44717792e-02, -1.53173171e-02, -1.34716146e-02,
6.72431313e-04, -1.31279239e-02, 1.95410103e-02, 8.77668557e-04,
-4.46675252e-03, -3.59900296e-02, -7.96748400e-02, -2.24557682e-03,
6.82948623e-03, 9.16492473e-03, 7.64212804e-04, 2.81372835e-04,
-2.21718214e-02, -7.32463878e-03, -3.13916639e-03, 1.40989078e-02,
8.69531592e-04, 6.18395628e-03, 3.09359375e-03, -1.62273063e-03,
-1.81020529e-04, -8.85605440e-03, -3.52023430e-02, 3.07588354e-02,
8.79613275e-04, 1.79869942e-02, 1.46495244e-02, -4.85088257e-03,
-3.63417645e-03, -1.43184152e-03, 3.72584537e-03, 1.49723524e-02,
1.18230851e-02, -5.99940633e-03, 1.20357645e-03, 4.32597334e-03,
-7.08900671e-03, -2.87136855e-03, -1.31057734e-02, 6.80287369e-03,
7.64887081e-04, -9.03168973e-03, -2.98980926e-03, -1.44707412e-02,
-8.48157724e-05, 4.62485813e-02, -8.42207205e-03, -1.21583730e-01,
-1.69677788e-03, -4.84203594e-03, -8.78408272e-03, -8.18974972e-02,
4.64980351e-03, -2.19167992e-02, -3.08722816e-03, -1.09999226e-02,
1.16145927e-02, -1.69398058e-02, 6.43228590e-02, 9.89711378e-03,
-1.94506701e-02, 1.65792853e-02, 4.67116339e-03, -5.69736473e-02,
-8.37437809e-03, -1.12876680e-03, 4.21729870e-03, 5.83686214e-03,
6.38038735e-04, 1.59960389e-02, 1.09745301e-02, 3.83243605e-04,
-8.73216465e-02, 3.84991872e-03, 2.24868301e-03, 8.21910799e-03,
1.75233446e-02, 3.22782376e-04, -1.21828035e-01, -5.73556870e-03,
-3.35951481e-04, 1.52382662e-03, -8.35970696e-03, 1.27844866e-02,
-4.76959301e-03, -2.53992574e-03, -1.60105270e-03, 2.01589130e-02,
-5.07437531e-03, 1.79226638e-03, -4.53328528e-03, 5.67664276e-04,
-2.51616351e-03, 1.22069418e-02, 8.47196672e-03, 1.02453716e-02,
5.38435159e-03, 2.99520907e-03, 2.78073899e-03, 1.76977459e-02,
1.48236239e-02, -1.45528376e-01, 3.87851335e-03, 1.42852003e-02,
2.61840597e-03, 1.18272370e-02, 4.32929164e-03, 1.07074659e-02,
-7.58202455e-04, 4.96439310e-03, 3.70181799e-02, -1.38960540e-01,
-7.00370781e-03, -1.27403131e-02, -5.37718879e-03, -1.02470759e-02,
-6.06298028e-03, -1.00048454e-02, -5.11866109e-03, 6.08263817e-03,
4.24736086e-03, -4.73625259e-03, 2.11990962e-04, 2.56018760e-03,
9.66703147e-03, -8.41557980e-03, -3.24191572e-03, -8.95409007e-03,
-1.19080013e-02, 9.59675666e-03, -4.40449733e-03, 6.89560198e-04,
-6.83376752e-03, -6.80298870e-03, 7.92070385e-03, -1.02428757e-02,
-3.82725964e-03, 1.28127681e-02, -2.40804814e-03, 4.09662817e-03,
-5.56241907e-03, -1.17346026e-01, -8.00196081e-03, 1.29663874e-03,
1.10061625e-02, -1.17006870e-02, -1.61563121e-02, -3.79574456e-04,
-9.67975520e-03, -2.44614016e-03, 2.91191484e-03, -1.25522225e-03,
5.71437413e-03, 3.36199673e-03, -1.04897069e-02, 6.90355338e-03,
-5.41149639e-03, -1.26767845e-03, 1.24809723e-02, 1.24099327e-03,
3.94557090e-03, -1.66563899e-03, -7.73626426e-03, 8.27431679e-03,
1.45018538e-02, -5.92773024e-04, 4.30521322e-03, -8.55698436e-03,
-9.26231686e-03, 1.06858984e-02, -2.15699635e-02, -9.08072386e-03,
-2.00625211e-02, -6.22118916e-03, -5.33714192e-03, 6.78832270e-03,
1.58700000e+03], dtype=float32)

python code:
vector2 = vector2.reshape((1,301))
vector1 = vector1.reshape((1,301))
index = faiss.IndexFlatL2(301)
index.add(vector2)
diss, inds = index.search(vector1,1)

when I run the faiss code, the result is as follows：
inds,diss = (array([[0]]), array([[1.]], dtype=float32))

When I run other Python code[np.sum(np.square(vector1-vector2))]
the result is :1.3264003

So, what I want to ask is why the output of the faiss code omits the decimal places

[mdouze]

This is because the distance computation is decomposed as described here: https://github.com/facebookresearch/faiss/wiki/Implementation-notes#matrix-multiplication-to-do-many-l2-distance-computations

So the inaccuracy is triggered by the fact that:

• there is one component of the vector that is much larger than the others (last component is 1580, the others are around 1e-2)

• Faiss uses the "matrix computation" trick because the dimension is not a multiple of 4 (this is an obsolete check that I will remove)

The fix to improve the numerical stability is to subtract the mean of the vectors. For example, subtracting 1500 from the last component fixes the discrepancy, see #297

[mdouze]

The enhancement is to remove the obsolete check. Done now in FB internal version.

[mdouze]

Will be shipped in 1.6.2.