faiss/benchs/kmeans_mnist.py
Matthijs Douze 6d0bc58db6 Implementation of PQ4 search with SIMD instructions (#1542)
Summary:
IndexPQ and IndexIVFPQ implementations with AVX shuffle instructions.

The training and computing of the codes does not change wrt. the original PQ versions but the code layout is "packed" so that it can be used efficiently by the SIMD computation kernels.

The main changes are:

- new IndexPQFastScan and IndexIVFPQFastScan objects

- simdib.h for an abstraction above the AVX2 intrinsics

- BlockInvertedLists for invlists that are 32-byte aligned and where codes are not sequential

- pq4_fast_scan.h/.cpp:  for packing codes and look-up tables + optmized distance comptuation kernels

- simd_result_hander.h: SIMD version of result collection in heaps / reservoirs

Misc changes:

- added contrib.inspect_tools to access fields in C++ objects

- moved .h and .cpp code for inverted lists to an invlists/ subdirectory, and made a .h/.cpp for InvertedListsIOHook

- added a new inverted lists type with 32-byte aligned codes (for consumption by SIMD)

- moved Windows-specific intrinsics to platfrom_macros.h

Pull Request resolved: https://github.com/facebookresearch/faiss/pull/1542

Test Plan:
```
buck test mode/opt  -j 4  //faiss/tests/:test_fast_scan_ivf //faiss/tests/:test_fast_scan
buck test mode/opt  //faiss/manifold/...
```

Reviewed By: wickedfoo

Differential Revision: D25175439

Pulled By: mdouze

fbshipit-source-id: ad1a40c0df8c10f4b364bdec7172e43d71b56c34
2020-12-03 10:06:38 -08:00

90 lines
2.1 KiB
Python

#! /usr/bin/env python2
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import print_function
import numpy as np
import time
import faiss
import sys
# Get command-line arguments
k = int(sys.argv[1])
ngpu = int(sys.argv[2])
# Load Leon's file format
def load_mnist(fname):
print("load", fname)
f = open(fname)
header = np.fromfile(f, dtype='int8', count=4*4)
header = header.reshape(4, 4)[:, ::-1].copy().view('int32')
print(header)
nim, xd, yd = [int(x) for x in header[1:]]
data = np.fromfile(f, count=nim * xd * yd,
dtype='uint8')
print(data.shape, nim, xd, yd)
data = data.reshape(nim, xd, yd)
return data
basedir = "/path/to/mnist/data"
x = load_mnist(basedir + 'mnist8m/mnist8m-patterns-idx3-ubyte')
print("reshape")
x = x.reshape(x.shape[0], -1).astype('float32')
def train_kmeans(x, k, ngpu):
"Runs kmeans on one or several GPUs"
d = x.shape[1]
clus = faiss.Clustering(d, k)
clus.verbose = True
clus.niter = 20
# otherwise the kmeans implementation sub-samples the training set
clus.max_points_per_centroid = 10000000
res = [faiss.StandardGpuResources() for i in range(ngpu)]
flat_config = []
for i in range(ngpu):
cfg = faiss.GpuIndexFlatConfig()
cfg.useFloat16 = False
cfg.device = i
flat_config.append(cfg)
if ngpu == 1:
index = faiss.GpuIndexFlatL2(res[0], d, flat_config[0])
else:
indexes = [faiss.GpuIndexFlatL2(res[i], d, flat_config[i])
for i in range(ngpu)]
index = faiss.IndexReplicas()
for sub_index in indexes:
index.addIndex(sub_index)
# perform the training
clus.train(x, index)
centroids = faiss.vector_float_to_array(clus.centroids)
obj = faiss.vector_float_to_array(clus.obj)
print("final objective: %.4g" % obj[-1])
return centroids.reshape(k, d)
print("run")
t0 = time.time()
train_kmeans(x, k, ngpu)
t1 = time.time()
print("total runtime: %.3f s" % (t1 - t0))