harsha-simhadri#1 T1 FAISS customized for own quantized dataset

benchmark/datasets.py

@@ -79,8 +79,17 @@ def ivecs_read(fname):
 
 def xbin_mmap(fname, dtype, maxn=-1):
  """ mmap the competition file format for a given type of items """
- n, d = map(int, np.fromfile(fname, dtype="uint32", count=2))
- assert os.stat(fname).st_size == 8 + n * d * np.dtype(dtype).itemsize
+ # for search
+ # n, d = map(int, np.fromfile(fname, dtype="uint32", count=2))
+ # for dim reduced file of dtype float32
+ n, d = map(int, np.fromfile(fname, dtype="float32", count=2))
+ n = 100000000
+ d = 32
+ dtype = "float32"
+ print(f"fname={fname} n={n} d={d} dtype={dtype} os.stat(fname).st_size={os.stat(fname).st_size} np.dtype(dtype).itemsize={np.dtype(dtype).itemsize}")
+ # offset = 8
+ offset = 0
+ assert os.stat(fname).st_size == offset + n * d * np.dtype(dtype).itemsize
  if maxn > 0:
  n = min(n, maxn)
  return np.memmap(fname, dtype=dtype, mode="r", offset=8, shape=(n, d))
@@ -278,6 +287,7 @@ def get_dataset_iterator(self, bs=512, split=(1,0)):
  i0, i1 = self.nb * rank // nsplit, self.nb * (rank + 1) // nsplit
  filename = self.get_dataset_fn()
  x = xbin_mmap(filename, dtype=self.dtype, maxn=self.nb)
+ print(f"x.shape={x.shape} self.nb={self.nb} self.d={self.d}")
  assert x.shape == (self.nb, self.d)
  for j0 in range(i0, i1, bs):
  j1 = min(j0 + bs, i1)
@@ -369,6 +379,33 @@ def __init__(self, nb_M=1000):
  def distance(self):
  return "euclidean"
 
+
+class BigANNDimReducedDataset(DatasetCompetitionFormat):
+ def __init__(self, nb_M=1000):
+ self.nb_M = nb_M
+ self.nb = 10**6 * nb_M
+ self.d = 32
+ self.nq = 10000
+ # use for indexing of the dimensionality reduced dataset
+ self.dtype = "float32"
+ # use for search over the index built from the dimensionality reduced dataset
+ # self.dtype = "uint8"
+ self.ds_fn = "base.1B.u8bin"
+ self.qs_fn = "query.public.10K.u8bin"
+ self.gt_fn = (
+ "GT.public.1B.ibin" if self.nb_M == 1000 else
+ subset_url + "GT_100M/bigann-100M" if self.nb_M == 100 else
+ subset_url + "GT_10M/bigann-10M" if self.nb_M == 10 else
+ None
+ )
+ # self.gt_fn = "https://comp21storage.blob.core.windows.net/publiccontainer/comp21/bigann/public_query_gt100.bin" if self.nb == 10**9 else None
+ self.base_url = "https://dl.fbaipublicfiles.com/billion-scale-ann-benchmarks/bigann/"
+ self.basedir = os.path.join(BASEDIR, "bigann")
+
+ def distance(self):
+ return "euclidean"
+
+
 class Deep1BDataset(DatasetCompetitionFormat):
  def __init__(self, nb_M=1000):
  self.nb_M = nb_M
@@ -600,6 +637,7 @@ def default_count(self):
  'bigann-1B': lambda : BigANNDataset(1000),
  'bigann-100M': lambda : BigANNDataset(100),
  'bigann-10M': lambda : BigANNDataset(10),
+ 'bigann-dim-reduced-100M': lambda: BigANNDimReducedDataset(100),
 
  'deep-1B': lambda : Deep1BDataset(),
  'deep-100M': lambda : Deep1BDataset(100),

run_t1_faiss_baseline_eval.sh

@@ -0,0 +1,37 @@
+params="
+nprobe=1,quantizer_efSearch=4
+nprobe=2,quantizer_efSearch=4
+nprobe=4,quantizer_efSearch=4
+nprobe=4,quantizer_efSearch=8
+nprobe=8,quantizer_efSearch=4
+nprobe=8,quantizer_efSearch=8
+nprobe=8,quantizer_efSearch=16
+nprobe=8,quantizer_efSearch=32
+nprobe=16,quantizer_efSearch=16
+nprobe=16,quantizer_efSearch=32
+nprobe=16,quantizer_efSearch=64
+nprobe=32,quantizer_efSearch=8
+nprobe=32,quantizer_efSearch=32
+nprobe=32,quantizer_efSearch=64
+nprobe=32,quantizer_efSearch=128
+nprobe=64,quantizer_efSearch=16
+nprobe=64,quantizer_efSearch=32
+nprobe=64,quantizer_efSearch=64
+nprobe=64,quantizer_efSearch=128
+nprobe=64,quantizer_efSearch=256
+nprobe=128,quantizer_efSearch=32
+nprobe=128,quantizer_efSearch=64
+nprobe=128,quantizer_efSearch=128
+nprobe=128,quantizer_efSearch=256
+nprobe=128,quantizer_efSearch=512
+nprobe=256,quantizer_efSearch=64
+nprobe=256,quantizer_efSearch=128
+nprobe=256,quantizer_efSearch=512
+nprobe=512,quantizer_efSearch=256
+nprobe=512,quantizer_efSearch=512
+nprobe=1024,quantizer_efSearch=256
+"
+
+python track1_baseline_faiss/baseline_faiss.py \
+ --dataset bigann-dim-reduced-100M --indexfile data/track1_baseline_faiss/bigann-100M.IVF1M_2level_PQ64x4fsr.faissindex \
+ --search --searchparams $params

run_t1_faiss_baseline_index.sh

@@ -0,0 +1,11 @@
+python -u track1_baseline_faiss/baseline_faiss.py --dataset bigann-dim-reduced-100M \
+ --indexkey OPQ64_128,IVF1048576_HNSW32,PQ64x4fsr \
+ --maxtrain 10000000 \
+ --two_level_clustering \
+ --build \
+ --add_splits 30 \
+ --indexfile data/track1_baseline_faiss_dim_reduction/bigann-100M.IVF1M_2level_PQ64x4fsr.faissindex \
+ --quantizer_efConstruction 200 \
+ --quantizer_add_efSearch 80
+
+