chunkdot

ChunkDot

Multi-threaded matrix multiplication and cosine similarity calculations for dense and sparse matrices. Appropriate for calculating the K most similar items for a large number of items by chunking the item matrix representation (embeddings) and using Numba to accelerate the calculations.

Use for:

• dense embeddings
• sparse embeddings
• similarity calculation versus other embeddings
• CosineSimilarityTopK scikit-learn transformer

Related blog posts

• Cosine Similarity for 1 Trillion Pairs of Vectors
• Bulk Similarity Calculations for Sparse Embeddings

Usage

pip install -U chunkdot

Dense embeddings

Calculate the 50 most similar and dissimilar items for 100K items.

import numpy as np
from chunkdot import cosine_similarity_top_k

embeddings = np.random.randn(100000, 256)
# using all you system's memory
cosine_similarity_top_k(embeddings, top_k=50)
# most dissimilar items using 20GB
cosine_similarity_top_k(embeddings, top_k=-50, max_memory=20E9)

<100000x100000 sparse matrix of type '<class 'numpy.float64'>'
 with 5000000 stored elements in Compressed Sparse Row format>

# with progress bar
cosine_similarity_top_k(embeddings, top_k=50, show_progress=True)

100%|███████████████████████████████████████████████████████████████| 129.0/129 [01:04<00:00,  1.80it/s]
<100000x100000 sparse matrix of type '<class 'numpy.float64'>'
  with 5000000 stored elements in Compressed Sparse Row format>

Execution time

from timeit import timeit
import numpy as np
from chunkdot import cosine_similarity_top_k

embeddings = np.random.randn(100000, 256)
timeit(lambda: cosine_similarity_top_k(embeddings, top_k=50, max_memory=20E9), number=1)

58.611996899999994

Sparse embeddings

Calculate the 50 most similar and dissimilar items for 100K items. Items represented by 10K dimensional vectors and an embeddings matrix of 0.005 density.

from scipy import sparse
from chunkdot import cosine_similarity_top_k

embeddings = sparse.rand(100000, 10000, density=0.005)
# using all you system's memory
cosine_similarity_top_k(embeddings, top_k=50)
# most dissimilar items using 20GB
cosine_similarity_top_k(embeddings, top_k=-50, max_memory=20E9)

<100000x100000 sparse matrix of type '<class 'numpy.float64'>'
 with 5000000 stored elements in Compressed Sparse Row format>

Execution time

from timeit import timeit
from scipy import sparse
from chunkdot import cosine_similarity_top_k

embeddings = sparse.rand(100000, 10000, density=0.005)
timeit(lambda: cosine_similarity_top_k(embeddings, top_k=50, max_memory=20E9), number=1)

51.87472256699999

Similarity calculation versus other embeddings

Given 20K items, for each item, find the 50 most similar items in a collection of other 10K items.

import numpy as np
from chunkdot import cosine_similarity_top_k

embeddings = np.random.randn(20000, 256)
other_embeddings = np.random.randn(10000, 256)

cosine_similarity_top_k(embeddings, embeddings_right=other_embeddings, top_k=10)

<20000x10000 sparse matrix of type '<class 'numpy.float64'>'
 with 200000 stored elements in Compressed Sparse Row format>

CosineSimilarityTopK scikit-learn transformer

Given a pandas DataFrame with 100K rows and

• 2 numerical columns
• 2 categorical columns with 500 categories each

use scikit-learn transformers, the standard scaler for the numerical columns and the one-hot encoder for the categorical columns, to form an embeddings matrix of dimensions 100K x 1002 and then calculate the top 50 most similar rows per each row.

import numpy as np
import pandas as pd

n_rows = 100000
n_categories = 500
df = pd.DataFrame(
    {
        "A_numeric": np.random.rand(n_rows),
        "B_numeric": np.random.rand(n_rows),
        "C_categorical": np.random.randint(n_categories, size=n_rows),
        "D_categorical": np.random.randint(n_categories, size=n_rows),
    }
)

from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OneHotEncoder, StandardScaler

from chunkdot import CosineSimilarityTopK

numeric_features = ["A_numeric", "B_numeric"]
numeric_transformer = Pipeline(steps=[("scaler", StandardScaler())])

categorical_features = ["C_categorical", "D_categorical"]
categorical_transformer = Pipeline(steps=[("encoder", OneHotEncoder())])

preprocessor = ColumnTransformer(
    transformers=[
        ("num", numeric_transformer, numeric_features),
        ("cat", categorical_transformer, categorical_features),
    ]
)

cos_sim = CosineSimilarityTopK(top_k=50)

pipe = Pipeline(steps=[("preprocessor", preprocessor), ("cos_sim", cos_sim)])
pipe.fit_transform(df)

<100000x100000 sparse matrix of type '<class 'numpy.float64'>'
	with 5000000 stored elements in Compressed Sparse Row format>

Execution time

from timeit import timeit

timeit(lambda: pipe.fit_transform(df), number=1)

24.45172154181637