A feature-packed Python package and vector storage file format for utilizing vector embeddings in machine learning models in a fast, efficient, and simple manner developed by Plasticity. It is primarily intended to be a simpler / faster alternative to Gensim, but can be used as a generic key-vector store for domains outside NLP.
Table of Contents
You can install this package with
pip install pymagnitude # Python 2.7 pip3 install pymagnitude # Python 3
Vector space embedding models have become increasingly common in machine learning and traditionally have been popular for natural language processing applications. A fast, lightweight tool to consume these large vector space embedding models efficiently is lacking.
The Magnitude file format (
.magnitude) for vector embeddings is intended to be a more efficient universal vector embedding format that allows for lazy-loading for faster cold starts in development, LRU memory caching for performance in production, multiple key queries, direct featurization to the inputs for a neural network, performant similiarity calculations, and other nice to have features for edge cases like handling out-of-vocabulary keys or misspelled keys and concatenating multiple vector models together. It also is intended to work with large vector models that may not fit in memory.
It uses SQLite, a fast, popular embedded database, as its underlying data store. It uses indexes for fast key lookups as well as uses memory mapping, SIMD instructions, and spatial indexing for fast similarity search in the vector space off-disk with good memory performance even between multiple processes. Moreover, memory maps are cached between runs so even after closing a process, speed improvements are reaped.
Benchmarks and Features
|Metric||Magnitude Light||Magnitude Medium||Magnitude Heavy|
|Initial load time||0.7210s||━ 1||━ 1|
|Cold single key query||0.0001s||━ 1||━ 1|
|Warm single key query |
(same key as cold query)
|0.00004s||━ 1||━ 1|
|Cold multiple key query |
|0.0442s||━ 1||━ 1|
|Warm multiple key query |
(n=25) (same keys as cold query)
|0.00004s||━ 1||━ 1|
(n=10) (worst case)
|247.05s||━ 1||━ 1|
(n=10) (average case) (w/ disk persistent cache)
|1.8217s||━ 1||━ 1|
(n=10) (different key than first query)
|0.2434s||━ 1||━ 1|
|Warm subsequent |
(n=10) (same key as first query)
(n=10, effort=1.0) (worst case)
(n=10, effort=1.0) (average case) (w/ disk persistent cache)
(n=10, effort=1.0) (different key than first query)
(n=10, effort=0.1) (different key than first query)
|Warm subsequent |
(n=10, effort=1.0) (same key as first query)
|Process memory (RAM) utilization||18KB||━ 1||━ 1|
|Process memory (RAM) utilization after 100 key queries||168KB||━ 1||━ 1|
|Process memory (RAM) utilization after 100 key queries + similarity search||342KB2||━ 1||━ 1|
|Integrity checks and tests|
|Universal format between word2vec (|
|Simple, Pythonic interface|
|Support for larger than memory models|
|Lazy loading whenever possible for speed and performance|
|Optimized for |
|Bulk and multiple key lookup with padding, truncation, placeholder, and featurization support|
|Concatenting multiple vector models together|
|Basic out-of-vocabulary key lookup |
(character n-gram feature hashing)
|Advanced out-of-vocabulary key lookup with support for misspellings |
(character n-gram feature hashing to similar in-vocabulary keys)
|Approximate most similar search with an annoy index|
|Built-in training for new models|
1: same value as previous column
mmap to read from disk, so the OS will still allocate pages of memory when memory is available, but it can be shared between processes and isn't managed within each process for extremely large files which is a performance win
*: All benchmarks were performed on the Google News pre-trained word vectors (
GoogleNews-vectors-negative300.bin) with a MacBook Pro (Retina, 15-inch, Mid 2014) 2.2GHz quad-core Intel Core i7 @ 16GB RAM on SSD over an average of trials where feasible.
Pre-converted Magnitude Formats of Popular Embeddings Models
Popular embedding models have been pre-converted to the
.magnitude format for immmediate download and usage:
(basic support for out-of-vocabulary keys)
(advanced support for out-of-vocabulary keys)
(advanced support for out-of-vocabulary keys and faster
|Google - word2vec||Google News 100B||300D||300D||300D|
|Stanford - GloVe||Wikipedia 2014 + Gigaword 5 6B||50D, 100D, 200D, 300D||50D, 100D, 200D, 300D||50D, 100D, 200D, 300D|
|Stanford - GloVe||Wikipedia 2014 + Gigaword 5 6B |
(lemmatized by Plasticity)
|50D, 100D, 200D, 300D||50D, 100D, 200D, 300D||50D, 100D, 200D, 300D|
|Stanford - GloVe||Common Crawl 840B||300D||300D||300D|
|Stanford - GloVe||Twitter 27B||25D, 50D, 100D, 200D||25D, 50D, 100D, 200D||25D, 50D, 100D, 200D|
|Facebook - fastText||English Wikipedia 2017 16B||300D||300D||300D|
|Facebook - fastText||English Wikipedia 2017 + subword 16B||300D||300D||300D|
|Facebook - fastText||Common Crawl 600B||300D||300D||300D|
There are instructions below for converting any
.vec file to a
Using the Library
Constructing a Magnitude Object
You can create a Magnitude object like so:
from pymagnitude import * vectors = Magnitude("/path/to/vectors.magnitude")
If needed, and included for convenience, you can also open a
.vec file directly with Magnitude. This is, however, less efficient and very slow for large models as it will convert the file to a
.magnitude file on the first run into a temporary directory. The temporary directory is not guaranteed to persist and does not persist when your computer reboots. You should pre-convert
.vec files with
python -m pymagnitude.converter typically for faster speeds, but this feature is useful for one-off use-cases. A warning will be generated when instantiating a Magnitude object directly with a
.vec. You can supress warnings by setting the
supress_warnings argument in the constructor to
- By default, lazy loading is enabled. You can pass in an optional
lazy_loadingargument to the constructor with the value
-1to disable lazy-loading and pre-load all vectors into memory (a la Gensim),
0(default) to enable lazy-loading with an unbounded in-memory LRU cache, or an integer greater than zero
Xto enable lazy-loading with an LRU cache that holds the
Xmost recently used vectors in memory.
- If you want the data for the
most_similarfunctions to be pre-loaded eagerly on initialization, set
- Note, even when
lazy_loadingis set to
eageris set to
Truedata will be pre-loaded into memory in a background thread to prevent the constructor from blocking for a few minutes for large models. If you really want blocking behavior, you can pass
- By default, NumPy arrays are returned for queries. Set the optional argument
Falseif you wish to recieve Python lists instead.
- By default, querying for keys is case-sensitive. Set the optional argument
Trueif you wish to perform case-insensitive searches.
- Optionally, you can include the
pad_to_lengthargument which will specify the length all examples should be padded to if passing in multple examples. Any examples that are longer than the pad length will be truncated.
- Optionally, you can set the
Trueif you want the beginning of the the list of keys in each example to be truncated instead of the end in case it is longer than
- Optionally, you can set the
Trueif you want the padding to appear at the beginning versus the end (which is the default).
- Optionally, you can pass in the
placeholdersargument, which will increase the dimensions of each vector by a
placeholdersamount, zero-padding those extra dimensions. This is useful, if you plan to add other values and information to the vectors and want the space for that pre-allocated in the vectors for efficiency.
- Optionally, you can pass in the
languageargument with an ISO 639-1 Language Code, which, if you are using Magnitude for word vectors, will ensure the library respects stemming and other language-specific features for that language. The default is
enfor English. You can also pass in
Noneif you are not using Magnitude for word vectors.
You can query the total number of vectors in the file like so:
You can query the dimensions of the vectors like so:
You can check if a key is in the vocabulary like so:
You can iterate through all keys and vectors like so:
You can query for the vector of a key like so:
You can index for the n-th key and vector like so:
You can query for the vector of multiple keys like so:
A 2D array (keys by vectors) will be returned.
You can query for the vector of multiple examples like so:
vectors.query([["I", "read", "a", "book"], ["I", "read", "a", "magazine"]])
A 3D array (examples by keys by vectors) will be returned. If
pad_to_length is not specified, and the size of each example is uneven, they will be padded to the length of the longest example.
You can index for the keys and vectors of multiple indices like so:
vectors[:42] # slice notation vectors[42, 1337, 2001] # tuple notation
You can query the distance of two or multiple keys like so:
vectors.distance("cat", "dog") vectors.distance("cat", ["dog", "tiger"])
You can query the similarity of two or multiple keys like so:
vectors.similarity("cat", "dog") vectors.similarity("cat", ["dog", "tiger"])
You can query for the most similar key out of a list of keys to a given key like so:
vectors.most_similar_to_given("cat", ["dog", "television", "laptop"]) # dog
You can query for which key doesn't match a list of keys to a given key like so:
vectors.doesnt_match(["breakfast", "cereal", "dinner", "lunch"]) # cereal
You can query for the most similar (nearest neighbors) keys like so:
vectors.most_similar("cat", topn = 100) # Most similar by key vectors.most_similar(vectors.query("cat"), topn = 100) # Most similar by vector
Optionally, you can pass a
max_distance argument to
most_similar. Since they are unit norm vectors, values from [0.0-2.0] are valid.
You can also query for the most similar keys giving positive and negative examples (which, incidentally, solves analogies) like so:
vectors.most_similar(positive = ["woman", "king"], negative = ["man"]) # queen
vectors.most_similar_cosmul function exists that uses the 3CosMul function from Levy and Goldberg:
vectors.most_similar_cosmul(positive = ["woman", "king"], negative = ["man"]) # queen
You can also query for the most similar keys using an approximate nearest neighbors index which is much faster, but doesn't guarantee the exact answer:
vectors.most_similar_approx("cat") vectors.most_similar_approx(positive = ["woman", "king"], negative = ["man"])
Optionally, you can pass an
effort argument with values between [0.0-1.0] to the
most_similar_approx function which will give you runtime trade-off. The default value for
effort is 1.0 which will take the longest, but will give the most accurate result.
You can query for all keys closer to a key than another key is like so:
vectors.closer_than("cat", "rabbit") # ["dog", ...]
You can access all of the underlying vectors in the model in a large
numpy.memmap array of size (
len(vectors) x vectors.emb_dim) like so:
You can clean up all associated resources, open files, and database connections like so:
Basic Out-of-Vocabulary Keys
For word vector representations, handling out-of-vocabulary keys is important to handling new words not in the trained model, handling mispellings and typos, and making models trained on the word vector representations more robust in general.
Out-of-vocabulary keys are handled by assigning them a random vector value. However, the randomness is deterministic. So if the same out-of-vocabulary key is encountered twice, it will be assigned the same random vector value for the sake of being able to train on those out-of-vocabulary keys. Moreover, if two out-of-vocabulary keys share similar character n-grams ("uberx", "uberxl") they will placed close to each other even if they are both not in the vocabulary:
vectors = Magnitude("/path/to/GoogleNews-vectors-negative300.magnitude") "uberx" in vectors # False "uberxl" in vectors # False vectors.query("uberx") # array([ 5.07109939e-02, -7.08248823e-02, -2.74812328e-02, ... ]) vectors.query("uberxl") # array([ 0.04734962, -0.08237578, -0.0333479, -0.00229564, ... ]) vectors.similarity("uberx", "uberxl") # 0.955000000200815
Advanced Out-of-Vocabulary Keys
If using a Magnitude file with advanced out-of-vocabulary support (Medium or Heavy), out-of-vocabulary keys will also be embedded close to similar keys (determined by string similarity) that are in the vocabulary:
vectors = Magnitude("/path/to/GoogleNews-vectors-negative300.magnitude") "uberx" in vectors # False "uberification" in vectors # False "uber" in vectors # True vectors.similarity("uberx", "uber") # 0.7383483267618451 vectors.similarity("uberification", "uber") # 0.745452837882727
Handling Misspellings and Typos
This also makes Magnitude robust to a lot of spelling errors:
vectors = Magnitude("/path/to/GoogleNews-vectors-negative300.magnitude") "missispi" in vectors # False vectors.similarity("missispi", "mississippi") # 0.35961736624824003 "discrimnatory" in vectors # False vectors.similarity("discrimnatory", "discriminatory") # 0.8309152561753461 "hiiiiiiiiii" in vectors # False vectors.similarity("hiiiiiiiiii", "hi") # 0.7069775034853861
Character n-grams are used to create this effect for out-of-vocabulary keys. The inspiration for this feature was taken from Facebook AI Research's Enriching Word Vectors with Subword Information, but instead of utilizing character n-grams at train time, character n-grams are used at inference so the effect can be somewhat replicated (but not perfectly replicated) in older models that were not trained with character n-grams like word2vec and GloVe.
Concatenation of Multiple Models
Optionally, you can combine vectors from multiple models to feed stronger information into a machine learning model like so:
from pymagnitude import * word2vec = Magnitude("/path/to/GoogleNews-vectors-negative300.magnitude") glove = Magnitude("/path/to/glove.6B.50d.magnitude") vectors = Magnitude(word2vec, glove) # concatenate word2vec with glove vectors.query("cat") # returns 350-dimensional NumPy array ('cat' from word2vec concatenated with 'cat' from glove) vectors.query(("cat", "cats")) # returns 350-dimensional NumPy array ('cat' from word2vec concatenated with 'cats' from glove)
You can concatenate more than two vector models, simply by passing more arguments to constructor.
Additional Featurization (Parts of Speech, etc.)
You can automatically create vectors from additional features you may have such as parts of speech, syntax dependency information, or any other information using the
from pymagnitude import * pos_vectors = FeaturizerMagnitude(100, namespace = "PartsOfSpeech") pos_vectors.dim # 4 - number of dims automatically determined by Magnitude from 100 pos_vectors.query("NN") # - array([ 0.08040417, -0.71705252, 0.61228951, 0.32322192]) pos_vectors.query("JJ") # - array([-0.11681135, 0.10259253, 0.8841201 , -0.44063763]) pos_vectors.query("NN") # - array([ 0.08040417, -0.71705252, 0.61228951, 0.32322192]) (deterministic hashing so the same value is returned every time for the same key) dependency_vectors = FeaturizerMagnitude(100, namespace = "SyntaxDependencies") dependency_vectors.dim # 4 - number of dims automatically determined by Magnitude from 100 dependency_vectors.query("nsubj") # - array([-0.81043793, 0.55401352, -0.10838071, 0.15656626]) dependency_vectors.query("prep") # - array([-0.30862918, -0.44487267, -0.0054573 , -0.84071788])
Magnitude will use the feature hashing trick internally to directly use the hash of the feature value to create a unique vector for that feature value.
The first argument to
FeaturizerMagnitude should be an approximate upper-bound on the number of values for the feature. Since there are < 100 parts of speech tags and < 100 syntax dependencies, we choose 100 for both in the example above. The value chosen will determine how many dimensions Magnitude will automatically assign to the particular the
FeaturizerMagnitude object to reduce the chance of a hash collision. The
namespace argument can be any string that describes your additional feature. It is optional, but highly recommended.
You can then concatenate these features for use with a standard Magnitude object:
from pymagnitude import * word2vec = Magnitude("/path/to/GoogleNews-vectors-negative300.magnitude") pos_vectors = FeaturizerMagnitude(100, namespace = "PartsOfSpeech") dependency_vectors = FeaturizerMagnitude(100, namespace = "SyntaxDependencies") vectors = Magnitude(word2vec, pos_vectors, dependency_vectors) # concatenate word2vec with pos and dependencies vectors.query([ ("I", "PRP", "nsubj"), ("saw", "VBD", "ROOT"), ("a", "DT", "det"), ("cat", "NN", "dobj"), (".", ".", "punct") ]) # array of size 5 x (300 + 4 + 4) or 5 x 308 # Or get a unique vector for every 'buffalo' in: # "Buffalo buffalo Buffalo buffalo buffalo buffalo Buffalo buffalo" # (https://en.wikipedia.org/wiki/Buffalo_buffalo_Buffalo_buffalo_buffalo_buffalo_Buffalo_buffalo) vectors.query([ ("Buffalo", "JJ", "amod"), ("buffalo", "NNS", "nsubj"), ("Buffalo", "JJ", "amod"), ("buffalo", "NNS", "nsubj"), ("buffalo", "VBP", "rcmod"), ("buffalo", "VB", "ROOT"), ("Buffalo", "JJ", "amod"), ("buffalo", "NNS", "dobj") ]) # array of size 8 x (300 + 4 + 4) or 8 x 308
A machine learning model, given this output, now has access to parts of speech information and syntax dependency information instead of just word vector information. In this case, this additional information can give neural networks stronger signal for semantic information and reduce the need for training data.
Using Magnitude with a ML library
Magnitude makes it very easy to quickly build and iterate on models that need to use vector representations by taking care of a lot of pre-processing code to convert a dataset of text (or keys) into vectors. Moreover, it can make these models more robust to out-of-vocabulary words and misspellings.
There is example code available using Magnitude to build an intent classification model for the ATIS (Airline Travel Information Systems) dataset (Train/Test), used for chatbots or conversational interfaces, in a few popular machine learning libraries below.
You can access a guide for using Magnitude with Keras (which supports TensorFlow, Theano, CNTK) at this Google Colaboratory Python notebook.
The PyTorch guide is coming soon.
The TFLearn guide is coming soon.
You can use the
MagnitudeUtils class for convenient access to functions that may be useful when creating machine learning models.
You can create a batch generator for
y data with
batchify, like so:
X = [.3, .2, .7, .8, .1] y = [0, 0, 1, 1, 0] batch_gen = MagnitudeUtils.batchify(X, y, 2) for X_batch, y_batch in batch_gen: print(X_batch, y_batch) # Returns: # 1st loop: X_batch = [.3, .2], y_batch = [0, 0] # 2nd loop: X_batch = [.7, .8], y_batch = [1, 1] # 3rd loop: X_batch = [.1], y_batch =  # next loop: repeats infinitely...
You can encode class labels to integers and back with
class_encoding, like so:
add_class, class_to_int, int_to_class = MagnitudeUtils.class_encoding() add_class("cat") # Returns: 0 add_class("dog") # Returns: 1 add_class("cat") # Returns: 0 class_to_int("dog") # Returns: 1 class_to_int("cat") # Returns: 0 int_to_class(1) # Returns: "dog" int_to_class(0) # Returns: "cat"
You can convert categorical data with class integers to one-hot NumPy arrays with
to_categorical, like so:
y = [1, 5, 2] MagnitudeUtils.to_categorical(y, num_classes = 6) # num_classes is optional # Returns: # array([[0., 1., 0., 0., 0., 0.] # [0., 0., 0., 0., 0., 1.] # [0., 0., 1., 0., 0., 0.]])
You can convert from one-hot NumPy arrays back to a 1D NumPy array of class integers with
from_categorical, like so:
y_c = [[0., 1., 0., 0., 0., 0.], [0., 0., 0., 0., 0., 1.]] MagnitudeUtils.from_categorical(y_c) # Returns: # array([1., 5.])
Concurrency and Parallelism
The library is thread safe (it uses a different connection to the underlying store per thread), is read-only, and it never writes to the file. Because of the light-memory usage, you can also run it in multiple processes (or use
multiprocessing) with different address spaces without having to duplicate the data in-memory like with other libraries and without having to create a multi-process shared variable since data is read off-disk and each process keeps its own LRU memory cache. For heavier functions, like
most_similar a shared memory mapped file is created to share memory between processes.
File Format and Converter
The Magnitude package uses the
.magnitude file format instead of
.vec as with other vector models like word2vec, GloVe, and fastText. There is an included command-line utility for converting word2vec, GloVe, fastText files to Magnitude files.
You can convert them like so:
python -m pymagnitude.converter -i <PATH TO FILE TO BE CONVERTED> -o <OUTPUT PATH FOR MAGNITUDE FILE>
The input format will automatically be determined by the extension / the contents of the input file. When the vectors are converted, they will also be unit-length normalized. You should only need to perform this conversion once for a model. After converting, the Magnitude file format is static and it will not be modified or written to make concurrent read access safe.
The flags for
pymagnitude.converter are specified below:
- You can pass in the
-hflag for help and to list all flags.
- You can use the
-p <PRECISION>flag to specify the decimal precision to retain (selecting a lower number will create smaller files). The actual underlying values are stored as integers instead of floats so this is essentially quantization for smaller model footprints.
- You can add an approximate nearest neighbors index to the file (increases size) with the
-aflag which will enable the use of the
-t <TREES>flag controls the number of trees in the approximate neigherest neighbors index (higher is more accurate) when used in conjunction with the
-aflag (if not supplied, the number of trees is automatically determined).
- You can pass the
-sflag to disable adding subword information to the file (which will make the file smaller), but disable advanced out-of-vocabulary key support.
Optionally, you can bulk convert many files by passing an input folder and output folder instead of an input file and output file. All
.vec files in the input folder will be converted to
.magnitude files in the the output folder. The output folder must exist before a bulk conversion operation.
Other documentation is not available at this time. See the source file directly (it is well commented) if you need more information about a method's arguments or want to see all supported features.
Currently, we only provide English word vector models on this page pre-converted to the
.magnitude format. You can, however, still use Magnitude with word vectors of other languages. Facebook has trained their fastText vectors for many different languages. You can down the
.vec file for any language you want and then convert it to
.magnitude with the converter.
Other Programming Languages
Currently, reading Magnitude files is only supported in Python, since it has become the de-facto language for machine learning. This is sufficient for most use cases. Extending the file format to other languages shouldn't be difficult as SQLite has a native C implementation and has bindings in most languages. The file format itself and the protocol for reading and searching is also fairly straightforward upon reading the source code of this repository.
Currently, natural language processing is the most popular domain that uses pre-trained vector embedding models for word vector representations. There are, however, other domains like computer vision that have started using pre-trained vector embedding models like Deep1B for image representation. This library intends to stay agnostic to various domains and instead provides a generic key-vector store and interface that is useful for all domains.
The main repository for this project can be found on GitLab. The GitHub repository is only a mirror. Pull requests for more tests, better error-checking, bug fixes, performance improvements, or documentation or adding additional utilties / functionalities are welcome on GitLab.
You can contact us at email@example.com.
Other Notable Projects
- spotify/annoy - Powers the approximate nearest neighbors algorithm behind
most_similar_approxin Magnitude using random-projection trees and hierarchical 2-means. Thanks to author Erik Bernhardsson for helping out with some of the integration details between Magnitude and Annoy.
Citing this Repository
Clicking on the badge will lead to a page that will help you generate proper BibTeX citations, JSON-LD citations, and other citations.
LICENSE and Attribution
This repository is licensed under the license found here.