embedding-as-service 1.0.0

embedding-as-service: one-stop solution to encode sentence to vectors using various embedding methods

embedding-as-service

One-Stop Solution to encode sentence to fixed length vectors from various embedding techniques
• Inspired from bert-as-service

What is it • Installation • Getting Started • Supported Embeddings • API •

What is it

Encoding/Embedding is a upstream task of encoding any inputs in the form of text, image, audio, video, transactional data to fixed length vector. Embeddings are quite popular in the field of NLP, there has been various Embeddings models being proposed in recent years by researchers, some of the famous one are bert, xlnet, word2vec etc. The goal of this repo is to build one stop solution for all embeddings techniques available, here we are starting with popular text embeddings for now and later on we aim to add as much technique for image, audio, video inputs also.

Finally, embedding-as-service help you to encode any given text to fixed length vector from supported embeddings and models.

:floppy_disk: Installation

Install the embedding-as-servive via pip.

$ pip install embedding-as-service

Note that the code MUST be running on Python >= 3.6. Again module does not support Python 2!

:zap: Getting Started

1. Intialise encoder using supported embedding and models from here

>>> from embedding_as_service.text.encode import Encoder  
>>> en = Encoder(embedding='bert', model='bert_base_cased', download=True)  

2. Get sentences tokens embedding

>>> vecs = en.encode(texts=['hello aman', 'how are you?'])  
>>> vecs  
array([[[ 1.7049843 ,  0.        ,  1.3486509 , ..., -1.3647075 ,  
 0.6958289 ,  1.8013777 ], ... [ 0.4913215 ,  0.60877025,  0.73050433, ..., -0.64490885, 0.8525057 ,  0.3080206 ]]], dtype=float32)  
>>> vecs.shape  
(2, 128, 768) # batch x max_sequence_length x embedding_size  

3. Using pooling strategy, click here for more.

>>> vecs = en.encode(texts=['hello aman', 'how are you?'], pooling='reduce_mean')  
>>> vecs  
array([[-0.33547154,  0.34566957,  1.1954105 , ...,  0.33702594,  
 1.0317835 , -0.785943  ], [-0.3439088 ,  0.36881036,  1.0612687 , ...,  0.28851607, 1.1107115 , -0.6253736 ]], dtype=float32)  

>>> vecs.shape  
(2, 768) # batch x embedding_size  

4. Use custom max_seq_length, default is 128

>>> vecs = en.encode(texts=['hello aman', 'how are you?'], max_seq_length=256)  
>>> vecs  
array([[ 0.48388457, -0.01327741, -0.76577514, ..., -0.54265064,  
 -0.5564591 ,  0.6454179 ], [ 0.53209245,  0.00526248, -0.71091074, ..., -0.5171917 , -0.40458363,  0.6779779 ]], dtype=float32)  

>>> vecs.shape  
(2, 256, 768) # batch x max_sequence_length x embedding_size  

5. Show embedding Tokens

>>> en.tokenize(texts=['hello aman', 'how are you?'])  
[['_hello', '_aman'], ['_how', '_are', '_you', '?']]  

6. Using your own tokenizer

>>> texts = ['hello aman!', 'how are you']  

# a naive whitespace tokenizer  
>>> tokens = [s.split() for s in texts]  
>>> vecs = en.encode(tokens, is_tokenized=True)  

:clipboard: API

1. class embedding_as_service.text.encoder.Encoder

Argument	Type	Default	Description
embedding	str	Required	embedding method to be used, check Embedding column here
model	str	Required	Model to be used for mentioned embedding, check Model column here
download	bool	False	Download model if model does not exists

2. def embedding_as_service.text.encoder.Encoder.encode

Argument	Type	Default	Description
Texts	List[str] or List[List[str]]	Required	List of sentences or list of list of sentence tokens in case of is_tokenized=True
pooling	str	(Optional)	Pooling methods to apply, here is available methods
max_seq_length	int	128	Maximum Sequence Length, default is 128
is_tokenized	bool	False	set as True in case of tokens are passed for encoding
batch_size	int	128	maximum number of sequences handled by encoder, larger batch will be partitioned into small batches.

2. def embedding_as_service.text.encoder.Encoder.tokenize

Argument	Type	Default	Description
Texts	List[str]	Required	List of sentences

:white_check_mark: Supported Embeddings and Models

Here are the list of supported embeddings and their respective models.

	Embedding	Model	Embedding dimensions	Paper
:one:	xlnet	xlnet_large_cased	1024	Read Paper :bookmark:
		xlnet_base_cased	768
:two:	bert	bert_base_uncased	768	Read Paper :bookmark:
		bert_base_cased	768
		bert_multi_cased	768
		bert_large_uncased	1024
		bert_large_cased	1024
:three:	elmo	elmo_bi_lm	512	Read Paper :bookmark:
:four:	ulmfit	ulmfit_forward	300	Read Paper :bookmark:
		ulmfit_backward	300
:five:	use	use_dan	512	Read Paper :bookmark:
		use_transformer_large	512
		use_transformer_lite	512
:six:	word2vec	google_news_300	300	Read Paper :bookmark:
:seven:	fasttext	wiki_news_300	300	Read Paper :bookmark:
		wiki_news_300_sub	300
		common_crawl_300	300
		common_crawl_300_sub	300
:eight:	glove	twitter_200	200	Read Paper :bookmark:
		twitter_100	100
		twitter_50	50
		twitter_25	25
		wiki_300	300
		wiki_200	200
		wiki_100	100
		wiki_50	50
		crawl_42B_300	300
		crawl_840B_300	300

:heavy_plus_sign: Pooling Strategies

Here is a table summarizes all supported pooling strategies

Strategy	Description
None	no pooling at all, useful when you want to use word embedding instead of sentence embedding. This will results in a [max_seq_len, embedding_size] encode matrix for a sequence.
reduce_mean	take the average of all token embeddings
reduce_min	take the minumun of all token embeddings
reduce_max	take the maximum of all token embeddings
reduce_mean_max	do reduce_mean and reduce_max separately and then concat them together
first_token	get the token embedding of first token of a sentence
last_token	get the token embedding of last token of a sentence