Performance testing of BERT based Sentence Transformers for sentence encoding

Download Code

import pandas as pd
from sentence_transformers import SentenceTransformer

sbert_model = SentenceTransformer('bert-base-nli-mean-tokens')

df = pd.read_csv('nytEditorialSnippets_GroundTruth.txt', sep = '\t')

df_10 = df[0:10]
df_100 = df[0:100]

# 1. Using Sentence Encoder in a function on every record

def get_embedding(input_sentence):
    return sbert_model.encode([input_sentence])[0]

%%timeit
df_out_1 = df_100['text'].apply(get_embedding)



9.15 s ± 317 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)



%%time
df_out_1 = df_100['text'].apply(get_embedding)



CPU times: user 18.3 s, sys: 112 ms, total: 18.5 s
Wall time: 9.25 s



%%time
df_out_1 = df_10['text'].apply(get_embedding)



CPU times: user 1.97 s, sys: 7.88 ms, total: 1.98 s
Wall time: 997 ms



df_out_1



0     [-0.7970602, 0.47616163, 0.2621567, 0.38846374...
1     [-0.32450542, -0.10945253, 0.6443658, 0.212320...
2     [-0.2602994, -0.0036350375, 1.2917686, 0.12602...
3     [0.5173101, -0.86385506, 1.5003084, 0.76273316...
4     [-0.19630705, 1.611963, 0.8502133, 0.059544455...
                            ...                        
95    [-0.7843676, 0.70446295, -0.86373883, 0.096476...
96    [-0.047543377, -1.0461698, 0.9984542, 0.776394...
97    [-0.5863306, 0.38590172, -0.15509816, 0.275745...
98    [0.22616625, 0.33843663, 0.030288033, 0.191214...
99    [0.40268317, 1.1528935, 0.3597172, 0.16918863,...
Name: text, Length: 100, dtype: object



type(df_out_1)



pandas.core.series.Series



# 2. Using Sentence Encoder on entire array of sentences at once

%%timeit
df_out_2 = sbert_model.encode(df_100['text'].values)



9.25 s ± 275 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)



%%time
df_out_2 = sbert_model.encode(df_100['text'].values)



CPU times: user 17.8 s, sys: 1 s, total: 18.8 s
Wall time: 9.41 s



df_out_2



array([[-0.7970604 ,  0.4761617 ,  0.262157  , ..., -0.2437577 ,
1.1380528 ,  0.28697717],
[-0.32450542, -0.10945235,  0.64436567, ...,  0.14436643,
-0.24656864, -0.18447737],
[-0.26029944, -0.00363465,  1.2917686 , ...,  1.1463983 ,
-1.0714562 , -0.09548129],
...,
[-0.5863306 ,  0.38590172, -0.15509816, ..., -0.2937488 ,
-0.3724223 ,  0.1826524 ],
[ 0.22616649,  0.33843663,  0.03028765, ...,  0.66338176,
-0.6620043 ,  0.09410357],
[ 0.40268335,  1.152894  ,  0.35971704, ..., -0.9203086 ,
0.17893644,  0.71039814]], dtype=float32)



type(df_out_2)


numpy.ndarray

df_out_2.shape

(100, 768)

Tags: Technology,Natural Language Processing,