Skip to content

Latest commit

 

History

History
104 lines (75 loc) · 3.24 KB

readme.md

File metadata and controls

104 lines (75 loc) · 3.24 KB
Raw

Made with Jupyter nbviewer Open in HTML

Introduction

The objective of this project is to perform sentiment analysis (only positive and negative) on the 515k hotel review dataset and build an API that can classify review text.

This project covers:

  • TF-IDF
  • count features
  • logistic regression
  • naive bayes
  • svm
  • xgboost
  • grid search
  • word vectors (Universal Sentence Encoder) model from Tensorflow HUB)
  • LSTM

The final LSTM model achieved an accuracy of ~81% in Test Dataset (75:25 split)

Sneak Peek Into Data

image

Dataset Source

Text Embedding Model

How to Use The Exported Model

# use the latest `tensorflow_text` version
# !pip install tensorflow_text

import numpy as np
from numpy import newaxis

import tensorflow as tf
import tensorflow_hub as hub
import tensorflow_text

# load Universal Sentence Encoder Multilingual Model, v3 is ~300 MB
module_url = 'https://tfhub.dev/google/universal-sentence-encoder-multilingual-large/3'
use = hub.load(module_url)

imported_model = tf.keras.models.load_model('/content/lstm_final_model.h5')

def predict_sentiment(txt):
  # generate embedding
  emb_txt = use(txt)
  
  # reshape to pass into the model
  emb_test_reshaped = emb_txt[:, newaxis, :]
  
  # predict sentiment score
  sentiment_val = np.argmax(imported_model.predict(emb_test_reshaped)) # for 1hot-encoded labels

  sentiment_val = (imported_model.predict(emb_test_reshaped) > 0.5).astype('int32') # for binary labels

  # sentiment score
  print("Score:", imported_model.predict(emb_test_reshaped).flatten()[0])
  
  # return sentiment value based on score
  return "Positive" if sentiment_val == 1 else "Negative"

sample_text = "I like the room service"
pred_sentiment = predict_sentiment(sample_text)
print(f"The sentiment of this sentence is : {pred_sentiment}")

# output : 
# The sentiment of this sentence is : Positive

FastAPI Project Structure

|-- model
  |-- lstm_sentiment_model.h5
|-- tfhub
  |--  universal-sentence-encoder-multilingual-large-v3
    |-- assets
    |-- variables
      |-- variables.data-00000-of-00001
      |-- variables.index
    |-- saved_model.pb
|-- main.py

Built With

tensorflow==2.7.0
tensorflow_text==0.12.0
tensorflow-hub==0.12.0
keras==2.7.0
keras-vis==0.4.1
scikit-learn==0.22.2.post1
fastapi==0.70.1
uvicorn==0.16.0