import streamlit as st from functions_preprocess import LinguisticPreprocessor, download_if_non_existent import pickle import nltk nltk.download('stopwords') nltk.download('punkt') download_if_non_existent('corpora/stopwords', 'stopwords') download_if_non_existent('taggers/averaged_perceptron_tagger', 'averaged_perceptron_tagger') download_if_non_existent('corpora/wordnet', 'wordnet') #################################################################### Streamlit interface st.title("Movie Reviews: An NLP Sentiment analysis") #################################################################### Cache the model loading @st.cache_data() def load_model(model_path): model_pkl_file = "sentiment_model.pkl" with open(model_pkl_file, 'rb') as file: model = pickle.load(file) return model def load_cnn(): model.load_state_dict(torch.load('model_cnn.pkl')) model.eval() return model def predict_sentiment(text, model): processor.transform(text) prediction = model.predict([text]) return prediction model_1 = load_model() model_2 = load_cnn() processor = LinguisticPreprocessor() ############################################################# Text input with st.expander("Model 1: SGD Classifier"): st.markdown("Give it a go by writing a positive or negative text, and analyze it!") # Text input inside the expander user_input = st.text_area("Enter text here...") if st.button('Analyze'): # Displaying output result = predict_sentiment(user_input, model_1) if result >= 0.5: st.write('The sentiment is: Positive 😀') else: st.write('The sentiment is: Negative 😞') with st.expander("Model 2: CNN Sentiment analysis"): st.markdown("Give it a go by writing a positive or negative text, and analyze it!") # Text input inside the expander user_input = st.text_area("Enter text here...") if st.button('Analyze'): # Displaying output result = predict_sentiment(user_input, model_2) if result >= 0.5: st.write('The sentiment is: Positive 😀') else: st.write('The sentiment is: Negative 😞') st.caption("Por @efeperro.")