pages/✨BERT.py

import streamlit as st
import base64
import streamlit as st
import plotly.express as px

df = px.data.iris()

@st.cache_data
def get_img_as_base64(file):
    with open(file, "rb") as f:
        data = f.read()
    return base64.b64encode(data).decode()


page_bg_img = f"""
<style>
[data-testid="stAppViewContainer"] > .main {{
background-image: url("https://i.ibb.co/kH8bcr4/1368432.jpg");
background-size: 115%;
background-position: top left;
background-repeat: no-repeat;
background-attachment: local;
}}

[data-testid="stSidebar"] > div:first-child {{
background-image: url("https://ibb.co/ZBkdJRg");
background-size: 115%;
background-position: center; 
background-repeat: no-repeat;
background-attachment: fixed;
}}

[data-testid="stHeader"] {{
background: rgba(0,0,0,0);
}}

[data-testid="stToolbar"] {{
right: 2rem;
}}

div.css-1n76uvr.e1tzin5v0 {{
background-color: rgba(238, 238, 238, 0.5);
border: 10px solid #EEEEEE;
padding: 5% 5% 5% 10%;
border-radius: 5px;
}}

</style>
"""
st.markdown(page_bg_img, unsafe_allow_html=True)

import tensorflow as tf
from tensorflow import keras
import numpy as np
import matplotlib.pyplot as plt

################################################################################################
import torch
import numpy as np
import transformers 
import pickle
from sklearn.linear_model import LogisticRegression

#def load_model():
model = transformers.DistilBertModel.from_pretrained(
"distilbert-base-uncased",
output_attentions = False,
output_hidden_states = False
)
#model.load_state_dict(torch.load('model/modelbert.pt', map_location=torch.device('cpu')))
tokenizer = transformers.DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
    #return model_finetuned, tokenizer

def preprocess_text(text_input, max_len, tokenizer):
    input_tokens = tokenizer(
        text_input, 
        return_tensors='pt', 
        padding=True, 
        max_length=max_len,
        truncation = True
        )
    return input_tokens

def predict_sentiment(model, input_tokens):
    #st.write(input_tokens)
    ans = {0: "NEGATIVE", 1: "POSITIVE"}
    last_hidden_states = model(**input_tokens)
    #st.write('last_hidden_states ok')
    vectors = last_hidden_states[0][:,0,:].detach().cpu().numpy()
    #st.write('vectors ok')
    #output = model(**input_tokens).last_hidden_states.detach().numpy()
    with open('model/modelbert1.pkl', 'rb') as file:
        cls = pickle.load(file)
    result = ans[cls.predict(vectors)[0]]
    return result

col1, col2, col3 = st.columns([1,5,1])
with col2:

    st.title('Text sentiment analysis')

col1, col2, col3 = st.columns([2,5,2])
with col2:
    
    max_len = st.slider('Maximum word length', 0, 500, 250)

    text_input = st.text_input("Enter some text")
    #model, tokenizer = load_model()

    if text_input:
        input_tokens = preprocess_text(text_input, max_len, tokenizer)
        output = predict_sentiment(model, input_tokens)
        st.write(output)



################################################################################################ 
#st.markdown("<div style='text-align: center; font-size: 25px;'> ", unsafe_allow_html=True)
#st.markdown("<div style='text-align: center; font-size: 25px;'> ", unsafe_allow_html=True)