app.py

import streamlit as st
import torch
from prediction_sinhala import MDFEND, TokenizerFromPreTrained  



# Set constants for model and tokenizer paths
MODEL_SAVE_PATH = "models/last-epoch-model-2024-03-08-15_34_03_6.pth"
BERT_MODEL_NAME = 'sinhala-nlp/sinbert-sold-si'
DOMAIN_NUM = 3
MAX_LEN = 160
BATCH_SIZE = 100

# Load model and tokenizer
@st.cache(allow_output_mutation=True)
def load_model():
    # Load the tokenizer from the pre-trained model name
    tokenizer = TokenizerFromPreTrained(MAX_LEN, BERT_MODEL_NAME)
    # Initialize and load the custom model from saved state
    model = MDFEND(BERT_MODEL_NAME, DOMAIN_NUM, expert_num=18, mlp_dims=[5080, 4020, 3010, 2024, 1012, 606, 400])
    model.load_state_dict(torch.load(MODEL_SAVE_PATH, map_location=torch.device('cpu')))
    model.eval()  # Set the model to evaluation mode
    return model, tokenizer

model, tokenizer = load_model()

# User input
text_input = st.text_area("Enter text here:")

# Prediction
if st.button("Predict"):
    if text_input:  # Check if input is not empty
        # Process the input text through the custom tokenizer
        inputs = tokenizer.tokenize(text_input)
        
        # Convert to tensor, add batch dimension, and send to same device as model
        inputs = torch.tensor(inputs).unsqueeze(0).to(model.device)
        
        with torch.no_grad():  # No gradient computation
            # Get model prediction
            output_prob = model.predict(inputs)
            
            # Interpret the output probability
            prediction = 1 if output_prob >= 0.5 else 0
            result = "offensive" if prediction == 1 else "not offensive"
            st.write(f"Prediction: {result}")
    else:
        st.error("Please enter some text to predict.")