app.py

import streamlit as st
from transformers import pipeline
from textblob import TextBlob
from transformers import BertForSequenceClassification, AdamW, BertConfig
st.set_page_config(layout='wide', initial_sidebar_state='expanded')
col1, col2= st.columns(2)
placeholder = st.empty()
placeholder2 = st.empty()
with col1:
    st.title("Spamd: Turkish Spam Detector")
    st.markdown("Message spam detection tool for Turkish language. Due the small size of the dataset, I decided to go with transformers         technology Google BERT. Using the Turkish pre-trained model BERTurk, I imporved the accuracy of the tool by 18 percent compared to the previous model which used fastText.")
with col2:
        text = placeholder.text_input("Enter the text you'd like to analyze for spam.", disabled=True, key="1")
        aButton = placeholder2.button('Analyze', disabled=True, key="1")
if st.button('Load Model', disabled=False):   
    with st.spinner('Wait for it...'):
        import torch
        import numpy as np
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained("dbmdz/bert-base-turkish-uncased")
        from transformers import AutoModel
        model = BertForSequenceClassification.from_pretrained("NimaKL/spamd_model")
        
        token_id = []
        attention_masks = []
        def preprocessing(input_text, tokenizer):
            '''
                  Returns <class transformers.tokenization_utils_base.BatchEncoding> with the following fields:
                    - input_ids: list of token ids
                    - token_type_ids: list of token type ids
                    - attention_mask: list of indices (0,1) specifying which tokens should considered by the model (return_attention_mask = True).
            '''
            return tokenizer.encode_plus(
                input_text,
                add_special_tokens = True,
            max_length = 32,
            pad_to_max_length = True,
            return_attention_mask = True,
            return_tensors = 'pt'
            )
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        with col1:
            st.success("Model Loaded!")
        def predict(new_sentence):
            # We need Token IDs and Attention Mask for inference on the new sentence
            test_ids = []
            test_attention_mask = []
            # Apply the tokenizer
            encoding = preprocessing(new_sentence, tokenizer)
            #Extract IDs and Attention Mask
            test_ids.append(encoding['input_ids'])
            test_attention_mask.append(encoding['attention_mask'])
            test_ids = torch.cat(test_ids, dim = 0)
            test_attention_mask = torch.cat(test_attention_mask, dim = 0)
            #Forward pass, calculate logit predictions
            with torch.no_grad():
                output = model(test_ids.to(device), token_type_ids = None, attention_mask = test_attention_mask.to(device))
                prediction = 'Spam' if np.argmax(output.logits.cpu().numpy()).flatten().item() == 1 else 'Normal'
                pred = 'Predicted Class: '+ prediction
                return pred
        placeholder.text_input("Enter the text you'd like to analyze for spam.", disabled=False, key="2")
        placeholder2.button('Analyze', disabled=False, key="2") 
if not flag:       
    with col2:
        if text or aButton:
            st.header(predict(text))