app.py

import torch
from transformers import BertTokenizer, BertModel
from huggingface_hub import PyTorchModelHubMixin
import numpy as np
import gradio as gr
import nltk
nltk.download('stopwords')
from nltk.corpus import stopwords
import re

device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
device

class BERTClass(torch.nn.Module, PyTorchModelHubMixin):
    def __init__(self):
        super(BERTClass, self).__init__()
        self.bert_model = BertModel.from_pretrained('digitalepidemiologylab/covid-twitter-bert-v2', return_dict=True)
        self.dropout = torch.nn.Dropout(0.3)
        self.linear = torch.nn.Linear(1024, 11)

    def forward(self, input_ids, attn_mask, token_type_ids):
        output = self.bert_model(
            input_ids,
            attention_mask=attn_mask,
            token_type_ids=token_type_ids
        )
        output_dropout = self.dropout(output.pooler_output)
        output = self.linear(output_dropout)
        return output

model = BERTClass()

model = model.from_pretrained("Asutosh2003/ct-bert-v2-vaccine-concern")
model.to(device)

tokenizer = BertTokenizer.from_pretrained('digitalepidemiologylab/covid-twitter-bert-v2')
MAX_LEN = 256


def rmTrash(raw_string, remuser, remstop, remurls):
    final_string = ""
    raw_string_2 = ""
    if remuser == True:
      for i in raw_string.split():
          if '@' not in i:
              raw_string_2 += ' ' + i
    else:
      raw_string_2 = raw_string
    raw_string_2 = re.sub(r'[^\w\s]', '', raw_string_2.lower())
    if remurls == True:
      raw_string_2 = re.sub(r'http\S+', '', raw_string_2.lower())
    if remstop == True:
      raw_string_tokens = raw_string_2.split()
      for token in raw_string_tokens:
          if (not(token in stopwords.words('english'))):
              final_string = final_string + ' ' + token
    else:
       final_string = raw_string_2
    return final_string


def return_vec(text):
    text = rmTrash(text,True,True,True)
    encodings = tokenizer.encode_plus(
      text,
      None,
      add_special_tokens=True,
      max_length=MAX_LEN,
      padding='max_length',
      return_token_type_ids=True,
      truncation=True,
      return_attention_mask=True,
      return_tensors='pt'
    )
    model.eval()
    with torch.no_grad():
      input_ids = encodings['input_ids'].to(device, dtype=torch.long)
      attention_mask = encodings['attention_mask'].to(device, dtype=torch.long)
      token_type_ids = encodings['token_type_ids'].to(device, dtype=torch.long)
      output = model(input_ids, attention_mask, token_type_ids)
      final_output = torch.sigmoid(output).cpu().detach().numpy().tolist()
    return list(final_output[0])


def filter_threshold_lst(vector, threshold_list):
    optimized_vector = []
    optimized_vector = [1 if val >= threshold else 0 for val, threshold in zip(vector, threshold_list)]
    optimized_vector.append(optimized_vector)

    return optimized_vector


def predict(text, threshold_lst):
  pred_lbl_lst = []
  labels = ('side-effect', 'ineffective', 'rushed', 'pharma', 'mandatory', 'unnecessary', 'political', 'ingredients', 'conspiracy', 'country', 'religious')
  prob_lst = return_vec(text)
  vec = filter_threshold_lst(prob_lst, threshold_lst)
  if vec[:11] == [0] * 11:
    pred_lbl_lst = ['none']
    vec = [0] * 11
    vec.append(1)
    return pred_lbl_lst, prob_lst
  for i in range(len(vec)):
    if vec[i] == 1:
      pred_lbl_lst.append(labels[i])
  return pred_lbl_lst, prob_lst

def gr_predict(text):
  thres = [0.616, 0.212, 0.051, 0.131, 0.212, 0.111, 0.071, 0.566, 0.061, 0.02, 0.081]
  out_lst, _ = predict(text,thres)
  out_str = ''
  for lbl in out_lst:
    out_str += lbl + ','
  out_str = out_str[:-1]

  return out_str

descr = """

This app uses [Covid-twitter-BERT-v2](https://huggingface.co/digitalepidemiologylab/covid-twitter-bert-v2)
fine tuned on a custom subset of [Caves dataset](https://arxiv.org/abs/2204.13746) sent by [FIRE 2023](http://fire.irsi.res.in/fire/2023/home)
conference to do multi-label classification of tweets expressing concerns towards vaccines. The different concerns/classes are
('side-effect', 'ineffective', 'rushed', 'pharma', 'mandatory', 'unnecessary', 'political', 'ingredients', 'conspiracy', 'country', 'religious'). 
Each tweet can be expressing multiple of these concerns. If a tweet is not expressing any concern falling into any of these categories 
it will be classified as 'None'.\n
[Source files](https://github.com/Ranjit246/AISoME_FIRE_2023)\n
Try it out with some ridiculous statements about vaccines. You can use the examples below as a start.


"""
# Gradio Interface
iface = gr.Interface(
    fn=gr_predict,
    inputs=gr.Textbox(),
    outputs=gr.Label(),  # Use Label widget for output
    examples=["This vaccine gave me mumps", "Chinese vaccine will infect our brain",
              "Trump is gonna use these vaccines to control us and become the president"],
    title="Vaccine Concerns ML",
    description=descr
)
# Launch the Gradio app
iface.launch(debug=True)