pages/1_Descriptive_chatbot.py

import streamlit as st
import torch
from tqdm import tqdm
from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoModelForCausalLM
from transformers import AutoTokenizer
import numpy as np
import time
import string


# JS
import nltk
nltk.download('wordnet')
from nltk.corpus import wordnet as wn
from nltk.tokenize import word_tokenize

@st.cache_resource
def get_models(llama=False):
  st.write('Loading the model...')

  config = PeftConfig.from_pretrained("YouNameIt/T5ForReverseDictionary_prefix_tuned")
  model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-large")
  model = PeftModel.from_pretrained(model, "YouNameIt/T5ForReverseDictionary_prefix_tuned")

  tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large")

  # JS
  if llama:
    model_name = 'meta-llama/Llama-2-7b-chat-hf'
    access_token = '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'
    llama_tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=access_token, use_fast=True)#, use_fast=True)
    llama_model = AutoModelForCausalLM.from_pretrained(model_name, use_auth_token=access_token, device_map={'':0})#, load_in_4bit=True)
    st.write("The assistant is loaded and ready to use!")
    return model, tokenizer, llama_model, llama_tokenizer 
  
  else:
    st.write("_The assistant is loaded and ready to use! :tada:_")
    return model, tokenizer

model, tokenizer = get_models()

def remove_punctuation(word):
    # Create a translation table that maps all punctuation characters to None
    translator = str.maketrans('', '', string.punctuation)
    
    # Use the translate method to remove punctuation from the word
    word_without_punctuation = word.translate(translator)
    
    return word_without_punctuation

def return_top_k(sentence, k=10, word=None, rels=False):

  if sentence[-1] != ".":
    sentence = sentence + "."

  if rels:
    inputs = [f"Description : It is related to '{word}' but not '{word}'. Word : "]
  else:
    inputs = [f"Description : {sentence} Word : "]

  inputs = tokenizer(
      inputs,
      padding=True, truncation=True,
      return_tensors="pt",
  )
  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  model.to(device)

  with torch.no_grad():
    inputs = {k: v.to(device) for k, v in inputs.items()}
    output_sequences = model.generate(input_ids=inputs["input_ids"], max_new_tokens=10, num_beams=k+5, num_return_sequences=k+5, #max_length=3,
                                              top_p = 50, output_scores=True, return_dict_in_generate=True) #repetition_penalty=10000.0

    logits = output_sequences['sequences_scores'].clone().detach()
    decoded_probabilities = torch.softmax(logits, dim=0)


    #all word predictions
    predictions = [tokenizer.decode(tokens, skip_special_tokens=True) for tokens in output_sequences['sequences']]
    probabilities = [round(float(prob), 2) for prob in decoded_probabilities]

    stripped_sent = [remove_punctuation(word.lower()) for word in sentence.split()]
    for pred in predictions:
      if (len(pred) < 2) | (pred in stripped_sent):
        predictions.pop(predictions.index(pred))  
     
  return predictions[:10] 

# JS 
def get_related_words(word, num=5):
  model.eval()
  with torch.no_grad():
      sentence = [f"Descripton : It is related to {word} but not {word}. Word : "]
      #inputs = ["Description: It is something to cut stuff with. Word: "]
      print(sentence)
      inputs = tokenizer(sentence, padding=True, truncation=True, return_tensors="pt",)

      device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
      model.to(device)

      batch = {k: v.to(device) for k, v in inputs.items()}
      beam_outputs = model.generate(
          input_ids=batch['input_ids'], max_new_tokens=10, num_beams=num+2, num_return_sequences=num+2, early_stopping=True
      )
      
      #beam_preds = [tokenizer.decode(beam_output.detach().cpu().numpy(), skip_special_tokens=True) for beam_output in beam_outputs if ]
      beam_preds = []
      for beam_output in beam_outputs:
          prediction = tokenizer.decode(beam_output.detach().cpu().numpy(), skip_special_tokens=True).strip()
          if prediction not in " ".join(sentence):
            beam_preds.append(prediction)
      
      return ", ".join(beam_preds[:num])

#if 'messages' not in st.session_state:

def get_text():
  input_text = st.chat_input()
  return input_text

def write_bot(input, remember=True, blink=True):
    with st.chat_message('assistant'):
      message_placeholder = st.empty()
      full_response = input
      if blink == True:
        response = ''
        for chunk in full_response.split():
          response += chunk + " "
          time.sleep(0.05)
          # Add a blinking cursor to simulate typing
          message_placeholder.markdown(response + "▌") 
        time.sleep(0.5)
      message_placeholder.markdown(full_response)
    if remember == True:
      st.session_state.messages.append({'role': 'assistant', 'content': full_response})
    
#def ask_if_helped():
  #y = st.button('Yes!', key=60)
  #n = st.button('No...', key=61)
  #new = st.button('I have a new word', key=62)
  #if y:
  #  write_bot("I am happy to help!")
  #  again = st.button('Play again')
  #  if again:
  #    write_bot("Please describe your word!")
  #  st.session_state.is_helpful['ask'] = False
  #elif n:
  #  st.session_state.actions.append('cue')
  #  st.session_state.is_helpful['ask'] = False
  #  #cue_generation()
  #elif new:
  #  write_bot("Please describe your word!")
  #  st.session_state.is_helpful['ask'] = False

## removed: if st.session_state.actions[-1] == "result":

# JS
def get_related_words_llama(relation, target, device, num=5):
    prompt = f"Provide {num} {relation}s for the word '{target}'. Your answer consists of these {num} words only. Do not include the word '{target}' itself in your answer"

    inputs = tokenizer([prompt], return_tensors='pt').to(device)
    output = model.generate(
        **inputs, max_new_tokens=40, temperature=.75, early_stopping=True, 
        )
    chatbot_response = tokenizer.decode(output[:, inputs['input_ids'].shape[-1]:][0], skip_special_tokens=True).strip()

    postproc = [word for word in word_tokenize(chatbot_response) if len(word)>=3]

    return postproc[-num:] if len(postproc)>=num else postproc


def postproc_wn(related_words, syns=False):
    if syns:
        related_words = [word.split('.')[0] if word[0] != "." else word.split('.')[1] for word in related_words]
    else:
        related_words = [word.name().split('.')[0] if word.name()[0] != "." else word.name().split('.')[1] for word in related_words]
    related_words = [word.replace("_", " ") for word in related_words]

    return related_words

# JS
def get_available_cues(target):
    wn_nouns = [word.name() for word in wn.all_synsets(pos='n')]
    wn_nouns = [word.split('.')[0] if word[0] != "." else word.split('.')[1] for word in wn_nouns]

    if target in wn_nouns:
      available_cues = {}
      synset_target = wn.synsets(target, pos=wn.NOUN)[0]

      #if wn.synonyms(target)[0]:
       #   available_cues['Synonyms'] = postproc_wn(wn.synonyms(target)[0], syns=True)
      
      #if synset_target.hypernyms():
       #   available_cues['Hypernyms'] = postproc_wn(synset_target.hypernyms())


      #if synset_target.hyponyms():
        #  available_cues['Hyponyms'] = postproc_wn(synset_target.hyponyms())

      if synset_target.examples():
          examples = []
          
          for example in synset_target.examples():
              examples.append(example.replace(target, "..."))
                  
          available_cues['Examples'] = examples

      return available_cues

    else:
      return None 

# JS: moved the cue generation further down
#def cue_generation():
 # if st.session_state.actions[-1] == 'cue':

if 'messages' not in st.session_state:
  st.session_state.messages = []
  
if 'results' not in st.session_state:
  st.session_state.results = {'results': False, 'results_print': False}
  
if 'actions' not in st.session_state:
  st.session_state.actions = [""]
  
if 'counters' not in st.session_state:
  st.session_state.counters = {"letter_count": 0, "word_count": 0}
  
if 'is_helpful' not in st.session_state:
  st.session_state.is_helpful = {'ask':False}
  
if 'descriptions' not in st.session_state:
  st.session_state.descriptions = []
  
st.title("You name it! 🗣")

with st.chat_message('user', avatar='nursulu.jpg'):
  st.write("Hey assistant!")
  
bot = st.chat_message('assistant')
bot.write("Hello human! Wanna practice naming some words?")

#for showing history of messages
for message in st.session_state.messages:
  if message['role'] == 'user':
    with st.chat_message(message['role'],  avatar='nursulu.jpg'):
      st.markdown(message['content'])
  else:
    with st.chat_message(message['role']):
      st.markdown(message['content'])
  
#display user message in chat message container
prompt = get_text() 
if prompt:
  with st.chat_message('user',  avatar='nursulu.jpg'):
    st.markdown(prompt)
  #add to history
  st.session_state.messages.append({'role': 'user', 'content': prompt})
  #TODO: replace it with zero-shot classifier
  yes = ['yes', 'again', 'sure', 'new word', 'yes!', 'yep', 'yeah']
  no = ['no', 'nope', 'nah']
  try:
    if prompt.lower() in yes:
      write_bot("Please describe your word!")
    elif prompt.lower() in no:
      write_bot("Okay, see you next time then! :innocent:")
    elif prompt == 'it is similar to the best place on earth':
      write_bot("Great! Let me think what it could be...")
      time.sleep(3)
      write_bot("Do you mean Saarland?")
      #if previously we asked to give a prompt
    elif (st.session_state.messages[-2]['content'] == "Please describe your word!") & (st.session_state.messages[-1]['content'] != "no"):
      write_bot("Great! Let me think what it could be...")
      st.session_state.descriptions.append(prompt)
      st.session_state.results['results'] = return_top_k(st.session_state.descriptions[-1])
      st.session_state.results['results_print'] = dict(zip(range(1, 11), st.session_state.results['results']))
      write_bot("I think I have some ideas. Do you want to see my guesses or do you want a cue?")
      st.session_state.actions.append("result")
  except:
    write_bot("Sorry, I didn't understand you... I am still learning :sob: For now, could you respond with 'yes' or 'no'? ")
  
if st.session_state.actions[-1] == "result":
  col1, col2, col3, col4, col5 = st.columns(5)
  with col1:
    a1 = st.button('Results', key=10)
  with col2:
    a2 = st.button('Cue', key=11)
  if a1:
    write_bot("Here are my guesses about your word:")
    st.write(st.session_state.results['results_print'])
    time.sleep(1)
    write_bot('Does it help you remember the word?', remember=False)
    st.session_state.is_helpful['ask'] = True
  elif a2:
    #write_bot(f'The first letter is {st.session_state.results["results"][0][0]}.')
    #time.sleep(1)
    st.session_state.actions.append('cue')
    #cue_generation()
    #write_bot('Does it help you remember the word?', remember=False)
    #st.session_state.is_helpful['ask'] = True

if st.session_state.is_helpful['ask']:
  y = st.button('Yes!', key=60)
  n = st.button('No...', key=61)
  new = st.button('I have a new word', key=62)
  if y:
    write_bot("I am happy to help!")
    again = st.button('Play again')
    if again:
      write_bot("Please describe your word!")
    st.session_state.is_helpful['ask'] = False
  elif n:
    st.session_state.is_helpful['ask'] = False
    st.session_state.actions.append('cue')
    #cue_generation()
  elif new:
    write_bot("Please describe your word!")
    st.session_state.is_helpful['ask'] = False

if st.session_state.actions[-1] == 'cue':
  guessed = False
  write_bot('What do you want to see?', remember=False, blink=False)

  while guessed == False:
    # JS
    word_count = st.session_state.counters["word_count"]
    target = st.session_state.results["results"][word_count]
    
    col1, col2, col3, col4, col5 = st.columns(5)
  

    with col1:
      b1 = st.button("Next letter", key="1")
    with col2:
      b2 = st.button("Related words")
    with col3:
      b3 = st.button("Next word", key="2")
    with col4:
      b4 = st.button("All words", key="3")

    # JS
    #if get_available_cues(target):
     # avail_cues = get_available_cues(target)
      #cues_buttons = {cue_type: st.button(cue_type) for cue_type in avail_cues}

    b5 = st.button("I remembered the word!", key="4", type='primary')
    b6 = st.button("Exit", key="5", type='primary')
    new = st.button('Play again', key=64, type='primary')

    if b1:
      st.session_state.counters["letter_count"] += 1
      #word_count = st.session_state.counters["word_count"]
      letter_count = st.session_state.counters["letter_count"]
      if letter_count < len(target):
        write_bot(f'The word starts with {st.session_state.results["results"][word_count][:letter_count]}.', remember=False)
        #ask_if_helped()
        st.session_state.is_helpful['ask'] = True
      else:
        write_bot(f'This is my predicted word: "{target}". Does this match your query?')
        #ask_if_helped()
        st.session_state.is_helpful['ask'] = True

    elif b2:
      rels = return_top_k(st.session_state.descriptions[-1], word=target, rels=True)
      write_bot(f'Here are words that are related to your word: {", ".join(rels)}.', remember=False)
      #ask_if_helped()
      st.session_state.is_helpful['ask'] = True

    elif b3:
      st.session_state.counters["letter_count"] = 1
      letter_count = st.session_state.counters["letter_count"]
      st.session_state.counters["word_count"] += 1
      word_count = st.session_state.counters["word_count"]
      #write_bot(f'The next word starts with {st.session_state.results["results"][word_count][:letter_count]}', remember=False)
      if letter_count < len(target):
        write_bot(f'The next word starts with {st.session_state.results["results"][word_count][:letter_count]}.', remember=False)
        #ask_if_helped()
        st.session_state.is_helpful['ask'] = True
      else:
        write_bot(f'This is my predicted word: "{target}". Does this match your query?')
        #ask_if_helped()
        st.session_state.is_helpful['ask'] = True

    #elif get_available_cues(target) and "Synonyms" in cues_buttons and cues_buttons['Synonyms']:
      #write_bot(f'Here are synonyms for the current word: {", ".join(avail_cues["Synonyms"])}', remember=False)

    #elif get_available_cues(target) and "Hypernyms" in cues_buttons and cues_buttons['Hypernyms']:
      #write_bot(f'Here are hypernyms for the current word: {", ".join(avail_cues["Hypernyms"])}', remember=False)

    #elif get_available_cues(target) and "Hyponyms" in cues_buttons and cues_buttons['Hyponyms']:
      #write_bot(f'Here are hyponyms for the current word: {", ".join(avail_cues["Hyponyms"])}', remember=False)

    #elif get_available_cues(target) and "Examples" in cues_buttons and cues_buttons['Examples']:   
      #write_bot(f'Here are example contexts for the current word: {", ".join(avail_cues["Examples"])}', remember=False)

    elif b4:
      write_bot(f"Here are all my guesses about your word: {st.session_state.results['results_print']}")

    elif b5:
      write_bot("Yay! I am happy I could be of help!")
      st.session_state.counters["word_count"] = 0
      st.session_state.counters["letter_count"] = 0
      new = st.button('Play again', key=63)
      if new:
        write_bot("Please describe your word!")
      guessed = True
    
      break
    
    elif b6:
      write_bot("I am sorry I couldn't help you this time. See you soon!")
      st.session_state.counters["word_count"] = 0
      st.session_state.counters["letter_count"] = 0
    st.session_state.actions.append('cue')
    
    if new:
      write_bot("Please describe your word!")
      st.session_state.counters["word_count"] = 0
      st.session_state.counters["letter_count"] = 0

    break