app.py

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import re


model_name = 'armandnlp/gpt2-TOD_finetuned_SGD'
tokenizer_TOD = AutoTokenizer.from_pretrained(model_name)
model_TOD = AutoModelForCausalLM.from_pretrained(model_name)


def generate_response(prompt):
    input_ids = tokenizer_TOD(prompt, return_tensors="pt").input_ids
    outputs = model_TOD.generate(input_ids, 
                                 do_sample=False, 
                                 max_length=1024, 
                                 eos_token_id=50262)
    return tokenizer_TOD.batch_decode(outputs)[0]

#<|context|> <|user|> I want to go to the restaurant.<|endofcontext|>


import gradio as gr

iface = gr.Interface(fn=generate_response,
                     inputs="text",
                     outputs="text",
                     title="gpt2-TOD",
                     examples=[["<|context|> <|user|> I'm super hungry ! I want to go to the restaurant.<|endofcontext|>"],
                                ["<|context|> <|user|> I want to go to the restaurant. <|system|> What food would you like to eat ? <|user|> Italian sounds good. <|endofcontext|>"]],
                     description="Passing in a task-oriented dialogue context generates a belief state, actions to take and a response based on those actions",
                     )

iface.launch()

"""

## Work in progress
## https://gradio.app/creating_a_chatbot/
## make chatbot interface
## can get input and responses for now
## would like to add belief state and actions to history response
## means modifying the history when appending input during next turn
## ie. keeping only the response and adding <|system|> token

ckpt = 'armandnlp/gpt2-TOD_finetuned_SGD'
tokenizer = AutoTokenizer.from_pretrained(ckpt)
model = AutoModelForCausalLM.from_pretrained(ckpt)

def predict(input, history=[]):
  # history: list of all token ids
  # response: list of tuples of strings corresponding to dialogue turns

  #model input and output with extra formatting
  new_user_input_ids = tokenizer.encode(' <|user|> '+input, return_tensors='pt')
  context = tokenizer.encode('<|context|>', return_tensors='pt')
  endofcontext = tokenizer.encode(' <|endofcontext|>', return_tensors='pt')
  model_input = torch.cat([context, torch.LongTensor(history), new_user_input_ids, endofcontext], dim=-1)
  out = model.generate(model_input, max_length=1024, eos_token_id=50262).tolist()[0]

  #history : format for next dialogue turn
  history = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1)
  string_out = tokenizer.decode(out)  
  response_only = string_out.split('<|response|>')[1].replace('<|endofresponse|>', '')
  resp_tokenized = tokenizer.encode(' <|system|> '+response_only, return_tensors='pt')
  history = torch.cat([history, resp_tokenized], dim=-1).tolist()


  # history with belief + action
  # output with belief + action + response

  
  #response: format printed output
  turns = tokenizer.decode(history[0])
  
  #turns  = "<|user|> I want to go to the restaurant. <|system|> What food would you like to eat ? <|user|> Italian sounds good. <|system|> Okay then !"
  turns = re.split('<\|system\|>|<\|user\|>', turns)[1:]
  #print(turns)
  response = [(turns[i], turns[i+1]) for i in range(0, len(turns)-1, 2)]
  #print(response)

  return response, history
#predict("I want to go to the restaurant.")


import gradio as gr
gr.Interface(fn=predict,
             inputs=["text", "state"],
             outputs=["chatbot", "state"]).launch()
"""