GPT2_Chatbot / app.py
gabriel lopez
working prototype
38fdc49
raw
history blame
2.28 kB
from transformers import TFAutoModelForCausalLM, AutoTokenizer
import tensorflow as tf
import gradio as gr
TITLE = "DialoGPT -- Chatbot"
DESCRIPTION = "<center>Have funny/existencial dialogs with non-human entities</center>"
EXAMPLES = [
["How will the world end?"],
["Does the universe have a purpose?"],
["Is the universe infinite?"],
["Was Einstein right about time being relative?"],
["What is Pythagoras theorem?"],
["What is the meaning of life?"],
]
ARTICLE = r"""<center>
This application allowsa you to talk with a machine.
In the back-end I'm using the DialoGPT model from microsoft.<br>
This model extends GPT2 towards the conversational neural response generetion domain.<br>
ArXiv paper: https://arxiv.org/abs/1911.00536<br>
Done by dr. Gabriel Lopez<br>
For more please visit: <a href='https://sites.google.com/view/dr-gabriel-lopez/home'>My Page</a><br>
</center>"""
tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-medium")
model = TFAutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-medium")
def chat_with_bot(user_input, chat_history_and_input=[]):
emb_user_input = tokenizer.encode(
user_input + tokenizer.eos_token, return_tensors="tf"
)
print("chat_history:", chat_history_and_input)
print("emb_user_input:", emb_user_input)
if chat_history_and_input == []:
bot_input_ids = emb_user_input # first iteration
else:
bot_input_ids = tf.concat(
[chat_history_and_input, emb_user_input], axis=-1
) # other iterations
chat_history_and_input = model.generate(
bot_input_ids, max_length=1000, pad_token_id=tokenizer.eos_token_id
).numpy()
# print
bot_response = tokenizer.decode(
chat_history_and_input[:, bot_input_ids.shape[-1] :][0],
skip_special_tokens=True,
)
print(f"{bot_response=}")
print(f"{chat_history_and_input=}")
return bot_response, chat_history_and_input
gr.Interface(
inputs=["text", "state"],
outputs=["text", "state"],
examples=EXAMPLES,
title=TITLE,
description=DESCRIPTION,
article=ARTICLE,
fn=chat_with_bot,
allow_flagging=False,
).launch()