gorkemgoknar's picture
Update app.py
392fb0c
raw
history blame
8.61 kB
import gradio as gr
import random
import torch
from transformers import AutoConfig
from transformers import GPT2Tokenizer, GPT2LMHeadModel
from itertools import chain
import tempfile
from typing import Optional
from TTS.config import load_config
import gradio as gr
import numpy as np
from TTS.utils.manage import ModelManager
from TTS.utils.synthesizer import Synthesizer
config = AutoConfig.from_pretrained('gorkemgoknar/gpt2chatbotenglish')
model = GPT2LMHeadModel.from_pretrained('gorkemgoknar/gpt2chatbotenglish', config=config)
tokenizer = GPT2Tokenizer.from_pretrained('gorkemgoknar/gpt2chatbotenglish')
tokenizer.model_max_length = 1024
#Dynamic Temperature
#See experiment https://www.linkedin.com/pulse/ai-goes-job-interview-g%25C3%25B6rkem-g%25C3%25B6knar
base_temperature = 1.3
dynamic_temperature_range = 0.15
rand_range = random.uniform(-1 * dynamic_temperature_range , dynamic_temperature_range )
temperature = base_temperature + rand_range
SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"]
#See document for experiment https://www.linkedin.com/pulse/ai-goes-job-interview-g%C3%B6rkem-g%C3%B6knar/
def get_chat_response(name,history=[], input_txt = "Hello , what is your name?"):
ai_history = history.copy()
#ai_history.append(input_txt)
ai_history_e = [tokenizer.encode(e) for e in ai_history]
personality = "My name is " + name
bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:-1])
#persona first, history next, input text must be at the end
#[[bos, persona] , [history] , [input]]
sequence = [[bos] + tokenizer.encode(personality)] + ai_history_e + [tokenizer.encode(input_txt)]
##[[bos, persona] , [speaker1 .., speakser2 .., speaker1 ... speaker2 ... , [input]]
sequence = [sequence[0]] + [[speaker2 if (len(sequence)-i) % 2 else speaker1] + s for i, s in enumerate(sequence[1:])]
sequence = list(chain(*sequence))
#bot_input_ids = tokenizer.encode(personality + tokenizer.eos_token + input_txt + tokenizer.eos_token , return_tensors='pt')
sequence_len = len(sequence)
#optimum response and speed
chat_history_ids = model.generate(
torch.tensor(sequence).unsqueeze(0), max_length=50,
pad_token_id=tokenizer.eos_token_id,
no_repeat_ngram_size=3,
do_sample=True,
top_k=60,
top_p=0.8,
temperature = 1.3
)
out_str = tokenizer.decode(chat_history_ids[0][sequence_len:], skip_special_tokens=True)
#out_str = tokenizer.decode(chat_history_ids[:, sequence.shape[-1]:][0], skip_special_tokens=False)
return out_str
##you can use anyone from below
'''
| Macleod | Moran | Brenda | Ramirez | Peter Parker | Quentin Beck | Andy
| Red | Norton | Willard | Chief | Chef | Kilgore | Kurtz | Westley | Buttercup
| Vizzini | Fezzik | Inigo | Man In Black | Taylor | Zira | Zaius | Cornelius
| Bud | Lindsey | Hippy | Erin | Ed | George | Donna | Trinity | Agent Smith
| Morpheus | Neo | Tank | Meryl | Truman | Marlon | Christof | Stromboli | Bumstead
| Schreber | Walker | Korben | Cornelius | Loc Rhod | Anakin | Obi-Wan | Palpatine
| Padme | Superman | Luthor | Dude | Walter | Donny | Maude | General | Starkiller
| Indiana | Willie | Short Round | John | Sarah | Terminator | Miller | Sarge | Reiben
| Jackson | Upham | Chuckie | Will | Lambeau | Sean | Skylar | Saavik | Spock
| Kirk | Bones | Khan | Kirk | Spock | Sybok | Scotty | Bourne | Pamela | Abbott
| Nicky | Marshall | Korshunov | Troy | Vig | Archie Gates | Doc | Interrogator
| Ellie | Ted | Peter | Drumlin | Joss | Macready | Childs | Nicholas | Conrad
| Feingold | Christine | Adam | Barbara | Delia | Lydia | Cathy | Charles | Otho
| Schaefer | Han | Luke | Leia | Threepio | Vader | Yoda | Lando | Elaine | Striker
| Dr. Rumack | Kramer | David | Saavik | Kirk | Kruge | Holden | Deckard | Rachael
| Batty | Sebastian | Sam | Frodo | Pippin | Gandalf | Kay | Edwards | Laurel
| Edgar | Zed | Jay | Malloy | Plissken | Steve Rogers | Tony Stark | Scott Lang
| Bruce Banner | Bruce | Edward | Two-Face | Batman | Chase | Alfred | Dick
| Riddler | Din Djarin | Greef Karga | Kuiil | Ig-11 | Cara Dune | Peli Motto
| Toro Calican | Ripley | Meredith | Dickie | Marge | Peter | Lambert | Kane
| Dallas | Ripley | Ash | Parker | Threepio | Luke | Leia | Ben | Han | Common Bob
| Common Alice | Jack | Tyler | Marla | Dana | Stantz | Venkman | Spengler | Louis
| Fry | Johns | Riddick | Kirk | Decker | Spock | "Ilia | Indy | Belloq | Marion
| Brother | Allnut | Rose | Qui-Gon | Jar Jar
'''
MODEL_NAME= "tts_models/multilingual/multi-dataset/your_tts"
def tts(text: str, speaker_idx: str=None):
if len(text) > MAX_TXT_LEN:
text = text[:MAX_TXT_LEN]
print(f"Input text was cutoff since it went over the {MAX_TXT_LEN} character limit.")
print(text, model_name)
# download model
model_path, config_path, model_item = manager.download_model(f"tts_models/{MODEL_NAME}")
vocoder_name: Optional[str] = model_item["default_vocoder"]
# download vocoder
vocoder_path = None
vocoder_config_path = None
if vocoder_name is not None:
vocoder_path, vocoder_config_path, _ = manager.download_model(vocoder_name)
# init synthesizer
synthesizer = Synthesizer(
model_path, config_path, None, None, vocoder_path, vocoder_config_path,
)
# synthesize
if synthesizer is None:
raise NameError("model not found")
wavs = synthesizer.tts(text, speaker_idx)
# return output
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
synthesizer.save_wav(wavs, fp)
return fp.name
def greet(character,message,history,voice):
#gradios set_state/get_state had problems on embedded html!
history = history or {"character": character, "message_history" : [] }
#gradios set_state/get_state does not persist session for now using global
#global history
if history["character"] != character:
#switching character
history = {"character": character, "message_history" : [] }
response = get_chat_response(character,history=history["message_history"],input_txt=message)
voice = tts(response)
history["message_history"].append((message, response))
html = "<div class='chatbot'>"
for user_msg, resp_msg in history["message_history"]:
html += f"<div class='user_msg'>You: {user_msg}</div>"
html += f"<div class='resp_msg'>{character}: {resp_msg}</div>"
html += "</div>"
return html,history,voice
personality_choices = ["Gandalf", "Riddick", "Macleod", "Morpheus", "Neo","Spock","Vader","Indy"]
examples= ["Gandalf", "What is your name?"]
css="""
.chatbox {display:flex;flex-direction:column}
.user_msg, .resp_msg {padding:4px;margin-bottom:4px;border-radius:4px;width:80%}
.user_msg {background-color:cornflowerblue;color:white;align-self:start}
.resp_msg {background-color:lightgray;align-self:self-end}
"""
#some selected ones are in for demo use
personality_choices = ["Gandalf", "Riddick", "Macleod", "Morpheus", "Neo","Spock","Vader","Indy", "Ig-11","Threepio","Tony Stark","Batman","Vizzini"]
title = "Metayazar - Movie Chatbot"
description = "Chat with your favorite movie characters. This space demo has simple interface and simple history as gradio's state did not work need to make a global history, you will have different responses as same gradio machine will be used! Test it out in metayazar.com/chatbot for more movie/character options and history memorized."
article = "<p style='text-align: center'><a href='https://www.linkedin.com/pulse/ai-goes-job-interview-g%C3%B6rkem-g%C3%B6knar/' target='_blank'>AI Goes to Job Interview</a> | <a href='https://www.metayazar.com/' target='_blank'>Metayazar AI Writer</a> |<a href='https://www.linkedin.com/in/goknar/' target='_blank'>Görkem Göknar</a></p>"
#History not implemented in this demo, use metayazar.com/chatbot for a movie and character dropdown chat interface
##interface = gr.Interface(fn=greet, inputs=[gr.inputs.Dropdown(personality_choices) ,"text"], title=title, description=description, outputs="text")
history = {"character": "None", "message_history" : [] }
interface= gr.Interface(fn=greet,
inputs=[gr.inputs.Dropdown(personality_choices) ,"text", "state"],
outputs=["html","state", gr.outputs.Audio(label="voice"),],
css=css, title=title, description=description,article=article )
if __name__ == "__main__":
interface.launch()