Spaces:
Runtime error
Runtime error
Create duplex.py
Browse files
duplex.py
ADDED
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
import random
|
4 |
+
import string
|
5 |
+
|
6 |
+
import gradio as gr
|
7 |
+
import torch
|
8 |
+
import requests
|
9 |
+
from transformers import pipeline, set_seed
|
10 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
11 |
+
import logging
|
12 |
+
|
13 |
+
import sys
|
14 |
+
import gradio as gr
|
15 |
+
from transformers import pipeline, AutoModelForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM
|
16 |
+
|
17 |
+
DEBUG = os.environ.get("DEBUG", "false")[0] in "ty1"
|
18 |
+
HF_AUTH_TOKEN = os.environ.get("HF_AUTH_TOKEN", None)
|
19 |
+
MAX_LENGTH = int(os.environ.get("MAX_LENGTH", 1024))
|
20 |
+
|
21 |
+
HEADER = """
|
22 |
+
# Poor Man's Duplex
|
23 |
+
""".strip()
|
24 |
+
|
25 |
+
FOOTER = """
|
26 |
+
<div align=center>
|
27 |
+
<img src="https://visitor-badge.glitch.me/badge?page_id=spaces/bertin-project/bertin-gpt-j-6B"/>
|
28 |
+
<div align=center>
|
29 |
+
""".strip()
|
30 |
+
|
31 |
+
asr_model_name_es = "jonatasgrosman/wav2vec2-large-xlsr-53-spanish"
|
32 |
+
model_instance_es = AutoModelForCTC.from_pretrained(asr_model_name_es)
|
33 |
+
processor_es = Wav2Vec2ProcessorWithLM.from_pretrained(asr_model_name_es)
|
34 |
+
asr_es = pipeline(
|
35 |
+
"automatic-speech-recognition",
|
36 |
+
model=model_instance_es,
|
37 |
+
tokenizer=processor_es.tokenizer,
|
38 |
+
feature_extractor=processor_es.feature_extractor,
|
39 |
+
decoder=processor_es.decoder
|
40 |
+
)
|
41 |
+
tts_model_name = "facebook/tts_transformer-es-css10"
|
42 |
+
speak_es = gr.Interface.load(f"huggingface/{tts_model_name}")
|
43 |
+
transcribe_es = lambda input_file: asr_es(input_file, chunk_length_s=5, stride_length_s=1)["text"]
|
44 |
+
def generate_es(text, **kwargs):
|
45 |
+
# max_length=100, top_k=100, top_p=50, temperature=0.95, do_sample=True, do_clean=True
|
46 |
+
api_uri = "https://hf.space/embed/bertin-project/bertin-gpt-j-6B/+/api/predict/"
|
47 |
+
response = requests.post(api_uri, data=json.dumps({"data": [text, 100, 100, 50, 0.95, True, True]}))
|
48 |
+
if response.ok:
|
49 |
+
print(response.json())
|
50 |
+
return response.json()["data"][0]
|
51 |
+
else:
|
52 |
+
return ""
|
53 |
+
|
54 |
+
asr_model_name_en = "jonatasgrosman/wav2vec2-large-xlsr-53-english"
|
55 |
+
model_instance_en = AutoModelForCTC.from_pretrained(asr_model_name_en)
|
56 |
+
processor_en = Wav2Vec2ProcessorWithLM.from_pretrained(asr_model_name_en)
|
57 |
+
asr_en = pipeline(
|
58 |
+
"automatic-speech-recognition",
|
59 |
+
model=model_instance_en,
|
60 |
+
tokenizer=processor_en.tokenizer,
|
61 |
+
feature_extractor=processor_en.feature_extractor,
|
62 |
+
decoder=processor_en.decoder
|
63 |
+
)
|
64 |
+
tts_model_name = "facebook/fastspeech2-en-200_speaker-cv4"
|
65 |
+
speak_en = gr.Interface.load(f"huggingface/{tts_model_name}")
|
66 |
+
transcribe_en = lambda input_file: asr_en(input_file, chunk_length_s=5, stride_length_s=1)["text"]
|
67 |
+
generate_iface = gr.Interface.load("huggingface/EleutherAI/gpt-j-6B")
|
68 |
+
|
69 |
+
def generate_en(text, **kwargs):
|
70 |
+
response = generate_iface(text)
|
71 |
+
print(response)
|
72 |
+
return response or ""
|
73 |
+
|
74 |
+
|
75 |
+
def select_lang(lang):
|
76 |
+
if lang.lower() == "spanish":
|
77 |
+
return generate_es, transcribe_es, speak_es
|
78 |
+
else:
|
79 |
+
return generate_en, transcribe_en, speak_en
|
80 |
+
|
81 |
+
|
82 |
+
def select_lang_vars(lang):
|
83 |
+
if lang.lower() == "spanish":
|
84 |
+
AGENT = "BERTIN"
|
85 |
+
USER = "ENTREVISTADOR"
|
86 |
+
CONTEXT = """La siguiente conversación es un extracto de una entrevista a {AGENT} celebrada en Madrid para Radio Televisión Española:
|
87 |
+
|
88 |
+
{USER}: Bienvenido, {AGENT}. Un placer tenerlo hoy con nosotros.
|
89 |
+
{AGENT}: Gracias. El placer es mío."""
|
90 |
+
else:
|
91 |
+
AGENT = "ELEUTHER"
|
92 |
+
USER = "INTERVIEWER"
|
93 |
+
CONTEXT = """The next conversation is an excerpt from an interview to {AGENT} that appeared in the New York Times:
|
94 |
+
|
95 |
+
{USER}: Welcome, {AGENT}. It is a pleasure to have you here today.
|
96 |
+
{AGENT}: Thanks. The pleasure is mine."""
|
97 |
+
|
98 |
+
return AGENT, USER, CONTEXT
|
99 |
+
|
100 |
+
|
101 |
+
|
102 |
+
def chat_with_gpt(lang, agent, user, context, audio_in, history):
|
103 |
+
generate, transcribe, speak = select_lang(lang)
|
104 |
+
AGENT, USER, _ = select_lang_vars(lang)
|
105 |
+
user_message = transcribe(audio_in)
|
106 |
+
# agent = AGENT
|
107 |
+
# user = USER
|
108 |
+
generation_kwargs = {
|
109 |
+
"max_length": 25,
|
110 |
+
# "top_k": top_k,
|
111 |
+
# "top_p": top_p,
|
112 |
+
# "temperature": temperature,
|
113 |
+
# "do_sample": do_sample,
|
114 |
+
# "do_clean": do_clean,
|
115 |
+
# "num_return_sequences": 1,
|
116 |
+
# "return_full_text": False,
|
117 |
+
}
|
118 |
+
message = user_message.split(" ", 1)[0].capitalize() + " " + user_message.split(" ", 1)[-1]
|
119 |
+
history = history or [] #[(f"{user}: Bienvenido. Encantado de tenerle con nosotros.", f"{agent}: Un placer, muchas gracias por la invitación.")]
|
120 |
+
context = context.format(USER=user or USER, AGENT=agent or AGENT).strip()
|
121 |
+
if context[-1] not in ".:":
|
122 |
+
context += "."
|
123 |
+
context_length = len(context.split())
|
124 |
+
history_take = 0
|
125 |
+
history_context = "\n".join(f"{user}: {history_message.capitalize()}.\n{agent}: {history_response}." for history_message, history_response in history[-len(history) + history_take:])
|
126 |
+
while len(history_context.split()) > MAX_LENGTH - (generation_kwargs["max_length"] + context_length):
|
127 |
+
history_take += 1
|
128 |
+
history_context = "\n".join(f"{user}: {history_message.capitalize()}.\n{agent}: {history_response}." for history_message, history_response in history[-len(history) + history_take:])
|
129 |
+
if history_take >= MAX_LENGTH:
|
130 |
+
break
|
131 |
+
context += history_context
|
132 |
+
for _ in range(5):
|
133 |
+
response = generate(f"{context}\n\n{user}: {message}.\n", **generation_kwargs)
|
134 |
+
if DEBUG:
|
135 |
+
print("\n-----" + response + "-----\n")
|
136 |
+
response = response.split("\n")[-1]
|
137 |
+
if agent in response and response.split(agent)[-1]:
|
138 |
+
response = response.split(agent)[-1]
|
139 |
+
if user in response and response.split(user)[-1]:
|
140 |
+
response = response.split(user)[-1]
|
141 |
+
if response and response[0] in string.punctuation:
|
142 |
+
response = response[1:].strip()
|
143 |
+
if response.strip().startswith(f"{user}: {message}"):
|
144 |
+
response = response.strip().split(f"{user}: {message}")[-1]
|
145 |
+
if response.replace(".", "").strip() and message.replace(".", "").strip() != response.replace(".", "").strip():
|
146 |
+
break
|
147 |
+
if DEBUG:
|
148 |
+
print()
|
149 |
+
print("CONTEXT:")
|
150 |
+
print(context)
|
151 |
+
print()
|
152 |
+
print("MESSAGE")
|
153 |
+
print(message)
|
154 |
+
print()
|
155 |
+
print("RESPONSE:")
|
156 |
+
print(response)
|
157 |
+
if not response.strip():
|
158 |
+
response = "Lo siento, no puedo hablar ahora" if lang.lower() == "Spanish" else "Sorry, can't talk right now"
|
159 |
+
history.append((user_message, response))
|
160 |
+
return history, history, speak(response)
|
161 |
+
|
162 |
+
|
163 |
+
with gr.Blocks() as demo:
|
164 |
+
gr.Markdown(HEADER)
|
165 |
+
lang = gr.Radio(label="Language", choices=["English", "Spanish"], default="English", type="value")
|
166 |
+
AGENT, USER, CONTEXT = select_lang_vars("English")
|
167 |
+
context = gr.Textbox(label="Context", lines=5, value=CONTEXT)
|
168 |
+
with gr.Row():
|
169 |
+
audio_in = gr.Audio(label="User", source="microphone", type="filepath")
|
170 |
+
audio_out = gr.Audio(label="Agent", interactive=False)
|
171 |
+
# chat_btn = gr.Button("Submit")
|
172 |
+
with gr.Row():
|
173 |
+
user = gr.Textbox(label="User", value=USER)
|
174 |
+
agent = gr.Textbox(label="Agent", value=AGENT)
|
175 |
+
lang.change(select_lang_vars, inputs=[lang], outputs=[agent, user, context])
|
176 |
+
history = gr.Variable(value=[])
|
177 |
+
chatbot = gr.Variable() # gr.Chatbot(color_map=("green", "gray"), visible=False)
|
178 |
+
# chat_btn.click(chat_with_gpt, inputs=[lang, agent, user, context, audio_in, history], outputs=[chatbot, history, audio_out])
|
179 |
+
audio_in.change(chat_with_gpt, inputs=[lang, agent, user, context, audio_in, history], outputs=[chatbot, history, audio_out])
|
180 |
+
gr.Markdown(FOOTER)
|
181 |
+
|
182 |
+
demo.launch()
|