Spaces:
Paused
Paused
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
3 |
+
import torch
|
4 |
+
import os
|
5 |
+
import copy
|
6 |
+
import re
|
7 |
+
import secrets
|
8 |
+
from pathlib import Path
|
9 |
+
from pydub import AudioSegment
|
10 |
+
|
11 |
+
# Initialize the model and tokenizer
|
12 |
+
torch.manual_seed(420)
|
13 |
+
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-Audio-Chat", trust_remote_code=True)
|
14 |
+
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-Audio-Chat", device_map="cuda", trust_remote_code=True).eval()
|
15 |
+
|
16 |
+
def _parse_text(text):
|
17 |
+
lines = text.split("\n")
|
18 |
+
lines = [line for line in lines if line != ""]
|
19 |
+
count = 0
|
20 |
+
for i, line in enumerate(lines):
|
21 |
+
if "```" in line:
|
22 |
+
count += 1
|
23 |
+
items = line.split("`")
|
24 |
+
if count % 2 == 1:
|
25 |
+
lines[i] = f'<pre><code class="language-{items[-1]}">'
|
26 |
+
else:
|
27 |
+
lines[i] = f"<br></code></pre>"
|
28 |
+
else:
|
29 |
+
if i > 0:
|
30 |
+
if count % 2 == 1:
|
31 |
+
line = line.replace("`", r"\`")
|
32 |
+
line = line.replace("<", "<")
|
33 |
+
line = line.replace(">", ">")
|
34 |
+
line = line.replace(" ", " ")
|
35 |
+
line = line.replace("*", "*")
|
36 |
+
line = line.replace("_", "_")
|
37 |
+
line = line.replace("-", "-")
|
38 |
+
line = line.replace(".", ".")
|
39 |
+
line = line.replace("!", "!")
|
40 |
+
line = line.replace("(", "(")
|
41 |
+
line = line.replace(")", ")")
|
42 |
+
line = line.replace("$", "$")
|
43 |
+
lines[i] = "<br>" + line
|
44 |
+
text = "".join(lines)
|
45 |
+
return text
|
46 |
+
|
47 |
+
def predict(_chatbot, task_history):
|
48 |
+
if not task_history:
|
49 |
+
return _chatbot
|
50 |
+
|
51 |
+
query = task_history[-1][0]
|
52 |
+
history_cp = copy.deepcopy(task_history)
|
53 |
+
history_filter = []
|
54 |
+
audio_idx = 1
|
55 |
+
pre = ""
|
56 |
+
last_audio = None
|
57 |
+
|
58 |
+
for i, (q, a) in enumerate(history_cp):
|
59 |
+
if isinstance(q, (tuple, list)):
|
60 |
+
last_audio = q[0]
|
61 |
+
q = f'Audio {audio_idx}: <audio>{q[0]}</audio>'
|
62 |
+
pre += q + '\n'
|
63 |
+
audio_idx += 1
|
64 |
+
else:
|
65 |
+
pre += q
|
66 |
+
history_filter.append((pre, a))
|
67 |
+
pre = ""
|
68 |
+
|
69 |
+
history, message = history_filter[:-1], history_filter[-1][0]
|
70 |
+
response, history = model.chat(tokenizer, message, history=history)
|
71 |
+
|
72 |
+
ts_pattern = r"<\|\d{1,2}\.\d+\|>"
|
73 |
+
all_time_stamps = re.findall(ts_pattern, response)
|
74 |
+
if (len(all_time_stamps) > 0) and (len(all_time_stamps) % 2 ==0) and last_audio:
|
75 |
+
ts_float = [ float(t.replace("<|","").replace("|>","")) for t in all_time_stamps]
|
76 |
+
ts_float_pair = [ts_float[i:i + 2] for i in range(0,len(all_time_stamps),2)]
|
77 |
+
# θ―»ει³ι’ζδ»Ά
|
78 |
+
format = os.path.splitext(last_audio)[-1].replace(".","")
|
79 |
+
audio_file = AudioSegment.from_file(last_audio, format=format)
|
80 |
+
chat_response_t = response.replace("<|", "").replace("|>", "")
|
81 |
+
chat_response = chat_response_t
|
82 |
+
temp_dir = secrets.token_hex(20)
|
83 |
+
temp_dir = Path(uploaded_file_dir) / temp_dir
|
84 |
+
temp_dir.mkdir(exist_ok=True, parents=True)
|
85 |
+
# ζͺει³ι’ζδ»Ά
|
86 |
+
for pair in ts_float_pair:
|
87 |
+
audio_clip = audio_file[pair[0] * 1000: pair[1] * 1000]
|
88 |
+
# δΏει³ι’ζδ»Ά
|
89 |
+
name = f"tmp{secrets.token_hex(5)}.{format}"
|
90 |
+
filename = temp_dir / name
|
91 |
+
audio_clip.export(filename, format=format)
|
92 |
+
_chatbot[-1] = (_parse_text(query), chat_response)
|
93 |
+
_chatbot.append((None, (str(filename),)))
|
94 |
+
else:
|
95 |
+
_chatbot[-1] = (_parse_text(query), response)
|
96 |
+
full_response = _parse_text(response)
|
97 |
+
task_history[-1] = (query, full_response)
|
98 |
+
print("Qwen-Audio-Chat: " + _parse_text(full_response))
|
99 |
+
return _chatbot
|
100 |
+
|
101 |
+
def regenerate(_chatbot, task_history):
|
102 |
+
if not task_history:
|
103 |
+
return _chatbot
|
104 |
+
item = task_history[-1]
|
105 |
+
if item[1] is None:
|
106 |
+
return _chatbot
|
107 |
+
task_history[-1] = (item[0], None)
|
108 |
+
chatbot_item = _chatbot.pop(-1)
|
109 |
+
if chatbot_item[0] is None:
|
110 |
+
_chatbot[-1] = (_chatbot[-1][0], None)
|
111 |
+
else:
|
112 |
+
_chatbot.append((chatbot_item[0], None))
|
113 |
+
return predict(_chatbot, task_history)
|
114 |
+
|
115 |
+
def add_text(history, task_history, text):
|
116 |
+
history = history + [(_parse_text(text), None)]
|
117 |
+
task_history = task_history + [(text, None)]
|
118 |
+
return history, task_history, ""
|
119 |
+
|
120 |
+
def add_file(history, task_history, file):
|
121 |
+
history = history + [((file.name,), None)]
|
122 |
+
task_history = task_history + [((file.name,), None)]
|
123 |
+
return history, task_history
|
124 |
+
|
125 |
+
def add_mic(history, task_history, file):
|
126 |
+
if file is None:
|
127 |
+
return history, task_history
|
128 |
+
os.rename(file, file + '.wav')
|
129 |
+
print("add_mic file:", file)
|
130 |
+
print("add_mic history:", history)
|
131 |
+
print("add_mic task_history:", task_history)
|
132 |
+
# history = history + [((file.name,), None)]
|
133 |
+
# task_history = task_history + [((file.name,), None)]
|
134 |
+
task_history = task_history + [((file + '.wav',), None)]
|
135 |
+
history = history + [((file + '.wav',), None)]
|
136 |
+
print("task_history", task_history)
|
137 |
+
return history, task_history
|
138 |
+
|
139 |
+
def reset_user_input():
|
140 |
+
return gr.update(value="")
|
141 |
+
|
142 |
+
def reset_state(task_history):
|
143 |
+
task_history.clear()
|
144 |
+
return []
|
145 |
+
|
146 |
+
iface = gr.Interface(
|
147 |
+
fn=chat_with_model,
|
148 |
+
inputs=[
|
149 |
+
gr.inputs.Audio(label="Audio Input"),
|
150 |
+
gr.inputs.Textbox(label="Text Query"),
|
151 |
+
gr.State()
|
152 |
+
],
|
153 |
+
outputs=[
|
154 |
+
"text",
|
155 |
+
gr.State()
|
156 |
+
],
|
157 |
+
title="Audio-Text Interaction Model",
|
158 |
+
description="This model can process an audio input along with a text query and provide a response.",
|
159 |
+
theme="default",
|
160 |
+
allow_flagging="never"
|
161 |
+
)
|
162 |
+
|
163 |
+
iface.launch()
|