Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -15,11 +15,14 @@ import torch
|
|
15 |
import numpy as np
|
16 |
from loguru import logger
|
17 |
from PIL import Image
|
18 |
-
from transformers import AutoProcessor, Gemma3ForConditionalGeneration, TextIteratorStreamer,
|
19 |
import time
|
20 |
import warnings
|
21 |
from typing import Dict, List, Optional, Union
|
22 |
import librosa
|
|
|
|
|
|
|
23 |
|
24 |
# CSV/TXT ๋ถ์
|
25 |
import pandas as pd
|
@@ -28,7 +31,7 @@ import PyPDF2
|
|
28 |
|
29 |
warnings.filterwarnings('ignore')
|
30 |
|
31 |
-
print("๐ฎ ๋ก๋ด ์๊ฐ ์์คํ
์ด๊ธฐํ (Gemma3-R1984-4B + Whisper)...")
|
32 |
|
33 |
##############################################################################
|
34 |
# ์์ ์ ์
|
@@ -44,7 +47,6 @@ SERPHOUSE_API_KEY = os.getenv("SERPHOUSE_API_KEY", "")
|
|
44 |
model = None
|
45 |
processor = None
|
46 |
whisper_model = None
|
47 |
-
whisper_processor = None
|
48 |
model_loaded = False
|
49 |
whisper_loaded = False
|
50 |
model_name = "Gemma3-R1984-4B"
|
@@ -63,7 +65,7 @@ def clear_cuda_cache():
|
|
63 |
##############################################################################
|
64 |
@spaces.GPU(duration=60)
|
65 |
def load_whisper():
|
66 |
-
global whisper_model,
|
67 |
|
68 |
if whisper_loaded:
|
69 |
logger.info("Whisper ๋ชจ๋ธ์ด ์ด๋ฏธ ๋ก๋๋์ด ์์ต๋๋ค.")
|
@@ -71,11 +73,13 @@ def load_whisper():
|
|
71 |
|
72 |
try:
|
73 |
logger.info("Whisper ๋ชจ๋ธ ๋ก๋ฉ ์์...")
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
|
|
|
|
79 |
)
|
80 |
whisper_loaded = True
|
81 |
logger.info("โ
Whisper ๋ชจ๋ธ ๋ก๋ฉ ์๋ฃ!")
|
@@ -87,42 +91,141 @@ def load_whisper():
|
|
87 |
##############################################################################
|
88 |
# ์ค๋์ค ์ฒ๋ฆฌ ํจ์
|
89 |
##############################################################################
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
90 |
@spaces.GPU(duration=30)
|
91 |
-
def
|
92 |
"""Whisper๋ฅผ ์ฌ์ฉํ ์ค๋์ค ์ ์ฌ"""
|
93 |
-
global whisper_model,
|
94 |
|
95 |
if not whisper_loaded:
|
96 |
if not load_whisper():
|
97 |
-
return
|
98 |
|
99 |
try:
|
100 |
-
if audio_data is None:
|
101 |
-
return None
|
102 |
-
|
103 |
-
# ์ค๋์ค ๋ฐ์ดํฐ ์ฒ๋ฆฌ
|
104 |
-
sample_rate, audio = audio_data
|
105 |
-
|
106 |
-
# 16kHz๋ก ๋ฆฌ์ํ๋ง
|
107 |
-
if sample_rate != 16000:
|
108 |
-
audio = librosa.resample(audio.astype(float), orig_sr=sample_rate, target_sr=16000)
|
109 |
-
|
110 |
-
# Whisper ์
๋ ฅ ์ฒ๋ฆฌ
|
111 |
-
inputs = whisper_processor(audio, sampling_rate=16000, return_tensors="pt")
|
112 |
-
inputs = {k: v.to(whisper_model.device) for k, v in inputs.items()}
|
113 |
-
|
114 |
# ์์ฑ ์ธ์
|
115 |
-
|
116 |
-
|
117 |
|
118 |
-
|
119 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
|
121 |
-
|
|
|
|
|
|
|
|
|
122 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
123 |
except Exception as e:
|
124 |
-
logger.error(f"์ค๋์ค
|
125 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
|
127 |
##############################################################################
|
128 |
# ํค์๋ ์ถ์ถ ํจ์
|
@@ -591,13 +694,19 @@ css = """
|
|
591 |
background: #e3f2fd;
|
592 |
color: #1565c0;
|
593 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
594 |
"""
|
595 |
|
596 |
with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as demo:
|
597 |
gr.HTML("""
|
598 |
<div class="robot-header">
|
599 |
<h1>๐ค ๋ก๋ด ์๊ฐ ์์คํ
</h1>
|
600 |
-
<h3>๐ฎ Gemma3-R1984-4B + ๐ท ์ค์๊ฐ ์น์บ + ๐ค ์์ฑ ์ธ์</h3>
|
601 |
<p>โก ๋ฉํฐ๋ชจ๋ฌ AI๋ก ๋ก๋ด ์์
๋ถ์!</p>
|
602 |
</div>
|
603 |
""")
|
@@ -636,6 +745,11 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
636 |
'<div class="audio-status">๐ค ์์ฑ ์ธ์: ๋นํ์ฑํ</div>'
|
637 |
)
|
638 |
|
|
|
|
|
|
|
|
|
|
|
639 |
# ๋ง์ง๋ง ์ธ์๋ ํ
์คํธ
|
640 |
last_transcript = gr.Textbox(
|
641 |
label="์ธ์๋ ์์ฑ",
|
@@ -657,9 +771,9 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
657 |
)
|
658 |
|
659 |
use_audio_toggle = gr.Checkbox(
|
660 |
-
label="๐ค ์์ฑ ์ธ์ ์ฌ์ฉ",
|
661 |
value=False,
|
662 |
-
info="
|
663 |
)
|
664 |
|
665 |
with gr.Row():
|
@@ -719,6 +833,19 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
719 |
label="๋ง์ดํฌ ์
๋ ฅ"
|
720 |
)
|
721 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
722 |
# ๋ฌธ์ ๋ถ์ ํญ (์จ๊น)
|
723 |
with gr.Tab("๐ ๋ฌธ์ ๋ถ์", visible=False):
|
724 |
with gr.Row():
|
@@ -752,8 +879,6 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
752 |
|
753 |
# ์ด๋ฒคํธ ํธ๋ค๋ฌ
|
754 |
webcam_state = gr.State(None)
|
755 |
-
audio_state = gr.State(None)
|
756 |
-
transcript_state = gr.State("")
|
757 |
|
758 |
def capture_webcam(frame):
|
759 |
"""์น์บ ํ๋ ์ ์บก์ฒ"""
|
@@ -763,15 +888,29 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
763 |
|
764 |
def clear_capture():
|
765 |
"""์บก์ฒ ์ด๊ธฐํ"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
766 |
return None, gr.update(visible=False), '<div class="status-box" style="background:#d4edda; color:#155724;">๐ฎ ์์คํ
์ค๋น</div>', ""
|
767 |
|
768 |
-
def analyze_with_task(image, prompt, task_type, use_search, thinking, tokens
|
769 |
"""ํน์ ํ์คํฌ๋ก ์ด๋ฏธ์ง ๋ถ์"""
|
|
|
|
|
770 |
if image is None:
|
771 |
return "โ ๋จผ์ ์ด๋ฏธ์ง๋ฅผ ์บก์ฒํ์ธ์.", '<div class="status-box" style="background:#f8d7da; color:#721c24;">โ ์ด๋ฏธ์ง ์์</div>'
|
772 |
|
773 |
status = f'<div class="status-box" style="background:#cce5ff; color:#004085;">๐ {task_type} ๋ถ์ ์ค...</div>'
|
774 |
|
|
|
|
|
|
|
|
|
|
|
775 |
result = analyze_image_for_robot(
|
776 |
image=image,
|
777 |
prompt=prompt,
|
@@ -800,27 +939,31 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
800 |
return formatted_result, complete_status
|
801 |
|
802 |
# ์๋ ์บก์ฒ ๋ฐ ๋ถ์ ํจ์
|
803 |
-
def auto_capture_and_analyze(webcam_frame,
|
804 |
-
"""์๋ ์บก์ฒ ๋ฐ ๋ถ์ (์ค๋์ค
|
|
|
|
|
805 |
if webcam_frame is None:
|
806 |
return (
|
807 |
None,
|
808 |
"์๋ ์บก์ฒ ๋๊ธฐ ์ค...",
|
809 |
'<div class="status-box" style="background:#fff3cd; color:#856404;">โณ ์น์บ ๋๊ธฐ ์ค</div>',
|
810 |
'<div class="auto-capture-status">๐ ์๋ ์บก์ฒ: ์น์บ ๋๊ธฐ ์ค</div>',
|
811 |
-
|
812 |
-
current_transcript
|
813 |
)
|
814 |
|
815 |
# ์บก์ฒ ์ํ
|
816 |
timestamp = time.strftime("%H:%M:%S")
|
817 |
|
818 |
-
#
|
819 |
-
|
820 |
-
|
821 |
-
|
822 |
-
|
823 |
-
|
|
|
|
|
|
|
824 |
|
825 |
# ์ด๋ฏธ์ง ๋ถ์ (์์
๊ณํ ๋ชจ๋๋ก)
|
826 |
result = analyze_image_for_robot(
|
@@ -830,7 +973,7 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
830 |
use_web_search=use_search,
|
831 |
enable_thinking=thinking,
|
832 |
max_new_tokens=tokens,
|
833 |
-
audio_transcript=
|
834 |
)
|
835 |
|
836 |
formatted_result = f"""๐ ์๋ ๋ถ์ ์๋ฃ ({timestamp})
|
@@ -838,13 +981,15 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
838 |
{result}
|
839 |
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ"""
|
840 |
|
|
|
|
|
|
|
841 |
return (
|
842 |
webcam_frame,
|
843 |
formatted_result,
|
844 |
'<div class="status-box" style="background:#d4edda; color:#155724;">โ
์๋ ๋ถ์ ์๋ฃ</div>',
|
845 |
f'<div class="auto-capture-status">๐ ์๋ ์บก์ฒ: ๋ง์ง๋ง ๋ถ์ {timestamp}</div>',
|
846 |
-
|
847 |
-
new_transcript if new_transcript else current_transcript
|
848 |
)
|
849 |
|
850 |
# ์น์บ ์คํธ๋ฆฌ๋ฐ
|
@@ -854,14 +999,17 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
854 |
outputs=[webcam_state]
|
855 |
)
|
856 |
|
857 |
-
# ์ค๋์ค ์คํธ๋ฆฌ๋ฐ
|
858 |
-
def
|
859 |
-
|
|
|
|
|
860 |
|
|
|
861 |
audio_input.stream(
|
862 |
-
fn=
|
863 |
inputs=[audio_input],
|
864 |
-
outputs=
|
865 |
)
|
866 |
|
867 |
# ์๋ ์บก์ฒ ๋ฒํผ
|
@@ -874,19 +1022,19 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
874 |
# ์ด๊ธฐํ ๋ฒํผ
|
875 |
clear_capture_btn.click(
|
876 |
fn=clear_capture,
|
877 |
-
outputs=[webcam_state, captured_image, status_display,
|
878 |
)
|
879 |
|
880 |
# ์์
๋ฒํผ๋ค
|
881 |
planning_btn.click(
|
882 |
-
fn=lambda img, p, s, t, tk
|
883 |
-
inputs=[captured_image, task_prompt, use_web_search, enable_thinking, max_tokens
|
884 |
outputs=[result_output, status_display]
|
885 |
)
|
886 |
|
887 |
grounding_btn.click(
|
888 |
-
fn=lambda img, p, s, t, tk
|
889 |
-
inputs=[captured_image, task_prompt, use_web_search, enable_thinking, max_tokens
|
890 |
outputs=[result_output, status_display]
|
891 |
)
|
892 |
|
@@ -924,14 +1072,29 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
924 |
|
925 |
# ์ค๋์ค ํ ๊ธ ์ด๋ฒคํธ
|
926 |
def toggle_audio(enabled):
|
|
|
|
|
927 |
if enabled:
|
928 |
# Whisper ๋ชจ๋ธ ๋ก๋
|
929 |
load_whisper()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
930 |
return (
|
931 |
gr.update(visible=True), # audio_input ํ์
|
932 |
-
'<div class="audio-status">๐ค ์์ฑ ์ธ์:
|
933 |
)
|
934 |
else:
|
|
|
|
|
|
|
|
|
|
|
|
|
935 |
return (
|
936 |
gr.update(visible=False), # audio_input ์จ๊น
|
937 |
'<div class="audio-status">๐ค ์์ฑ ์ธ์: ๋นํ์ฑํ</div>'
|
@@ -946,13 +1109,18 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
946 |
# ํ์ด๋จธ ํฑ ์ด๋ฒคํธ
|
947 |
timer.tick(
|
948 |
fn=auto_capture_and_analyze,
|
949 |
-
inputs=[webcam_state,
|
950 |
-
outputs=[captured_image, result_output, status_display, auto_capture_status,
|
951 |
)
|
952 |
|
953 |
# ์ด๊ธฐ ๋ชจ๋ธ ๋ก๋
|
954 |
def initial_load():
|
955 |
load_model()
|
|
|
|
|
|
|
|
|
|
|
956 |
return "์์คํ
์ค๋น ์๋ฃ! ๐"
|
957 |
|
958 |
demo.load(
|
@@ -961,8 +1129,8 @@ with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as dem
|
|
961 |
)
|
962 |
|
963 |
if __name__ == "__main__":
|
964 |
-
print("๐ ๋ก๋ด ์๊ฐ ์์คํ
์์ (Gemma3-R1984-4B + Whisper)...")
|
965 |
-
demo.launch(
|
966 |
server_name="0.0.0.0",
|
967 |
server_port=7860,
|
968 |
share=False,
|
|
|
15 |
import numpy as np
|
16 |
from loguru import logger
|
17 |
from PIL import Image
|
18 |
+
from transformers import AutoProcessor, Gemma3ForConditionalGeneration, TextIteratorStreamer, pipeline
|
19 |
import time
|
20 |
import warnings
|
21 |
from typing import Dict, List, Optional, Union
|
22 |
import librosa
|
23 |
+
import scipy.signal as sps
|
24 |
+
from threading import Thread, Lock
|
25 |
+
import queue
|
26 |
|
27 |
# CSV/TXT ๋ถ์
|
28 |
import pandas as pd
|
|
|
31 |
|
32 |
warnings.filterwarnings('ignore')
|
33 |
|
34 |
+
print("๐ฎ ๋ก๋ด ์๊ฐ ์์คํ
์ด๊ธฐํ (Gemma3-R1984-4B + Whisper + 10์ด ๊ต๋ ๋
น์)...")
|
35 |
|
36 |
##############################################################################
|
37 |
# ์์ ์ ์
|
|
|
47 |
model = None
|
48 |
processor = None
|
49 |
whisper_model = None
|
|
|
50 |
model_loaded = False
|
51 |
whisper_loaded = False
|
52 |
model_name = "Gemma3-R1984-4B"
|
|
|
65 |
##############################################################################
|
66 |
@spaces.GPU(duration=60)
|
67 |
def load_whisper():
|
68 |
+
global whisper_model, whisper_loaded
|
69 |
|
70 |
if whisper_loaded:
|
71 |
logger.info("Whisper ๋ชจ๋ธ์ด ์ด๋ฏธ ๋ก๋๋์ด ์์ต๋๋ค.")
|
|
|
73 |
|
74 |
try:
|
75 |
logger.info("Whisper ๋ชจ๋ธ ๋ก๋ฉ ์์...")
|
76 |
+
# ํ์ดํ๋ผ์ธ ๋ฐฉ์์ผ๋ก ๋ก๋
|
77 |
+
device = 0 if torch.cuda.is_available() else "cpu"
|
78 |
+
whisper_model = pipeline(
|
79 |
+
task="automatic-speech-recognition",
|
80 |
+
model="openai/whisper-base",
|
81 |
+
chunk_length_s=30,
|
82 |
+
device=device,
|
83 |
)
|
84 |
whisper_loaded = True
|
85 |
logger.info("โ
Whisper ๋ชจ๋ธ ๋ก๋ฉ ์๋ฃ!")
|
|
|
91 |
##############################################################################
|
92 |
# ์ค๋์ค ์ฒ๋ฆฌ ํจ์
|
93 |
##############################################################################
|
94 |
+
import scipy.signal as sps
|
95 |
+
from threading import Thread, Lock
|
96 |
+
import queue
|
97 |
+
|
98 |
+
# ์ค๋์ค ๋ฒํผ ๊ด๋ฆฌ
|
99 |
+
audio_buffer_lock = Lock()
|
100 |
+
audio_buffer_a = []
|
101 |
+
audio_buffer_b = []
|
102 |
+
current_buffer = 'a' # ํ์ฌ ๋
น์ ์ค์ธ ๋ฒํผ
|
103 |
+
processing_queue = queue.Queue() # ์ฒ๋ฆฌ ๋๊ธฐ ํ
|
104 |
+
last_transcription = "" # ๋ง์ง๋ง ์ ์ฌ ๊ฒฐ๊ณผ
|
105 |
+
|
106 |
+
def resample_audio(audio: np.ndarray, orig_sr: int, target_sr: int = 16000) -> np.ndarray:
|
107 |
+
"""์ค๋์ค ๋ฆฌ์ํ๋ง"""
|
108 |
+
if orig_sr == target_sr:
|
109 |
+
return audio.astype(np.float32)
|
110 |
+
|
111 |
+
# scipy๋ฅผ ์ฌ์ฉํ ๋ฆฌ์ํ๋ง
|
112 |
+
number_of_samples = round(len(audio) * float(target_sr) / orig_sr)
|
113 |
+
audio_resampled = sps.resample(audio, number_of_samples)
|
114 |
+
return audio_resampled.astype(np.float32)
|
115 |
+
|
116 |
@spaces.GPU(duration=30)
|
117 |
+
def transcribe_audio_whisper(audio_array: np.ndarray, sr: int = 16000):
|
118 |
"""Whisper๋ฅผ ์ฌ์ฉํ ์ค๋์ค ์ ์ฌ"""
|
119 |
+
global whisper_model, whisper_loaded
|
120 |
|
121 |
if not whisper_loaded:
|
122 |
if not load_whisper():
|
123 |
+
return None
|
124 |
|
125 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
# ์์ฑ ์ธ์
|
127 |
+
result = whisper_model({"array": audio_array, "sampling_rate": sr})
|
128 |
+
transcription = result["text"].strip()
|
129 |
|
130 |
+
return transcription if transcription else None
|
131 |
+
|
132 |
+
except Exception as e:
|
133 |
+
logger.error(f"Whisper ์ค๋์ค ์ ์ฌ ์ค๋ฅ: {e}")
|
134 |
+
return None
|
135 |
+
|
136 |
+
def accumulate_audio(audio_chunk):
|
137 |
+
"""์ค๋์ค ์ฒญํฌ๋ฅผ ๋ฒํผ์ ๋์ """
|
138 |
+
global current_buffer, audio_buffer_a, audio_buffer_b
|
139 |
+
|
140 |
+
if audio_chunk is None:
|
141 |
+
return
|
142 |
+
|
143 |
+
sr, audio = audio_chunk
|
144 |
+
|
145 |
+
# ์คํ
๋ ์ค๋ฅผ ๋ชจ๋
ธ๋ก ๋ณํ
|
146 |
+
if audio.ndim > 1:
|
147 |
+
audio = audio.mean(axis=1)
|
148 |
+
|
149 |
+
with audio_buffer_lock:
|
150 |
+
if current_buffer == 'a':
|
151 |
+
audio_buffer_a.append((audio, sr))
|
152 |
+
else:
|
153 |
+
audio_buffer_b.append((audio, sr))
|
154 |
+
|
155 |
+
def switch_buffers():
|
156 |
+
"""๋ฒํผ ์ ํ ๋ฐ ์ฒ๋ฆฌ ํ์ ์ถ๊ฐ"""
|
157 |
+
global current_buffer, audio_buffer_a, audio_buffer_b
|
158 |
+
|
159 |
+
with audio_buffer_lock:
|
160 |
+
if current_buffer == 'a':
|
161 |
+
# A ๋ฒํผ๋ฅผ ์ฒ๋ฆฌ ํ์ ์ถ๊ฐ
|
162 |
+
if audio_buffer_a:
|
163 |
+
processing_queue.put(('a', audio_buffer_a.copy()))
|
164 |
+
audio_buffer_a.clear()
|
165 |
+
current_buffer = 'b'
|
166 |
+
else:
|
167 |
+
# B ๋ฒํผ๋ฅผ ์ฒ๋ฆฌ ํ์ ์ถ๊ฐ
|
168 |
+
if audio_buffer_b:
|
169 |
+
processing_queue.put(('b', audio_buffer_b.copy()))
|
170 |
+
audio_buffer_b.clear()
|
171 |
+
current_buffer = 'a'
|
172 |
+
|
173 |
+
def process_audio_buffer(buffer_data):
|
174 |
+
"""๋ฒํผ์ ์ค๋์ค ๋ฐ์ดํฐ ์ฒ๋ฆฌ"""
|
175 |
+
buffer_name, audio_chunks = buffer_data
|
176 |
+
|
177 |
+
if not audio_chunks:
|
178 |
+
return None
|
179 |
+
|
180 |
+
try:
|
181 |
+
# ๋ชจ๋ ์ฒญํฌ๋ฅผ ํ๋๋ก ๊ฒฐํฉ
|
182 |
+
combined_audio = []
|
183 |
+
sample_rate = 16000
|
184 |
|
185 |
+
for audio, sr in audio_chunks:
|
186 |
+
# 16kHz๋ก ๋ฆฌ์ํ๋ง
|
187 |
+
if sr != 16000:
|
188 |
+
audio = resample_audio(audio, sr, 16000)
|
189 |
+
combined_audio.append(audio)
|
190 |
|
191 |
+
# ๊ฒฐํฉ
|
192 |
+
if combined_audio:
|
193 |
+
full_audio = np.concatenate(combined_audio)
|
194 |
+
|
195 |
+
# Whisper๋ก ์ ์ฌ
|
196 |
+
transcription = transcribe_audio_whisper(full_audio, 16000)
|
197 |
+
|
198 |
+
if transcription:
|
199 |
+
logger.info(f"๋ฒํผ {buffer_name} ์ ์ฌ ์๋ฃ: {transcription[:50]}...")
|
200 |
+
return transcription
|
201 |
+
|
202 |
except Exception as e:
|
203 |
+
logger.error(f"์ค๋์ค ๋ฒํผ ์ฒ๋ฆฌ ์ค๋ฅ: {e}")
|
204 |
+
|
205 |
+
return None
|
206 |
+
|
207 |
+
# ๋ฐฑ๊ทธ๋ผ์ด๋ ์ฒ๋ฆฌ ์ค๋ ๋
|
208 |
+
def audio_processing_worker():
|
209 |
+
"""๋ฐฑ๊ทธ๋ผ์ด๋์์ ์ค๋์ค ๋ฒํผ ์ฒ๋ฆฌ"""
|
210 |
+
global last_transcription
|
211 |
+
|
212 |
+
while True:
|
213 |
+
try:
|
214 |
+
# ์ฒ๋ฆฌํ ๋ฒํผ ๊ฐ์ ธ์ค๊ธฐ
|
215 |
+
buffer_data = processing_queue.get(timeout=1)
|
216 |
+
|
217 |
+
# ์ค๋์ค ์ฒ๋ฆฌ
|
218 |
+
result = process_audio_buffer(buffer_data)
|
219 |
+
|
220 |
+
if result:
|
221 |
+
# ๊ฒฐ๊ณผ๋ฅผ ์ ์ญ ๋ณ์์ ์ ์ฅ (๋์ค์ ์ฌ์ฉ)
|
222 |
+
with audio_buffer_lock:
|
223 |
+
last_transcription = result
|
224 |
+
|
225 |
+
except queue.Empty:
|
226 |
+
continue
|
227 |
+
except Exception as e:
|
228 |
+
logger.error(f"์ค๋์ค ์ฒ๋ฆฌ ์์ปค ์ค๋ฅ: {e}")
|
229 |
|
230 |
##############################################################################
|
231 |
# ํค์๋ ์ถ์ถ ํจ์
|
|
|
694 |
background: #e3f2fd;
|
695 |
color: #1565c0;
|
696 |
}
|
697 |
+
.buffer-info {
|
698 |
+
font-size: 0.9em;
|
699 |
+
color: #666;
|
700 |
+
text-align: center;
|
701 |
+
margin-top: 5px;
|
702 |
+
}
|
703 |
"""
|
704 |
|
705 |
with gr.Blocks(title="๐ค ๋ก๋ด ์๊ฐ ์์คํ
(Gemma3-4B)", css=css) as demo:
|
706 |
gr.HTML("""
|
707 |
<div class="robot-header">
|
708 |
<h1>๐ค ๋ก๋ด ์๊ฐ ์์คํ
</h1>
|
709 |
+
<h3>๐ฎ Gemma3-R1984-4B + ๐ท ์ค์๊ฐ ์น์บ + ๐ค 10์ด ๊ต๋ ์์ฑ ์ธ์</h3>
|
710 |
<p>โก ๋ฉํฐ๋ชจ๋ฌ AI๋ก ๋ก๋ด ์์
๋ถ์!</p>
|
711 |
</div>
|
712 |
""")
|
|
|
745 |
'<div class="audio-status">๐ค ์์ฑ ์ธ์: ๋นํ์ฑํ</div>'
|
746 |
)
|
747 |
|
748 |
+
# ๋ฒํผ ์ ๋ณด
|
749 |
+
gr.HTML(
|
750 |
+
'<div class="buffer-info">A/B ๋ฒํผ ๊ต๋ ๋
น์์ผ๋ก ๋๊น ์๋ ์ธ์</div>'
|
751 |
+
)
|
752 |
+
|
753 |
# ๋ง์ง๋ง ์ธ์๋ ํ
์คํธ
|
754 |
last_transcript = gr.Textbox(
|
755 |
label="์ธ์๋ ์์ฑ",
|
|
|
771 |
)
|
772 |
|
773 |
use_audio_toggle = gr.Checkbox(
|
774 |
+
label="๐ค ์์ฑ ์ธ์ ์ฌ์ฉ (10์ด ๊ต๋ ๋
น์)",
|
775 |
value=False,
|
776 |
+
info="10์ด๋ง๋ค ๊ต๋๋ก ๋
น์ํ์ฌ ๋๊น ์์ด ์ธ์"
|
777 |
)
|
778 |
|
779 |
with gr.Row():
|
|
|
833 |
label="๋ง์ดํฌ ์
๋ ฅ"
|
834 |
)
|
835 |
|
836 |
+
# ์ค๋์ค ์คํธ๋ฆฌ๋ฐ ์ฒ๋ฆฌ
|
837 |
+
def audio_stream_callback(audio_chunk):
|
838 |
+
"""์ค๋์ค ์คํธ๋ฆผ ์ฝ๋ฐฑ - ๋ฒํผ์ ๋์ """
|
839 |
+
accumulate_audio(audio_chunk)
|
840 |
+
return None # ์ํ ์
๋ฐ๏ฟฝ๏ฟฝ๏ฟฝํธ ์์
|
841 |
+
|
842 |
+
# ์ค๋์ค ์คํธ๋ฆฌ๋ฐ ์ฐ๊ฒฐ
|
843 |
+
audio_input.stream(
|
844 |
+
fn=audio_stream_callback,
|
845 |
+
inputs=[audio_input],
|
846 |
+
outputs=None
|
847 |
+
)
|
848 |
+
|
849 |
# ๋ฌธ์ ๋ถ์ ํญ (์จ๊น)
|
850 |
with gr.Tab("๐ ๋ฌธ์ ๋ถ์", visible=False):
|
851 |
with gr.Row():
|
|
|
879 |
|
880 |
# ์ด๋ฒคํธ ํธ๋ค๋ฌ
|
881 |
webcam_state = gr.State(None)
|
|
|
|
|
882 |
|
883 |
def capture_webcam(frame):
|
884 |
"""์น์บ ํ๋ ์ ์บก์ฒ"""
|
|
|
888 |
|
889 |
def clear_capture():
|
890 |
"""์บก์ฒ ์ด๊ธฐํ"""
|
891 |
+
global last_transcription, audio_buffer_a, audio_buffer_b
|
892 |
+
|
893 |
+
with audio_buffer_lock:
|
894 |
+
last_transcription = ""
|
895 |
+
audio_buffer_a.clear()
|
896 |
+
audio_buffer_b.clear()
|
897 |
+
|
898 |
return None, gr.update(visible=False), '<div class="status-box" style="background:#d4edda; color:#155724;">๐ฎ ์์คํ
์ค๋น</div>', ""
|
899 |
|
900 |
+
def analyze_with_task(image, prompt, task_type, use_search, thinking, tokens):
|
901 |
"""ํน์ ํ์คํฌ๋ก ์ด๋ฏธ์ง ๋ถ์"""
|
902 |
+
global last_transcription
|
903 |
+
|
904 |
if image is None:
|
905 |
return "โ ๋จผ์ ์ด๋ฏธ์ง๋ฅผ ์บก์ฒํ์ธ์.", '<div class="status-box" style="background:#f8d7da; color:#721c24;">โ ์ด๋ฏธ์ง ์์</div>'
|
906 |
|
907 |
status = f'<div class="status-box" style="background:#cce5ff; color:#004085;">๐ {task_type} ๋ถ์ ์ค...</div>'
|
908 |
|
909 |
+
# ํ์ฌ ์ ์ฌ ํ
์คํธ ๊ฐ์ ธ์ค๊ธฐ
|
910 |
+
transcript = ""
|
911 |
+
with audio_buffer_lock:
|
912 |
+
transcript = last_transcription
|
913 |
+
|
914 |
result = analyze_image_for_robot(
|
915 |
image=image,
|
916 |
prompt=prompt,
|
|
|
939 |
return formatted_result, complete_status
|
940 |
|
941 |
# ์๋ ์บก์ฒ ๋ฐ ๋ถ์ ํจ์
|
942 |
+
def auto_capture_and_analyze(webcam_frame, task_prompt, use_search, thinking, tokens, use_audio):
|
943 |
+
"""์๋ ์บก์ฒ ๋ฐ ๋ถ์ (10์ด๋ง๋ค ์ค๋์ค ๋ฒํผ ์ ํ)"""
|
944 |
+
global last_transcription
|
945 |
+
|
946 |
if webcam_frame is None:
|
947 |
return (
|
948 |
None,
|
949 |
"์๋ ์บก์ฒ ๋๊ธฐ ์ค...",
|
950 |
'<div class="status-box" style="background:#fff3cd; color:#856404;">โณ ์น์บ ๋๊ธฐ ์ค</div>',
|
951 |
'<div class="auto-capture-status">๐ ์๋ ์บก์ฒ: ์น์บ ๋๊ธฐ ์ค</div>',
|
952 |
+
""
|
|
|
953 |
)
|
954 |
|
955 |
# ์บก์ฒ ์ํ
|
956 |
timestamp = time.strftime("%H:%M:%S")
|
957 |
|
958 |
+
# ๋ฒํผ ์ ํ (10์ด๋ง๋ค)
|
959 |
+
if use_audio:
|
960 |
+
switch_buffers()
|
961 |
+
|
962 |
+
# ๋ง์ง๋ง ์ ์ฌ ๊ฒฐ๊ณผ ๊ฐ์ ธ์ค๊ธฐ
|
963 |
+
audio_transcript = ""
|
964 |
+
if use_audio:
|
965 |
+
with audio_buffer_lock:
|
966 |
+
audio_transcript = last_transcription
|
967 |
|
968 |
# ์ด๋ฏธ์ง ๋ถ์ (์์
๊ณํ ๋ชจ๋๋ก)
|
969 |
result = analyze_image_for_robot(
|
|
|
973 |
use_web_search=use_search,
|
974 |
enable_thinking=thinking,
|
975 |
max_new_tokens=tokens,
|
976 |
+
audio_transcript=audio_transcript if audio_transcript else None
|
977 |
)
|
978 |
|
979 |
formatted_result = f"""๐ ์๋ ๋ถ์ ์๋ฃ ({timestamp})
|
|
|
981 |
{result}
|
982 |
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ"""
|
983 |
|
984 |
+
# ๋ง์ง๋ง ์ธ์๋ ํ
์คํธ ์
๋ฐ์ดํธ
|
985 |
+
transcript_display = audio_transcript if audio_transcript else "์์ฑ ์ธ์ ๋๊ธฐ ์ค..."
|
986 |
+
|
987 |
return (
|
988 |
webcam_frame,
|
989 |
formatted_result,
|
990 |
'<div class="status-box" style="background:#d4edda; color:#155724;">โ
์๋ ๋ถ์ ์๋ฃ</div>',
|
991 |
f'<div class="auto-capture-status">๐ ์๋ ์บก์ฒ: ๋ง์ง๋ง ๋ถ์ {timestamp}</div>',
|
992 |
+
transcript_display
|
|
|
993 |
)
|
994 |
|
995 |
# ์น์บ ์คํธ๋ฆฌ๋ฐ
|
|
|
999 |
outputs=[webcam_state]
|
1000 |
)
|
1001 |
|
1002 |
+
# ์ค๋์ค ์คํธ๋ฆฌ๋ฐ ์ฒ๋ฆฌ
|
1003 |
+
def audio_stream_callback(audio_chunk):
|
1004 |
+
"""์ค๋์ค ์คํธ๋ฆผ ์ฝ๋ฐฑ - ๋ฒํผ์ ๋์ """
|
1005 |
+
accumulate_audio(audio_chunk)
|
1006 |
+
return None # ์ํ ์
๋ฐ์ดํธ ์์
|
1007 |
|
1008 |
+
# ์ค๋์ค ์คํธ๋ฆฌ๋ฐ ์ฐ๊ฒฐ
|
1009 |
audio_input.stream(
|
1010 |
+
fn=audio_stream_callback,
|
1011 |
inputs=[audio_input],
|
1012 |
+
outputs=None
|
1013 |
)
|
1014 |
|
1015 |
# ์๋ ์บก์ฒ ๋ฒํผ
|
|
|
1022 |
# ์ด๊ธฐํ ๋ฒํผ
|
1023 |
clear_capture_btn.click(
|
1024 |
fn=clear_capture,
|
1025 |
+
outputs=[webcam_state, captured_image, status_display, last_transcript]
|
1026 |
)
|
1027 |
|
1028 |
# ์์
๋ฒํผ๋ค
|
1029 |
planning_btn.click(
|
1030 |
+
fn=lambda img, p, s, t, tk: analyze_with_task(img, p, "planning", s, t, tk),
|
1031 |
+
inputs=[captured_image, task_prompt, use_web_search, enable_thinking, max_tokens],
|
1032 |
outputs=[result_output, status_display]
|
1033 |
)
|
1034 |
|
1035 |
grounding_btn.click(
|
1036 |
+
fn=lambda img, p, s, t, tk: analyze_with_task(img, p, "grounding", s, t, tk),
|
1037 |
+
inputs=[captured_image, task_prompt, use_web_search, enable_thinking, max_tokens],
|
1038 |
outputs=[result_output, status_display]
|
1039 |
)
|
1040 |
|
|
|
1072 |
|
1073 |
# ์ค๋์ค ํ ๊ธ ์ด๋ฒคํธ
|
1074 |
def toggle_audio(enabled):
|
1075 |
+
global audio_buffer_a, audio_buffer_b, current_buffer, last_transcription
|
1076 |
+
|
1077 |
if enabled:
|
1078 |
# Whisper ๋ชจ๋ธ ๋ก๋
|
1079 |
load_whisper()
|
1080 |
+
# ๋ฒํผ ์ด๊ธฐํ
|
1081 |
+
with audio_buffer_lock:
|
1082 |
+
audio_buffer_a.clear()
|
1083 |
+
audio_buffer_b.clear()
|
1084 |
+
current_buffer = 'a'
|
1085 |
+
last_transcription = ""
|
1086 |
+
|
1087 |
return (
|
1088 |
gr.update(visible=True), # audio_input ํ์
|
1089 |
+
'<div class="audio-status">๐ค ์์ฑ ์ธ์: ํ์ฑํ๋จ (10์ด ๊ต๋ ๋
น์)</div>'
|
1090 |
)
|
1091 |
else:
|
1092 |
+
# ๋ฒํผ ์ด๊ธฐํ
|
1093 |
+
with audio_buffer_lock:
|
1094 |
+
audio_buffer_a.clear()
|
1095 |
+
audio_buffer_b.clear()
|
1096 |
+
last_transcription = ""
|
1097 |
+
|
1098 |
return (
|
1099 |
gr.update(visible=False), # audio_input ์จ๊น
|
1100 |
'<div class="audio-status">๐ค ์์ฑ ์ธ์: ๋นํ์ฑํ</div>'
|
|
|
1109 |
# ํ์ด๋จธ ํฑ ์ด๋ฒคํธ
|
1110 |
timer.tick(
|
1111 |
fn=auto_capture_and_analyze,
|
1112 |
+
inputs=[webcam_state, task_prompt, use_web_search, enable_thinking, max_tokens, use_audio_toggle],
|
1113 |
+
outputs=[captured_image, result_output, status_display, auto_capture_status, last_transcript]
|
1114 |
)
|
1115 |
|
1116 |
# ์ด๊ธฐ ๋ชจ๋ธ ๋ก๋
|
1117 |
def initial_load():
|
1118 |
load_model()
|
1119 |
+
|
1120 |
+
# ์ค๋์ค ์์ปค ์ค๋ ๋ ์์
|
1121 |
+
audio_worker_thread = Thread(target=audio_processing_worker, daemon=True)
|
1122 |
+
audio_worker_thread.start()
|
1123 |
+
|
1124 |
return "์์คํ
์ค๋น ์๋ฃ! ๐"
|
1125 |
|
1126 |
demo.load(
|
|
|
1129 |
)
|
1130 |
|
1131 |
if __name__ == "__main__":
|
1132 |
+
print("๐ ๋ก๋ด ์๊ฐ ์์คํ
์์ (Gemma3-R1984-4B + Whisper 10์ด ๊ต๋ ๋
น์)...")
|
1133 |
+
demo.queue().launch(
|
1134 |
server_name="0.0.0.0",
|
1135 |
server_port=7860,
|
1136 |
share=False,
|