Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -2,10 +2,8 @@
|
|
2 |
import gradio as gr
|
3 |
import numpy as np
|
4 |
import torch
|
5 |
-
import soundfile as sf
|
6 |
import os
|
7 |
import yaml
|
8 |
-
import spaces
|
9 |
from dotenv import load_dotenv
|
10 |
from threading import Thread
|
11 |
from gradio_webrtc import WebRTC
|
@@ -13,168 +11,145 @@ from twilio.rest import Client
|
|
13 |
|
14 |
# --- TTS & AI Imports ---
|
15 |
from parler_tts import ParlerTTSForConditionalGeneration
|
16 |
-
from transformers import AutoTokenizer, AutoFeatureExtractor
|
17 |
-
from streamer import ParlerTTSStreamer
|
18 |
|
|
|
19 |
from src.detection.factory import get_detector
|
20 |
from src.alerting.alert_system import get_alerter
|
21 |
|
22 |
# --- Load Configuration and Environment Variables ---
|
23 |
-
# This part is the same as our Streamlit app
|
24 |
load_dotenv()
|
25 |
config_path = 'config.yaml'
|
26 |
with open(config_path, 'r') as f:
|
27 |
config = yaml.safe_load(f)
|
28 |
-
secrets = {
|
29 |
-
"gemini_api_key": os.getenv("GEMINI_API_KEY"),
|
30 |
-
}
|
31 |
|
32 |
# --- Initialize Backend Components ---
|
33 |
print("Initializing detector and alerter...")
|
34 |
detector = get_detector(config)
|
35 |
alerter = get_alerter(config, secrets["gemini_api_key"])
|
36 |
-
print("Initialization complete.
|
37 |
|
|
|
38 |
account_sid = os.environ.get("TURN_USERNAME")
|
39 |
auth_token = os.environ.get("TURN_CREDENTIAL")
|
40 |
-
|
41 |
if account_sid and auth_token:
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
"
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
|
|
56 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
57 |
if device == "cpu":
|
58 |
-
print("\nWARNING: Running Parler-TTS on a CPU
|
59 |
torch_dtype = torch.float16 if device != "cpu" else torch.float32
|
60 |
|
61 |
-
|
62 |
-
# Using a smaller, faster model suitable for real-time alerts
|
63 |
repo_id = "parler-tts/parler_tts_mini_v0.1"
|
64 |
model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id, torch_dtype=torch_dtype).to(device)
|
65 |
tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
66 |
feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id)
|
67 |
print("Parler-TTS model loaded.")
|
68 |
|
69 |
-
# --- Audio Streaming Generator
|
70 |
-
@spaces.GPU
|
71 |
def stream_alert_audio(text_prompt):
|
72 |
-
"""
|
73 |
-
A generator function that yields audio chunks for a given text prompt.
|
74 |
-
This is the core of the streaming implementation.
|
75 |
-
"""
|
76 |
sampling_rate = model.config.sampling_rate
|
77 |
-
description = "
|
78 |
-
|
79 |
prompt_ids = tokenizer(text_prompt, return_tensors="pt").input_ids.to(device)
|
80 |
description_ids = tokenizer(description, return_tensors="pt").input_ids.to(device)
|
81 |
-
|
82 |
-
# Setup the streamer
|
83 |
streamer = ParlerTTSStreamer(model, device, play_steps=int(sampling_rate * 2.0))
|
84 |
-
|
85 |
-
generation_kwargs = dict(
|
86 |
-
input_ids=description_ids,
|
87 |
-
prompt_input_ids=prompt_ids,
|
88 |
-
streamer=streamer,
|
89 |
-
do_sample=True,
|
90 |
-
temperature=1.0, # Increase for more vocal variety
|
91 |
-
repetition_penalty=1.2,
|
92 |
-
)
|
93 |
-
|
94 |
-
# Run generation in a separate thread to not block the UI
|
95 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
96 |
-
|
97 |
try:
|
98 |
thread.start()
|
99 |
print(f"Audio stream started for: '{text_prompt}'")
|
100 |
-
# Yield audio chunks as they become available
|
101 |
for new_audio_chunk in streamer:
|
102 |
yield (sampling_rate, new_audio_chunk)
|
103 |
finally:
|
104 |
-
# CRITICAL: This block runs after the generator is exhausted (audio finishes)
|
105 |
-
# We reset the alerter state so that a new alert can be triggered later.
|
106 |
print("Audio stream finished. Resetting alerter state.")
|
107 |
alerter.reset_alert()
|
108 |
-
|
109 |
-
# ---
|
110 |
-
|
111 |
-
def
|
112 |
"""
|
113 |
-
|
114 |
-
|
115 |
"""
|
116 |
-
if
|
117 |
-
return np.zeros((480, 640, 3), dtype=np.uint8),
|
118 |
|
|
|
119 |
processed_frame, indicators, _ = detector.process_frame(frame)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
drowsiness_level = indicators.get("drowsiness_level", "Awake")
|
121 |
lighting = indicators.get("lighting", "Good")
|
122 |
score = indicators.get("details", {}).get("Score", 0)
|
123 |
|
124 |
-
|
125 |
-
status_text = f"Lighting: {lighting}\n"
|
126 |
if lighting == "Low":
|
127 |
-
status_text
|
128 |
-
else:
|
129 |
-
status_text += f"Status: {drowsiness_level}\nScore: {score:.2f}"
|
130 |
|
131 |
-
# --- Alert Trigger Logic ---
|
132 |
audio_output = None
|
133 |
if drowsiness_level != "Awake":
|
134 |
-
# alerter.trigger_alert() returns the alert TEXT if not on cooldown, otherwise None.
|
135 |
alert_text = alerter.trigger_alert(level=drowsiness_level)
|
136 |
if alert_text:
|
137 |
-
# If we got text, it means we can start an alert.
|
138 |
-
# We return the generator function itself. Gradio will handle it.
|
139 |
-
audio_output = stream_alert_audio(alert_text)
|
140 |
-
else:
|
141 |
-
alert_text = "WAKE UP"
|
142 |
audio_output = stream_alert_audio(alert_text)
|
143 |
-
|
144 |
-
# On subsequent frames where the user is drowsy, trigger_alert() will return None
|
145 |
-
# due to the cooldown, preventing a new stream from starting, which is what we want.
|
146 |
-
|
147 |
-
return processed_frame, status_text, audio_output
|
148 |
-
|
149 |
|
150 |
# --- Gradio UI Definition ---
|
151 |
with gr.Blocks(theme=gr.themes.Default(primary_hue="blue")) as app:
|
152 |
-
gr.Markdown("# 🚗 Drive Paddy - Drowsiness Detection (
|
153 |
-
gr.Markdown("
|
|
|
|
|
|
|
154 |
|
155 |
with gr.Row():
|
156 |
with gr.Column(scale=2):
|
157 |
-
|
|
|
158 |
with gr.Column(scale=1):
|
159 |
-
processed_output = gr.Image(label="Processed Feed")
|
160 |
status_output = gr.Textbox(label="Live Status", lines=3, interactive=False)
|
|
|
161 |
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
)
|
169 |
-
|
170 |
-
webcam_input.stream(
|
171 |
-
fn=process_live_frame,
|
172 |
-
inputs=[webcam_input],
|
173 |
-
outputs=[status_output, audio_alert_output],
|
174 |
-
time_limit=10
|
175 |
)
|
176 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
|
178 |
-
# --- Launch the App ---
|
179 |
if __name__ == "__main__":
|
180 |
-
|
|
|
|
2 |
import gradio as gr
|
3 |
import numpy as np
|
4 |
import torch
|
|
|
5 |
import os
|
6 |
import yaml
|
|
|
7 |
from dotenv import load_dotenv
|
8 |
from threading import Thread
|
9 |
from gradio_webrtc import WebRTC
|
|
|
11 |
|
12 |
# --- TTS & AI Imports ---
|
13 |
from parler_tts import ParlerTTSForConditionalGeneration
|
14 |
+
from transformers import AutoTokenizer, AutoFeatureExtractor
|
15 |
+
from streamer import ParlerTTSStreamer
|
16 |
|
17 |
+
# --- Local Project Imports ---
|
18 |
from src.detection.factory import get_detector
|
19 |
from src.alerting.alert_system import get_alerter
|
20 |
|
21 |
# --- Load Configuration and Environment Variables ---
|
|
|
22 |
load_dotenv()
|
23 |
config_path = 'config.yaml'
|
24 |
with open(config_path, 'r') as f:
|
25 |
config = yaml.safe_load(f)
|
26 |
+
secrets = {"gemini_api_key": os.getenv("GEMINI_API_KEY")}
|
|
|
|
|
27 |
|
28 |
# --- Initialize Backend Components ---
|
29 |
print("Initializing detector and alerter...")
|
30 |
detector = get_detector(config)
|
31 |
alerter = get_alerter(config, secrets["gemini_api_key"])
|
32 |
+
print("Initialization complete.")
|
33 |
|
34 |
+
# --- Twilio TURN Server Setup ---
|
35 |
account_sid = os.environ.get("TURN_USERNAME")
|
36 |
auth_token = os.environ.get("TURN_CREDENTIAL")
|
37 |
+
rtc_configuration = None
|
38 |
if account_sid and auth_token:
|
39 |
+
try:
|
40 |
+
client = Client(account_sid, auth_token)
|
41 |
+
token = client.tokens.create()
|
42 |
+
rtc_configuration = {"iceServers": token.ice_servers}
|
43 |
+
print("Twilio TURN server configured successfully.")
|
44 |
+
except Exception as e:
|
45 |
+
print(f"Warning: Failed to create Twilio token. Using public STUN server. Error: {e}")
|
46 |
+
# Fallback to a public STUN server if Twilio fails or is not configured
|
47 |
+
if rtc_configuration is None:
|
48 |
+
print("Using public STun server.")
|
49 |
+
rtc_configuration = {"iceServers": [{"urls": ["stun:stun.l.google.com:19302"]}]}
|
50 |
+
|
51 |
+
|
52 |
+
# --- Parler-TTS Model Setup ---
|
53 |
+
print("Loading Parler-TTS model...")
|
54 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
55 |
if device == "cpu":
|
56 |
+
print("\nWARNING: Running Parler-TTS on a CPU is slow. A GPU is highly recommended.\n")
|
57 |
torch_dtype = torch.float16 if device != "cpu" else torch.float32
|
58 |
|
|
|
|
|
59 |
repo_id = "parler-tts/parler_tts_mini_v0.1"
|
60 |
model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id, torch_dtype=torch_dtype).to(device)
|
61 |
tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
62 |
feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id)
|
63 |
print("Parler-TTS model loaded.")
|
64 |
|
65 |
+
# --- Audio Streaming Generator ---
|
|
|
66 |
def stream_alert_audio(text_prompt):
|
67 |
+
"""A generator that streams audio chunks for a given text prompt."""
|
|
|
|
|
|
|
68 |
sampling_rate = model.config.sampling_rate
|
69 |
+
description = "A female speaker with a clear and urgent voice."
|
|
|
70 |
prompt_ids = tokenizer(text_prompt, return_tensors="pt").input_ids.to(device)
|
71 |
description_ids = tokenizer(description, return_tensors="pt").input_ids.to(device)
|
|
|
|
|
72 |
streamer = ParlerTTSStreamer(model, device, play_steps=int(sampling_rate * 2.0))
|
73 |
+
generation_kwargs = dict(input_ids=description_ids, prompt_input_ids=prompt_ids, streamer=streamer, do_sample=True, temperature=1.0, repetition_penalty=1.2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
|
|
75 |
try:
|
76 |
thread.start()
|
77 |
print(f"Audio stream started for: '{text_prompt}'")
|
|
|
78 |
for new_audio_chunk in streamer:
|
79 |
yield (sampling_rate, new_audio_chunk)
|
80 |
finally:
|
|
|
|
|
81 |
print("Audio stream finished. Resetting alerter state.")
|
82 |
alerter.reset_alert()
|
83 |
+
|
84 |
+
# --- Decoupled Processing Functions ---
|
85 |
+
|
86 |
+
def process_video_and_update_state(frame_dict: dict, state: dict):
|
87 |
"""
|
88 |
+
HIGH-FREQUENCY LOOP: Processes video, updates shared state, and returns the processed frame.
|
89 |
+
This function's speed directly impacts video latency.
|
90 |
"""
|
91 |
+
if not frame_dict or "video" not in frame_dict or frame_dict["video"] is None:
|
92 |
+
return np.zeros((480, 640, 3), dtype=np.uint8), state
|
93 |
|
94 |
+
frame = frame_dict["video"]
|
95 |
processed_frame, indicators, _ = detector.process_frame(frame)
|
96 |
+
state['indicators'] = indicators
|
97 |
+
return processed_frame, state
|
98 |
+
|
99 |
+
def update_ui_from_state(state: dict):
|
100 |
+
"""
|
101 |
+
LOW-FREQUENCY LOOP: Reads from state to update status text and trigger audio.
|
102 |
+
This runs independently of the video loop.
|
103 |
+
"""
|
104 |
+
indicators = state.get('indicators', {})
|
105 |
drowsiness_level = indicators.get("drowsiness_level", "Awake")
|
106 |
lighting = indicators.get("lighting", "Good")
|
107 |
score = indicators.get("details", {}).get("Score", 0)
|
108 |
|
109 |
+
status_text = f"Lighting: {lighting}\nStatus: {drowsiness_level}\nScore: {score:.2f}"
|
|
|
110 |
if lighting == "Low":
|
111 |
+
status_text = "Lighting: Low\nDetection paused due to low light."
|
|
|
|
|
112 |
|
|
|
113 |
audio_output = None
|
114 |
if drowsiness_level != "Awake":
|
|
|
115 |
alert_text = alerter.trigger_alert(level=drowsiness_level)
|
116 |
if alert_text:
|
|
|
|
|
|
|
|
|
|
|
117 |
audio_output = stream_alert_audio(alert_text)
|
118 |
+
return status_text, audio_output
|
|
|
|
|
|
|
|
|
|
|
119 |
|
120 |
# --- Gradio UI Definition ---
|
121 |
with gr.Blocks(theme=gr.themes.Default(primary_hue="blue")) as app:
|
122 |
+
gr.Markdown("# 🚗 Drive Paddy - Drowsiness Detection (WebRTC)")
|
123 |
+
gr.Markdown("Low-latency video processing via WebRTC, with decoupled UI updates for smooth performance.")
|
124 |
+
|
125 |
+
# Shared state object to pass data between the two processing loops
|
126 |
+
shared_state = gr.State(value={'indicators': {}})
|
127 |
|
128 |
with gr.Row():
|
129 |
with gr.Column(scale=2):
|
130 |
+
# This WebRTC component is now correctly used for both input and output of the video stream.
|
131 |
+
webcam = WebRTC(label="Live Camera Feed", rtc_configuration=rtc_configuration)
|
132 |
with gr.Column(scale=1):
|
|
|
133 |
status_output = gr.Textbox(label="Live Status", lines=3, interactive=False)
|
134 |
+
audio_alert_output = gr.Audio(label="Alert System", autoplay=True, visible=False, streaming=True)
|
135 |
|
136 |
+
# LOOP 1: High-Frequency Video Stream (as fast as possible)
|
137 |
+
# This takes video from the webcam, processes it, and sends it right back.
|
138 |
+
webcam.stream(
|
139 |
+
fn=process_video_and_update_state,
|
140 |
+
inputs=[webcam, shared_state],
|
141 |
+
outputs=[webcam, shared_state],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
)
|
143 |
|
144 |
+
# LOOP 2: Low-Frequency UI Updates (4 times per second)
|
145 |
+
# This runs on a timer, reads the shared state, and updates the other UI elements.
|
146 |
+
app.load(
|
147 |
+
fn=update_ui_from_state,
|
148 |
+
inputs=[shared_state],
|
149 |
+
outputs=[status_output, audio_alert_output],
|
150 |
+
every=0.25 # Run this function every 250 milliseconds
|
151 |
+
)
|
152 |
|
|
|
153 |
if __name__ == "__main__":
|
154 |
+
print("Starting Drive Paddy WebRTC Application...")
|
155 |
+
app.launch(debug=True)
|