Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,97 +1,60 @@
|
|
| 1 |
import os
|
| 2 |
-
import json
|
| 3 |
-
import re
|
| 4 |
-
from collections.abc import Iterator
|
| 5 |
-
from threading import Thread
|
| 6 |
-
|
| 7 |
import gradio as gr
|
| 8 |
-
import spaces
|
| 9 |
import torch
|
| 10 |
import librosa
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
pipeline,
|
| 15 |
-
AutoModelForMultimodalLM,
|
| 16 |
-
BatchFeature,
|
| 17 |
-
StoppingCriteria
|
| 18 |
-
)
|
| 19 |
-
from transformers.generation.streamers import TextIteratorStreamer
|
| 20 |
|
| 21 |
-
#
|
| 22 |
-
|
| 23 |
-
# ==========================================
|
| 24 |
-
whisper_model_id = "openai/whisper-large-v3"
|
| 25 |
|
|
|
|
| 26 |
if torch.cuda.is_available() and torch.cuda.is_bf16_supported():
|
| 27 |
-
|
| 28 |
-
print("
|
| 29 |
else:
|
| 30 |
-
|
| 31 |
-
print("
|
| 32 |
|
|
|
|
| 33 |
attn_implementation = "sdpa"
|
| 34 |
try:
|
| 35 |
import flash_attn
|
| 36 |
attn_implementation = "flash_attention_2"
|
| 37 |
-
print("
|
| 38 |
except ImportError:
|
| 39 |
-
print("
|
| 40 |
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
|
|
|
| 44 |
low_cpu_mem_usage=True,
|
| 45 |
use_safetensors=True,
|
| 46 |
attn_implementation=attn_implementation
|
| 47 |
)
|
| 48 |
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
|
| 53 |
-
|
| 54 |
-
"automatic-speech-recognition",
|
| 55 |
-
model=whisper_model,
|
| 56 |
-
tokenizer=whisper_processor.tokenizer,
|
| 57 |
-
feature_extractor=whisper_processor.feature_extractor,
|
| 58 |
-
dtype=whisper_dtype,
|
| 59 |
-
device=whisper_device,
|
| 60 |
-
)
|
| 61 |
-
|
| 62 |
-
# ==========================================
|
| 63 |
-
# 2. تنظیمات و بارگذاری مدل زبانی چندرسانهای (Gemma)
|
| 64 |
-
# ==========================================
|
| 65 |
-
gemma_model_id = "google/gemma-4-e4b-it"
|
| 66 |
-
|
| 67 |
-
gemma_processor = AutoProcessor.from_pretrained(gemma_model_id, use_fast=False)
|
| 68 |
-
gemma_model = AutoModelForMultimodalLM.from_pretrained(
|
| 69 |
-
gemma_model_id,
|
| 70 |
-
device_map="auto",
|
| 71 |
-
dtype=torch.bfloat16
|
| 72 |
-
)
|
| 73 |
-
|
| 74 |
-
IMAGE_FILE_TYPES = (".jpg", ".jpeg", ".png", ".webp")
|
| 75 |
-
AUDIO_FILE_TYPES = (".wav", ".mp3", ".flac", ".ogg")
|
| 76 |
-
VIDEO_FILE_TYPES = (".mp4", ".mov", ".avi", ".webm")
|
| 77 |
-
MAX_INPUT_TOKENS = int(os.getenv("MAX_INPUT_TOKENS", "10_000"))
|
| 78 |
|
| 79 |
-
|
| 80 |
-
|
| 81 |
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
|
|
|
|
|
|
|
|
|
| 87 |
)
|
| 88 |
|
| 89 |
-
|
| 90 |
-
# ==========================================
|
| 91 |
-
# توابع کمکی تبدیل صدا به متن (Whisper)
|
| 92 |
-
# ==========================================
|
| 93 |
-
prompt_text = "کلمات دقیقاً همانطور که تلفظ میشوند، با رعایت حروف اضافه و ساختار عامیانه، محاورهای و شکسته نوشته شوند."
|
| 94 |
-
|
| 95 |
LANGUAGE_MAPPING = {
|
| 96 |
"Auto-Detect (تشخیص خودکار)": None,
|
| 97 |
"Persian / فارسی": "fa",
|
|
@@ -108,6 +71,7 @@ LANGUAGE_MAPPING = {
|
|
| 108 |
"Korean / کرهای": "ko"
|
| 109 |
}
|
| 110 |
|
|
|
|
| 111 |
def format_timestamp(seconds):
|
| 112 |
hours = int(seconds // 3600)
|
| 113 |
minutes = int((seconds % 3600) // 60)
|
|
@@ -115,49 +79,7 @@ def format_timestamp(seconds):
|
|
| 115 |
millis = int(round((seconds % 1) * 1000))
|
| 116 |
return f"{hours:02d}:{minutes:02d}:{secs:02d},{millis:03d}"
|
| 117 |
|
| 118 |
-
|
| 119 |
-
"""تولید استاندارد فایل زیرنویس SRT بر پایه دادههای جیسون بازنویسی شده کلمات"""
|
| 120 |
-
srt_lines = []
|
| 121 |
-
segment_idx = 1
|
| 122 |
-
current_segment_words = []
|
| 123 |
-
current_start = None
|
| 124 |
-
|
| 125 |
-
for item in word_items:
|
| 126 |
-
text = item.get("word", "").strip()
|
| 127 |
-
start = item.get("start")
|
| 128 |
-
end = item.get("end")
|
| 129 |
-
|
| 130 |
-
if not text or start is None:
|
| 131 |
-
continue
|
| 132 |
-
if end is None:
|
| 133 |
-
end = start + 0.3
|
| 134 |
-
|
| 135 |
-
if current_start is None:
|
| 136 |
-
current_start = start
|
| 137 |
-
|
| 138 |
-
current_segment_words.append(text)
|
| 139 |
-
duration = end - current_start
|
| 140 |
-
|
| 141 |
-
if len(current_segment_words) >= max_words_per_segment or duration >= max_duration_per_segment:
|
| 142 |
-
line_text = " ".join(current_segment_words)
|
| 143 |
-
srt_lines.append(f"{segment_idx}")
|
| 144 |
-
srt_lines.append(f"{format_timestamp(current_start)} --> {format_timestamp(end)}")
|
| 145 |
-
srt_lines.append(line_text)
|
| 146 |
-
srt_lines.append("")
|
| 147 |
-
segment_idx += 1
|
| 148 |
-
current_segment_words = []
|
| 149 |
-
current_start = None
|
| 150 |
-
|
| 151 |
-
if current_segment_words:
|
| 152 |
-
last_end = word_items[-1].get("end") if word_items and word_items[-1].get("end") else current_start + 1.0
|
| 153 |
-
line_text = " ".join(current_segment_words)
|
| 154 |
-
srt_lines.append(f"{segment_idx}")
|
| 155 |
-
srt_lines.append(f"{format_timestamp(current_start)} --> {format_timestamp(last_end)}")
|
| 156 |
-
srt_lines.append(line_text)
|
| 157 |
-
srt_lines.append("")
|
| 158 |
-
|
| 159 |
-
return "\n".join(srt_lines)
|
| 160 |
-
|
| 161 |
def create_srt_from_chunks(chunks, max_words_per_segment=6, max_duration_per_segment=3.0):
|
| 162 |
srt_lines = []
|
| 163 |
segment_idx = 1
|
|
@@ -175,7 +97,7 @@ def create_srt_from_chunks(chunks, max_words_per_segment=6, max_duration_per_seg
|
|
| 175 |
if start is None:
|
| 176 |
continue
|
| 177 |
if end is None:
|
| 178 |
-
end = start + 0.3
|
| 179 |
|
| 180 |
if current_start is None:
|
| 181 |
current_start = start
|
|
@@ -183,6 +105,7 @@ def create_srt_from_chunks(chunks, max_words_per_segment=6, max_duration_per_seg
|
|
| 183 |
current_segment_words.append(text)
|
| 184 |
duration = end - current_start
|
| 185 |
|
|
|
|
| 186 |
if len(current_segment_words) >= max_words_per_segment or duration >= max_duration_per_segment:
|
| 187 |
line_text = " ".join(current_segment_words)
|
| 188 |
srt_lines.append(f"{segment_idx}")
|
|
@@ -203,6 +126,7 @@ def create_srt_from_chunks(chunks, max_words_per_segment=6, max_duration_per_seg
|
|
| 203 |
|
| 204 |
return "\n".join(srt_lines)
|
| 205 |
|
|
|
|
| 206 |
def get_raw_word_timestamps(chunks):
|
| 207 |
word_timestamps = []
|
| 208 |
for chunk in chunks:
|
|
@@ -216,392 +140,123 @@ def get_raw_word_timestamps(chunks):
|
|
| 216 |
})
|
| 217 |
return json.dumps(word_timestamps, ensure_ascii=False, indent=2)
|
| 218 |
|
| 219 |
-
@spaces.GPU
|
| 220 |
def transcribe_audio(audio_path, language_selection, optimize_persian):
|
| 221 |
if audio_path is None:
|
| 222 |
return "Please upload or record an audio file first. / لطفاً ابتدا فایل صوتی را آپلود یا ضبط کنید.", "", ""
|
| 223 |
|
| 224 |
try:
|
| 225 |
-
#
|
| 226 |
audio_data, sampling_rate = librosa.load(audio_path, sr=16000)
|
|
|
|
|
|
|
| 227 |
lang_code = LANGUAGE_MAPPING.get(language_selection, None)
|
| 228 |
|
|
|
|
| 229 |
generate_kwargs = {
|
| 230 |
"task": "transcribe",
|
| 231 |
"do_sample": False,
|
| 232 |
-
"num_beams": 5
|
| 233 |
}
|
| 234 |
|
|
|
|
| 235 |
if lang_code is not None:
|
| 236 |
generate_kwargs["language"] = lang_code
|
| 237 |
|
|
|
|
| 238 |
if optimize_persian:
|
| 239 |
-
prompt_ids =
|
| 240 |
-
prompt_ids = prompt_ids.to(
|
| 241 |
generate_kwargs["prompt_ids"] = prompt_ids
|
| 242 |
-
generate_kwargs["language"] = "fa"
|
| 243 |
|
| 244 |
-
|
|
|
|
| 245 |
{"raw": audio_data, "sampling_rate": sampling_rate},
|
| 246 |
chunk_length_s=30,
|
| 247 |
stride_length_s=(6, 6),
|
| 248 |
batch_size=24,
|
| 249 |
-
return_timestamps="word",
|
| 250 |
generate_kwargs=generate_kwargs
|
| 251 |
)
|
| 252 |
|
| 253 |
full_text = result["text"].strip()
|
| 254 |
chunks = result.get("chunks", [])
|
| 255 |
|
|
|
|
| 256 |
srt_content = create_srt_from_chunks(chunks)
|
| 257 |
word_timestamps_json = get_raw_word_timestamps(chunks)
|
| 258 |
|
| 259 |
-
#
|
| 260 |
-
if word_timestamps_json:
|
| 261 |
-
correction_prompt = (
|
| 262 |
-
"You are an expert subtitle corrector. Below is a JSON array representing words with their exact start and end timestamps in Persian. "
|
| 263 |
-
"Your task is ONLY to fix spelling, grammatical, or listening errors inside the 'word' field of each item. "
|
| 264 |
-
"CRITICAL RULES:\n"
|
| 265 |
-
"1. Keep the JSON structure exactly the same. Do NOT translate or summarize.\n"
|
| 266 |
-
"2. The start and end timestamps must NOT be modified under any circumstance.\n"
|
| 267 |
-
"3. Do NOT add or remove items. The list length must remain exactly identical.\n"
|
| 268 |
-
"4. Output ONLY the corrected valid JSON array starting with '[' and ending with ']'. "
|
| 269 |
-
"Do NOT include any markdown code blocks, conversational text, explanations, or introductory sentences. Output raw JSON only.\n\n"
|
| 270 |
-
f"JSON:\n{word_timestamps_json}"
|
| 271 |
-
)
|
| 272 |
-
|
| 273 |
-
messages = [{"role": "user", "content": [{"type": "text", "text": correction_prompt}]}]
|
| 274 |
-
|
| 275 |
-
inputs = gemma_processor.apply_chat_template(
|
| 276 |
-
messages,
|
| 277 |
-
tokenize=True,
|
| 278 |
-
add_generation_prompt=True,
|
| 279 |
-
return_dict=True,
|
| 280 |
-
return_tensors="pt"
|
| 281 |
-
)
|
| 282 |
-
inputs = inputs.to(device=gemma_model.device, dtype=torch.bfloat16)
|
| 283 |
-
|
| 284 |
-
with torch.inference_mode():
|
| 285 |
-
outputs = gemma_model.generate(
|
| 286 |
-
**inputs,
|
| 287 |
-
max_new_tokens=4000,
|
| 288 |
-
disable_compile=True
|
| 289 |
-
)
|
| 290 |
-
|
| 291 |
-
input_length = inputs["input_ids"].shape[1]
|
| 292 |
-
generated_tokens = outputs[0][input_length:]
|
| 293 |
-
corrected_text = gemma_processor.decode(generated_tokens, skip_special_tokens=True).strip()
|
| 294 |
-
|
| 295 |
-
# پاکسازی و استخراج ایمن بلاک JSON
|
| 296 |
-
start_idx = corrected_text.find('[')
|
| 297 |
-
end_idx = corrected_text.rfind(']')
|
| 298 |
-
|
| 299 |
-
if start_idx != -1 and end_idx != -1:
|
| 300 |
-
clean_json_str = corrected_text[start_idx:end_idx+1]
|
| 301 |
-
try:
|
| 302 |
-
corrected_data = json.loads(clean_json_str)
|
| 303 |
-
|
| 304 |
-
# بازسازی تمام خروجیها از روی دیتای تصحیحشده
|
| 305 |
-
full_text = " ".join([item["word"] for item in corrected_data])
|
| 306 |
-
srt_content = create_srt_from_json_data(corrected_data)
|
| 307 |
-
word_timestamps_json = json.dumps(corrected_data, ensure_ascii=False, indent=2)
|
| 308 |
-
except Exception as je:
|
| 309 |
-
print(f"Failed to parse corrected JSON from Gemma: {je}")
|
| 310 |
-
# در صورت هرگونه خطا، از نتایج خام Whisper استفاده میشود تا برنامه متوقف نشود
|
| 311 |
-
|
| 312 |
return full_text, srt_content, word_timestamps_json
|
| 313 |
|
| 314 |
except Exception as e:
|
| 315 |
return f"An error occurred during processing / خطایی در حین پردازش رخ داد: {str(e)}", "", ""
|
| 316 |
|
| 317 |
-
|
| 318 |
-
|
| 319 |
-
# توابع کمکی چتبات (Gemma)
|
| 320 |
-
# ==========================================
|
| 321 |
-
def _strip_special_tokens(text: str) -> str:
|
| 322 |
-
for tok in _STRIP_TOKENS:
|
| 323 |
-
text = text.replace(tok, "")
|
| 324 |
-
return text
|
| 325 |
-
|
| 326 |
-
def _classify_file(path: str) -> str | None:
|
| 327 |
-
lower = path.lower()
|
| 328 |
-
if lower.endswith(IMAGE_FILE_TYPES):
|
| 329 |
-
return "image"
|
| 330 |
-
if lower.endswith(AUDIO_FILE_TYPES):
|
| 331 |
-
return "audio"
|
| 332 |
-
if lower.endswith(VIDEO_FILE_TYPES):
|
| 333 |
-
return "video"
|
| 334 |
-
return None
|
| 335 |
-
|
| 336 |
-
def process_new_user_message(message: dict) -> list[dict]:
|
| 337 |
-
content: list[dict] = []
|
| 338 |
-
for path in message.get("files", []):
|
| 339 |
-
kind = _classify_file(path)
|
| 340 |
-
if kind:
|
| 341 |
-
content.append({"type": kind, "url": path})
|
| 342 |
-
content.append({"type": "text", "text": message.message.get("text", "") if hasattr(message, 'message') else message.get("text", "")})
|
| 343 |
-
return content
|
| 344 |
-
|
| 345 |
-
def process_history(history: list[dict]) -> list[dict]:
|
| 346 |
-
messages: list[dict] = []
|
| 347 |
-
for item in history:
|
| 348 |
-
if item["role"] == "assistant":
|
| 349 |
-
if (item.get("metadata") or {}).get("title") == "Reasoning":
|
| 350 |
-
continue
|
| 351 |
-
text_parts = [p["text"] for p in item["content"] if p.get("type") == "text"]
|
| 352 |
-
messages.append(
|
| 353 |
-
{
|
| 354 |
-
"role": "assistant",
|
| 355 |
-
"content": [{"type": "text", "text": " ".join(text_parts)}],
|
| 356 |
-
}
|
| 357 |
-
)
|
| 358 |
-
else:
|
| 359 |
-
user_content: list[dict] = []
|
| 360 |
-
for part in item["content"]:
|
| 361 |
-
if part.get("type") == "text":
|
| 362 |
-
user_content.append({"type": "text", "text": part["text"]})
|
| 363 |
-
elif part.get("type") == "file":
|
| 364 |
-
filepath = part["file"]["path"]
|
| 365 |
-
kind = _classify_file(filepath)
|
| 366 |
-
if kind:
|
| 367 |
-
user_content.append({"type": kind, "url": filepath})
|
| 368 |
-
if user_content:
|
| 369 |
-
messages.append({"role": "user", "content": user_content})
|
| 370 |
-
return messages
|
| 371 |
-
|
| 372 |
-
class StopOnSignal(StoppingCriteria):
|
| 373 |
-
def __init__(self) -> None:
|
| 374 |
-
self.stopped = False
|
| 375 |
-
|
| 376 |
-
def __call__(self, input_ids: torch.Tensor, scores: torch.Tensor, **kwargs: object) -> bool:
|
| 377 |
-
return self.stopped
|
| 378 |
-
|
| 379 |
-
@spaces.GPU(duration=120)
|
| 380 |
-
@torch.inference_mode()
|
| 381 |
-
def _generate_on_gpu(inputs: BatchFeature, max_new_tokens: int, thinking: bool) -> Iterator[str]:
|
| 382 |
-
inputs = inputs.to(device=gemma_model.device, dtype=torch.bfloat16)
|
| 383 |
-
|
| 384 |
-
streamer = TextIteratorStreamer(
|
| 385 |
-
gemma_processor,
|
| 386 |
-
timeout=30.0,
|
| 387 |
-
skip_prompt=True,
|
| 388 |
-
skip_special_tokens=not thinking,
|
| 389 |
-
)
|
| 390 |
-
stop_criteria = StopOnSignal()
|
| 391 |
-
generate_kwargs = {
|
| 392 |
-
**inputs,
|
| 393 |
-
"streamer": streamer,
|
| 394 |
-
"stopping_criteria": [stop_criteria],
|
| 395 |
-
"max_new_tokens": max_new_tokens,
|
| 396 |
-
"disable_compile": True,
|
| 397 |
-
}
|
| 398 |
-
|
| 399 |
-
exception_holder: list[Exception] = []
|
| 400 |
-
|
| 401 |
-
def _generate() -> None:
|
| 402 |
-
try:
|
| 403 |
-
gemma_model.generate(**generate_kwargs)
|
| 404 |
-
except Exception as e:
|
| 405 |
-
exception_holder.append(e)
|
| 406 |
-
|
| 407 |
-
thread = Thread(target=_generate)
|
| 408 |
-
thread.start()
|
| 409 |
-
|
| 410 |
-
chunks: list[str] = []
|
| 411 |
-
try:
|
| 412 |
-
for text in streamer:
|
| 413 |
-
chunks.append(text)
|
| 414 |
-
accumulated = "".join(chunks)
|
| 415 |
-
if thinking:
|
| 416 |
-
yield _strip_special_tokens(accumulated)
|
| 417 |
-
else:
|
| 418 |
-
yield accumulated
|
| 419 |
-
except GeneratorExit:
|
| 420 |
-
stop_criteria.stopped = True
|
| 421 |
-
for _ in streamer:
|
| 422 |
-
pass
|
| 423 |
-
thread.join()
|
| 424 |
-
raise
|
| 425 |
-
|
| 426 |
-
thread.join()
|
| 427 |
-
if exception_holder:
|
| 428 |
-
msg = f"Generation failed: {exception_holder[0]}"
|
| 429 |
-
raise gr.Error(msg)
|
| 430 |
-
|
| 431 |
-
def validate_input(message: dict) -> dict:
|
| 432 |
-
has_text = bool(message.get("text", "").strip())
|
| 433 |
-
has_files = bool(message.get("files"))
|
| 434 |
-
if not (has_text or has_files):
|
| 435 |
-
return gr.validate(False, "Please enter a message or upload a file.")
|
| 436 |
-
|
| 437 |
-
files = message.get("files", [])
|
| 438 |
-
kinds = [_classify_file(f) for f in files]
|
| 439 |
-
kinds = [k for k in kinds if k is not None]
|
| 440 |
-
unique_kinds = set(kinds)
|
| 441 |
-
|
| 442 |
-
if len(unique_kinds) > 1:
|
| 443 |
-
return gr.validate(False, "Please upload only one type of media at a time.")
|
| 444 |
-
if kinds.count("audio") > 1:
|
| 445 |
-
return gr.validate(False, "Only one audio file can be uploaded at a time.")
|
| 446 |
-
if kinds.count("video") > 1:
|
| 447 |
-
return gr.validate(False, "Only one video file can be uploaded at a time.")
|
| 448 |
-
|
| 449 |
-
return gr.validate(True, "")
|
| 450 |
-
|
| 451 |
-
def _has_media_type(messages: list[dict], media_type: str) -> bool:
|
| 452 |
-
return any(c.get("type") == media_type for m in messages for c in m["content"])
|
| 453 |
-
|
| 454 |
-
def generate(
|
| 455 |
-
message: dict,
|
| 456 |
-
history: list[dict],
|
| 457 |
-
thinking: bool = False,
|
| 458 |
-
max_new_tokens: int = 1024,
|
| 459 |
-
max_soft_tokens: int = 280,
|
| 460 |
-
system_prompt: str = "",
|
| 461 |
-
) -> Iterator[str]:
|
| 462 |
-
messages: list[dict] = []
|
| 463 |
-
if system_prompt:
|
| 464 |
-
messages.append({"role": "system", "content": [{"type": "text", "text": system_prompt}]})
|
| 465 |
-
|
| 466 |
-
messages.extend(process_history(history))
|
| 467 |
-
messages.append({"role": "user", "content": process_new_user_message(message)})
|
| 468 |
-
|
| 469 |
-
template_kwargs: dict = {
|
| 470 |
-
"tokenize": True,
|
| 471 |
-
"return_dict": True,
|
| 472 |
-
"return_tensors": "pt",
|
| 473 |
-
"add_generation_prompt": True,
|
| 474 |
-
"load_audio_from_video": _has_media_type(messages, "video"),
|
| 475 |
-
"processor_kwargs": {"images_kwargs": {"max_soft_tokens": max_soft_tokens}},
|
| 476 |
-
}
|
| 477 |
-
if thinking:
|
| 478 |
-
template_kwargs["enable_thinking"] = True
|
| 479 |
-
|
| 480 |
-
inputs = gemma_processor.apply_chat_template(messages, **template_kwargs)
|
| 481 |
-
|
| 482 |
-
n_tokens = inputs["input_ids"].shape[1]
|
| 483 |
-
if n_tokens > MAX_INPUT_TOKENS:
|
| 484 |
-
msg = f"Input too long ({n_tokens} tokens). Maximum is {MAX_INPUT_TOKENS} tokens."
|
| 485 |
-
raise gr.Error(msg)
|
| 486 |
-
|
| 487 |
-
yield from _generate_on_gpu(inputs=inputs, max_new_tokens=max_new_tokens, thinking=thinking)
|
| 488 |
-
|
| 489 |
-
|
| 490 |
-
# ==========================================
|
| 491 |
-
# رابط کاربری مشترک (Gradio Blocks)
|
| 492 |
-
# ==========================================
|
| 493 |
-
examples_chat = [
|
| 494 |
-
[{"text": "What is the capital of France?", "files": []}],
|
| 495 |
-
[{"text": "Explain quantum entanglement in simple terms.", "files": []}],
|
| 496 |
-
[{"text": "Describe this image.", "files": ["https://news.bbc.co.uk/media/images/38107000/jpg/_38107299_ronaldogoal_ap_300.jpg"]}],
|
| 497 |
-
[{"text": "Transcribe the audio.", "files": ["https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3"]}],
|
| 498 |
-
[{"text": "What is happening in this video?", "files": ["https://huggingface.co/datasets/merve/vlm_test_images/resolve/main/concert.mp4"]}],
|
| 499 |
-
]
|
| 500 |
-
|
| 501 |
-
with gr.Blocks(theme=gr.themes.Default(primary_hue="blue")) as demo:
|
| 502 |
gr.Markdown(
|
| 503 |
"""
|
| 504 |
-
#
|
| 505 |
-
###
|
|
|
|
|
|
|
|
|
|
| 506 |
"""
|
| 507 |
)
|
| 508 |
|
| 509 |
-
with gr.
|
| 510 |
-
|
| 511 |
-
|
| 512 |
-
|
| 513 |
-
|
| 514 |
-
|
| 515 |
-
|
| 516 |
-
|
| 517 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 518 |
)
|
| 519 |
-
|
| 520 |
-
|
| 521 |
-
|
| 522 |
-
|
| 523 |
-
|
|
|
|
| 524 |
)
|
| 525 |
-
|
| 526 |
-
|
| 527 |
-
label="
|
| 528 |
-
|
|
|
|
|
|
|
| 529 |
)
|
| 530 |
-
|
| 531 |
-
submit_btn = gr.Button("Transcribe & Correct / شروع تبدیل و اصلاح هوشمند", variant="primary")
|
| 532 |
-
|
| 533 |
-
with gr.Column(scale=1):
|
| 534 |
-
with gr.Tabs():
|
| 535 |
-
with gr.Tab("Full Text / متن کامل"):
|
| 536 |
-
text_output = gr.Textbox(
|
| 537 |
-
label="Transcription Output (AI Corrected) / متن پیوسته (اصلاح شده)",
|
| 538 |
-
lines=14,
|
| 539 |
-
buttons=["copy"],
|
| 540 |
-
interactive=False
|
| 541 |
-
)
|
| 542 |
-
with gr.Tab("SRT Subtitles / زیرنویس استاندارد"):
|
| 543 |
-
srt_output = gr.Textbox(
|
| 544 |
-
label="Subtitles (SRT Format) / فایل زیرنویس",
|
| 545 |
-
lines=14,
|
| 546 |
-
buttons=["copy"],
|
| 547 |
-
interactive=False
|
| 548 |
-
)
|
| 549 |
-
with gr.Tab("Word Timestamps / زمانبندی کلمات (JSON)"):
|
| 550 |
-
json_output = gr.Textbox(
|
| 551 |
-
label="Word-Level Timestamps (JSON) / زمان خام کلمات",
|
| 552 |
-
lines=14,
|
| 553 |
-
buttons=["copy"],
|
| 554 |
-
interactive=False
|
| 555 |
-
)
|
| 556 |
|
| 557 |
-
|
| 558 |
-
|
| 559 |
-
|
| 560 |
-
|
| 561 |
-
|
| 562 |
-
|
| 563 |
-
|
| 564 |
-
# -------- TAB 2: Chat Interface --------
|
| 565 |
-
with gr.Tab("AI Chatbot / چتبات چندرسانهای"):
|
| 566 |
-
gr.ChatInterface(
|
| 567 |
-
fn=generate,
|
| 568 |
-
validator=validate_input,
|
| 569 |
-
chatbot=gr.Chatbot(
|
| 570 |
-
scale=1,
|
| 571 |
-
latex_delimiters=[
|
| 572 |
-
{"left": "$$", "right": "$$", "display": True},
|
| 573 |
-
{"left": "$", "right": "$", "display": False},
|
| 574 |
-
{"left": "\\(", "right": "\\)", "display": False},
|
| 575 |
-
{"left": "\\[", "right": "\\]", "display": True},
|
| 576 |
-
],
|
| 577 |
-
reasoning_tags=[(THINKING_START, THINKING_END)],
|
| 578 |
-
),
|
| 579 |
-
textbox=gr.MultimodalTextbox(
|
| 580 |
-
sources=["upload", "microphone"],
|
| 581 |
-
file_types=[*IMAGE_FILE_TYPES, *AUDIO_FILE_TYPES, *VIDEO_FILE_TYPES],
|
| 582 |
-
file_count="multiple",
|
| 583 |
-
autofocus=True,
|
| 584 |
-
stop_btn=True,
|
| 585 |
-
),
|
| 586 |
-
multimodal=True,
|
| 587 |
-
additional_inputs=[
|
| 588 |
-
gr.Checkbox(label="Thinking", value=False),
|
| 589 |
-
gr.Slider(label="Max New Tokens", minimum=100, maximum=4000, step=10, value=2000),
|
| 590 |
-
gr.Dropdown(
|
| 591 |
-
label="Image Token Budget",
|
| 592 |
-
info="Higher values preserve more visual detail. Lower values are faster.",
|
| 593 |
-
choices=[70, 140, 280, 560, 1120],
|
| 594 |
-
value=280,
|
| 595 |
-
),
|
| 596 |
-
gr.Textbox(label="System Prompt", value=""),
|
| 597 |
-
],
|
| 598 |
-
additional_inputs_accordion=gr.Accordion("Settings", open=False),
|
| 599 |
-
title="",
|
| 600 |
-
examples=examples_chat,
|
| 601 |
-
run_examples_on_click=False,
|
| 602 |
-
cache_examples=False,
|
| 603 |
-
delete_cache=(1800, 1800),
|
| 604 |
-
)
|
| 605 |
|
| 606 |
if __name__ == "__main__":
|
| 607 |
-
|
|
|
|
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import gradio as gr
|
|
|
|
| 3 |
import torch
|
| 4 |
import librosa
|
| 5 |
+
import json
|
| 6 |
+
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
|
| 7 |
+
import spaces
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
|
| 9 |
+
# انتخاب نسخه مرجع جهانی چندزبانه OpenAI Whisper V3
|
| 10 |
+
model_id = "openai/whisper-large-v3"
|
|
|
|
|
|
|
| 11 |
|
| 12 |
+
# ۱. انتخاب بهینهترین فرمت عددی برای سختافزار A100/H200
|
| 13 |
if torch.cuda.is_available() and torch.cuda.is_bf16_supported():
|
| 14 |
+
torch_dtype = torch.bfloat16
|
| 15 |
+
print("Using optimized BFloat16 for A100/H200")
|
| 16 |
else:
|
| 17 |
+
torch_dtype = torch.float16
|
| 18 |
+
print("Using Float16")
|
| 19 |
|
| 20 |
+
# ۲. فعالسازی مکانیسم شتابدهنده توجه (Attention)
|
| 21 |
attn_implementation = "sdpa"
|
| 22 |
try:
|
| 23 |
import flash_attn
|
| 24 |
attn_implementation = "flash_attention_2"
|
| 25 |
+
print("Using Flash Attention 2 for maximum hardware utilization")
|
| 26 |
except ImportError:
|
| 27 |
+
print("Flash Attention 2 not found. Falling back to high-performance PyTorch SDPA")
|
| 28 |
|
| 29 |
+
# ۳. بارگذاری بهینه مدل چندزبانه با تکیه بر Safetensors و Memory optimization
|
| 30 |
+
model = AutoModelForSpeechSeq2Seq.from_pretrained(
|
| 31 |
+
model_id,
|
| 32 |
+
torch_dtype=torch_dtype,
|
| 33 |
low_cpu_mem_usage=True,
|
| 34 |
use_safetensors=True,
|
| 35 |
attn_implementation=attn_implementation
|
| 36 |
)
|
| 37 |
|
| 38 |
+
# انتقال مدل به حافظه گرافیکی در صورت در دسترس بودن
|
| 39 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 40 |
+
model.to(device)
|
| 41 |
|
| 42 |
+
processor = AutoProcessor.from_pretrained(model_id)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
|
| 44 |
+
# تعریف متن راهنمای پرامپت فارسی به صورت سراسری
|
| 45 |
+
prompt_text = "کلمات دقیقاً همانطور که تلفظ میشوند، با رعایت حروف اضافه و ساختار عامیانه، محاورهای و شکسته نوشته شوند."
|
| 46 |
|
| 47 |
+
# ۴. ساخت پایپلاین تشخیص گفتار چندزبانه با استفاده از پارامتر اصلاح شدهی dtype
|
| 48 |
+
pipe = pipeline(
|
| 49 |
+
"automatic-speech-recognition",
|
| 50 |
+
model=model,
|
| 51 |
+
tokenizer=processor.tokenizer,
|
| 52 |
+
feature_extractor=processor.feature_extractor,
|
| 53 |
+
dtype=torch_dtype,
|
| 54 |
+
device=device,
|
| 55 |
)
|
| 56 |
|
| 57 |
+
# نقشه کدهای زبانی برای منوی کشویی انتخاب زبان
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
LANGUAGE_MAPPING = {
|
| 59 |
"Auto-Detect (تشخیص خودکار)": None,
|
| 60 |
"Persian / فارسی": "fa",
|
|
|
|
| 71 |
"Korean / کرهای": "ko"
|
| 72 |
}
|
| 73 |
|
| 74 |
+
# تابع کمکی برای فرمتبندی زمان در ساختار استاندارد زیرنویس (HH:MM:SS,mmm)
|
| 75 |
def format_timestamp(seconds):
|
| 76 |
hours = int(seconds // 3600)
|
| 77 |
minutes = int((seconds % 3600) // 60)
|
|
|
|
| 79 |
millis = int(round((seconds % 1) * 1000))
|
| 80 |
return f"{hours:02d}:{minutes:02d}:{secs:02d},{millis:03d}"
|
| 81 |
|
| 82 |
+
# مبدل دادههای کلمه به کلمه ویسپر به ساختار استاندارد خوانای SRT
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
def create_srt_from_chunks(chunks, max_words_per_segment=6, max_duration_per_segment=3.0):
|
| 84 |
srt_lines = []
|
| 85 |
segment_idx = 1
|
|
|
|
| 97 |
if start is None:
|
| 98 |
continue
|
| 99 |
if end is None:
|
| 100 |
+
end = start + 0.3 # مقدار جایگزین در صورت عدم ثبت زمان پایان توکن
|
| 101 |
|
| 102 |
if current_start is None:
|
| 103 |
current_start = start
|
|
|
|
| 105 |
current_segment_words.append(text)
|
| 106 |
duration = end - current_start
|
| 107 |
|
| 108 |
+
# دستهبندی کلمات در قالب خطوط زیرنویس خوانا (مثلا حداکثر ۶ کلمه یا ۳ ثانیه برای هر خط)
|
| 109 |
if len(current_segment_words) >= max_words_per_segment or duration >= max_duration_per_segment:
|
| 110 |
line_text = " ".join(current_segment_words)
|
| 111 |
srt_lines.append(f"{segment_idx}")
|
|
|
|
| 126 |
|
| 127 |
return "\n".join(srt_lines)
|
| 128 |
|
| 129 |
+
# تابع تولید دیتای خام زمانبندی کلمات (JSON) برای دسترسی توسعهدهندگان
|
| 130 |
def get_raw_word_timestamps(chunks):
|
| 131 |
word_timestamps = []
|
| 132 |
for chunk in chunks:
|
|
|
|
| 140 |
})
|
| 141 |
return json.dumps(word_timestamps, ensure_ascii=False, indent=2)
|
| 142 |
|
| 143 |
+
@spaces.GPU
|
| 144 |
def transcribe_audio(audio_path, language_selection, optimize_persian):
|
| 145 |
if audio_path is None:
|
| 146 |
return "Please upload or record an audio file first. / لطفاً ابتدا فایل صوتی را آپلود یا ضبط کنید.", "", ""
|
| 147 |
|
| 148 |
try:
|
| 149 |
+
# ۱. بارگذاری ایمن فایل صوتی به فرکانس ۱۶۰۰۰ هرتز (سازگار با همه فرمتها مانند میکروفون وبام یا m4a)
|
| 150 |
audio_data, sampling_rate = librosa.load(audio_path, sr=16000)
|
| 151 |
+
|
| 152 |
+
# ۲. پیکربندی پویا بر اساس زبان انتخاب شده توسط کاربر
|
| 153 |
lang_code = LANGUAGE_MAPPING.get(language_selection, None)
|
| 154 |
|
| 155 |
+
# تنظیمات پایه تولید متن با استفاده از روش پیشرفته Beam Search برای افزایش حداکثری دقت تفکیک کلمات
|
| 156 |
generate_kwargs = {
|
| 157 |
"task": "transcribe",
|
| 158 |
"do_sample": False,
|
| 159 |
+
"num_beams": 5 # ارزیابی موازی ۵ مسیر واژگانی برای کاهش خطاهای شنیداری صوتی
|
| 160 |
}
|
| 161 |
|
| 162 |
+
# اگر کاربر زبانی به جز تشخیص خودکار انتخاب کرده باشد
|
| 163 |
if lang_code is not None:
|
| 164 |
generate_kwargs["language"] = lang_code
|
| 165 |
|
| 166 |
+
# ۳. اعمال پرامپت اختصاصی فارسی فقط در صورت تمایل کاربر (به عنوان تانسور ۱ بعدی پایتورچِ روی GPU)
|
| 167 |
if optimize_persian:
|
| 168 |
+
prompt_ids = processor.get_prompt_ids(prompt_text, return_tensors="pt")
|
| 169 |
+
prompt_ids = prompt_ids.to(device)
|
| 170 |
generate_kwargs["prompt_ids"] = prompt_ids
|
| 171 |
+
generate_kwargs["language"] = "fa" # اجبار به فارسی در صورت فعال بودن این گزینه
|
| 172 |
|
| 173 |
+
# ۴. پردازش نهایی با حداکثر ظرفیت سختافزاری روی A100/H200 و درخواست ثبت تایماستمپ کلمات
|
| 174 |
+
result = pipe(
|
| 175 |
{"raw": audio_data, "sampling_rate": sampling_rate},
|
| 176 |
chunk_length_s=30,
|
| 177 |
stride_length_s=(6, 6),
|
| 178 |
batch_size=24,
|
| 179 |
+
return_timestamps="word", # فعالسازی ثبت دقیق زمانبندی برای هر کلمه
|
| 180 |
generate_kwargs=generate_kwargs
|
| 181 |
)
|
| 182 |
|
| 183 |
full_text = result["text"].strip()
|
| 184 |
chunks = result.get("chunks", [])
|
| 185 |
|
| 186 |
+
# ۵. تولید ساختارهای سهگانه دیتای خروجی
|
| 187 |
srt_content = create_srt_from_chunks(chunks)
|
| 188 |
word_timestamps_json = get_raw_word_timestamps(chunks)
|
| 189 |
|
| 190 |
+
# بازگرداندن هر سه خروجی در قالب تاپل
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 191 |
return full_text, srt_content, word_timestamps_json
|
| 192 |
|
| 193 |
except Exception as e:
|
| 194 |
return f"An error occurred during processing / خطایی در حین پردازش رخ داد: {str(e)}", "", ""
|
| 195 |
|
| 196 |
+
# رابط کاربری چندزبانه و مدرن بر پایه Gradio 6.0
|
| 197 |
+
with gr.Blocks() as demo:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
gr.Markdown(
|
| 199 |
"""
|
| 200 |
+
# 🌐 Global Speech-to-Text & Subtitle Studio (Whisper Large V3)
|
| 201 |
+
### سیستم هوشمند و بینالمللی تبدیل گفتار به متن و زیرنویس خودکار (بهینهسازی شده روی A100/H200)
|
| 202 |
+
|
| 203 |
+
*Supports over 99 languages with automatic detection and Word-Level Timestamps.*
|
| 204 |
+
*پشتیبانی از بیش از ۹۹ زبان زنده دنیا به همراه تشخیص خودکار زبان و زمانبندی دقیق کلمه به کلمه.*
|
| 205 |
"""
|
| 206 |
)
|
| 207 |
|
| 208 |
+
with gr.Row():
|
| 209 |
+
with gr.Column(scale=1):
|
| 210 |
+
audio_input = gr.Audio(
|
| 211 |
+
sources=["upload", "microphone"],
|
| 212 |
+
type="filepath",
|
| 213 |
+
label="Audio Input / ورودی صدا"
|
| 214 |
+
)
|
| 215 |
+
|
| 216 |
+
language_dd = gr.Dropdown(
|
| 217 |
+
choices=list(LANGUAGE_MAPPING.keys()),
|
| 218 |
+
value="Auto-Detect (تشخیص خودکار)",
|
| 219 |
+
label="Select Language / انتخاب زبان صوتی"
|
| 220 |
+
)
|
| 221 |
+
|
| 222 |
+
optimize_fa_chk = gr.Checkbox(
|
| 223 |
+
label="Optimize for Spoken Persian / بهینهسازی برای محاوره عامیانه فارسی (فقط برای ویسهای فارسی)",
|
| 224 |
+
value=False
|
| 225 |
+
)
|
| 226 |
+
|
| 227 |
+
submit_btn = gr.Button("Transcribe / شروع فرآیند تبدیل", variant="primary")
|
| 228 |
+
|
| 229 |
+
with gr.Column(scale=1):
|
| 230 |
+
with gr.Tabs():
|
| 231 |
+
with gr.Tab("Full Text / متن کامل"):
|
| 232 |
+
text_output = gr.Textbox(
|
| 233 |
+
label="Transcription Output / متن پیوسته",
|
| 234 |
+
lines=14,
|
| 235 |
+
buttons=["copy"],
|
| 236 |
+
interactive=False
|
| 237 |
)
|
| 238 |
+
with gr.Tab("SRT Subtitles / زیرنویس استاندارد"):
|
| 239 |
+
srt_output = gr.Textbox(
|
| 240 |
+
label="Subtitles (SRT Format) / فایل زیرنویس",
|
| 241 |
+
lines=14,
|
| 242 |
+
buttons=["copy"],
|
| 243 |
+
interactive=False
|
| 244 |
)
|
| 245 |
+
with gr.Tab("Word Timestamps / زمانبندی کلمات (JSON)"):
|
| 246 |
+
json_output = gr.Textbox(
|
| 247 |
+
label="Word-Level Timestamps (JSON) / زمان خام کلمات",
|
| 248 |
+
lines=14,
|
| 249 |
+
buttons=["copy"],
|
| 250 |
+
interactive=False
|
| 251 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 252 |
|
| 253 |
+
submit_btn.click(
|
| 254 |
+
fn=transcribe_audio,
|
| 255 |
+
inputs=[audio_input, language_dd, optimize_fa_chk],
|
| 256 |
+
outputs=[text_output, srt_output, json_output],
|
| 257 |
+
api_name="transcribe" # نام اختصاصی متد برای فراخوانی API کلاینت
|
| 258 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 259 |
|
| 260 |
if __name__ == "__main__":
|
| 261 |
+
# فعالسازی تم آبیرنگ در زمان راهاندازی بر اساس استانداردهای Gradio 6.0
|
| 262 |
+
demo.launch(theme=gr.themes.Default(primary_hue="blue"))
|