Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -6,7 +6,6 @@ import tempfile
|
|
6 |
import pandas as pd
|
7 |
import requests
|
8 |
from bs4 import BeautifulSoup
|
9 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
10 |
import torch
|
11 |
import whisper
|
12 |
from moviepy.editor import VideoFileClip
|
@@ -16,6 +15,11 @@ import docx
|
|
16 |
import yt_dlp
|
17 |
from functools import lru_cache
|
18 |
import gc
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
# Configure logging
|
21 |
logging.basicConfig(
|
@@ -24,6 +28,11 @@ logging.basicConfig(
|
|
24 |
)
|
25 |
logger = logging.getLogger(__name__)
|
26 |
|
|
|
|
|
|
|
|
|
|
|
27 |
class ModelManager:
|
28 |
_instance = None
|
29 |
|
@@ -37,127 +46,126 @@ class ModelManager:
|
|
37 |
if not self._initialized:
|
38 |
self.tokenizer = None
|
39 |
self.model = None
|
40 |
-
self.news_generator = None
|
41 |
self.whisper_model = None
|
42 |
self._initialized = True
|
|
|
43 |
|
44 |
-
@spaces.GPU(
|
45 |
-
def
|
46 |
-
"""Initialize
|
47 |
try:
|
48 |
-
|
49 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
50 |
|
51 |
-
HUGGINGFACE_TOKEN = os.environ.get('HUGGINGFACE_TOKEN')
|
52 |
-
if not HUGGINGFACE_TOKEN:
|
53 |
-
raise ValueError("HUGGINGFACE_TOKEN environment variable not set")
|
54 |
-
|
55 |
-
logger.info("Starting model initialization...")
|
56 |
-
model_name = "meta-llama/Llama-2-7b-chat-hf"
|
57 |
-
|
58 |
-
# Load tokenizer
|
59 |
logger.info("Loading tokenizer...")
|
60 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
61 |
-
|
62 |
token=HUGGINGFACE_TOKEN,
|
63 |
use_fast=True,
|
64 |
-
model_max_length=512
|
65 |
)
|
66 |
self.tokenizer.pad_token = self.tokenizer.eos_token
|
67 |
-
|
68 |
-
#
|
69 |
-
|
70 |
-
|
71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
token=HUGGINGFACE_TOKEN,
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
use_safetensors=True,
|
77 |
-
# ZeroGPU specific settings
|
78 |
-
max_memory={0: "6GB"},
|
79 |
-
offload_folder="offload",
|
80 |
-
offload_state_dict=True
|
81 |
)
|
82 |
-
|
83 |
-
#
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
"
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
temperature=0.7,
|
95 |
-
top_p=0.95,
|
96 |
-
repetition_penalty=1.2,
|
97 |
-
num_return_sequences=1,
|
98 |
-
early_stopping=True
|
99 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
|
101 |
-
|
|
|
|
|
|
|
102 |
logger.info("Loading Whisper model...")
|
|
|
103 |
self.whisper_model = whisper.load_model(
|
104 |
"tiny",
|
105 |
device="cuda" if torch.cuda.is_available() else "cpu",
|
106 |
download_root="/tmp/whisper"
|
107 |
)
|
108 |
-
|
109 |
-
|
110 |
return True
|
111 |
-
|
112 |
except Exception as e:
|
113 |
-
logger.error(f"Error
|
114 |
-
self.reset_models()
|
115 |
raise
|
116 |
|
117 |
-
def
|
118 |
-
"""
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
|
127 |
-
|
128 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
import gc
|
145 |
-
gc.collect()
|
146 |
-
|
147 |
-
except Exception as e:
|
148 |
-
logger.error(f"Error during model reset: {str(e)}")
|
149 |
-
|
150 |
-
def check_models_initialized(self):
|
151 |
-
"""Check if all models are properly initialized"""
|
152 |
-
if None in (self.tokenizer, self.model, self.news_generator, self.whisper_model):
|
153 |
-
logger.warning("Models not initialized, attempting to initialize...")
|
154 |
-
self.initialize_models()
|
155 |
-
|
156 |
-
def get_models(self):
|
157 |
-
"""Get initialized models, initializing if necessary"""
|
158 |
-
self.check_models_initialized()
|
159 |
-
return self.tokenizer, self.model, self.news_generator, self.whisper_model
|
160 |
-
|
161 |
# Create global model manager instance
|
162 |
model_manager = ModelManager()
|
163 |
|
@@ -188,7 +196,7 @@ def convert_video_to_audio(video_file):
|
|
188 |
try:
|
189 |
video = VideoFileClip(video_file)
|
190 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as temp_file:
|
191 |
-
video.audio.write_audiofile(temp_file.name)
|
192 |
logger.info(f"Video converted to audio: {temp_file.name}")
|
193 |
return temp_file.name
|
194 |
except Exception as e:
|
@@ -208,30 +216,40 @@ def preprocess_audio(audio_file):
|
|
208 |
logger.error(f"Error preprocessing audio: {str(e)}")
|
209 |
raise
|
210 |
|
211 |
-
@spaces.GPU(
|
212 |
def transcribe_audio(file):
|
213 |
"""Transcribe an audio or video file."""
|
214 |
try:
|
215 |
-
|
216 |
|
217 |
if isinstance(file, str) and file.startswith('http'):
|
218 |
file_path = download_social_media_video(file)
|
219 |
elif isinstance(file, str) and file.lower().endswith(('.mp4', '.avi', '.mov', '.mkv')):
|
220 |
file_path = convert_video_to_audio(file)
|
|
|
|
|
221 |
else:
|
222 |
-
|
223 |
|
224 |
logger.info(f"Transcribing audio: {file_path}")
|
225 |
if not os.path.exists(file_path):
|
226 |
raise FileNotFoundError(f"Audio file not found: {file_path}")
|
227 |
|
228 |
with torch.inference_mode():
|
229 |
-
result = whisper_model.transcribe(file_path)
|
230 |
if not result:
|
231 |
raise RuntimeError("Transcription failed to produce results")
|
232 |
|
233 |
transcription = result.get("text", "Error in transcription")
|
234 |
logger.info(f"Transcription completed: {transcription[:50]}...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
return transcription
|
236 |
except Exception as e:
|
237 |
logger.error(f"Error transcribing: {str(e)}")
|
@@ -247,7 +265,7 @@ def read_document(document_path):
|
|
247 |
elif document_path.endswith(".docx"):
|
248 |
doc = docx.Document(document_path)
|
249 |
return "\n".join([paragraph.text for paragraph in doc.paragraphs])
|
250 |
-
elif document_path.endswith(".xlsx"):
|
251 |
return pd.read_excel(document_path).to_string()
|
252 |
elif document_path.endswith(".csv"):
|
253 |
return pd.read_csv(document_path).to_string()
|
@@ -260,17 +278,42 @@ def read_document(document_path):
|
|
260 |
@lru_cache(maxsize=32)
|
261 |
def read_url(url):
|
262 |
"""Read the content of a URL."""
|
|
|
|
|
|
|
263 |
try:
|
264 |
-
|
|
|
|
|
|
|
265 |
response.raise_for_status()
|
266 |
soup = BeautifulSoup(response.content, 'html.parser')
|
267 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
268 |
except Exception as e:
|
269 |
logger.error(f"Error reading URL: {str(e)}")
|
270 |
return f"Error reading URL: {str(e)}"
|
271 |
|
272 |
def process_social_content(url):
|
273 |
"""Process social media content."""
|
|
|
|
|
|
|
274 |
try:
|
275 |
text_content = read_url(url)
|
276 |
try:
|
@@ -287,11 +330,20 @@ def process_social_content(url):
|
|
287 |
logger.error(f"Error processing social content: {str(e)}")
|
288 |
return None
|
289 |
|
290 |
-
@spaces.GPU(
|
291 |
def generate_news(instructions, facts, size, tone, *args):
|
|
|
292 |
try:
|
293 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
294 |
|
|
|
295 |
knowledge_base = {
|
296 |
"instructions": instructions,
|
297 |
"facts": facts,
|
@@ -301,78 +353,123 @@ def generate_news(instructions, facts, size, tone, *args):
|
|
301 |
"social_content": []
|
302 |
}
|
303 |
|
|
|
304 |
num_audios = 5 * 3
|
305 |
num_social_urls = 3 * 3
|
306 |
num_urls = 5
|
307 |
|
|
|
308 |
audios = args[:num_audios]
|
309 |
social_urls = args[num_audios:num_audios+num_social_urls]
|
310 |
urls = args[num_audios+num_social_urls:num_audios+num_social_urls+num_urls]
|
311 |
documents = args[num_audios+num_social_urls+num_urls:]
|
312 |
|
|
|
|
|
313 |
for url in urls:
|
314 |
-
if url:
|
315 |
content = read_url(url)
|
316 |
if content and not content.startswith("Error"):
|
317 |
knowledge_base["url_content"].append(content)
|
318 |
|
|
|
|
|
319 |
for document in documents:
|
320 |
if document is not None:
|
321 |
content = read_document(document.name)
|
322 |
if content and not content.startswith("Error"):
|
323 |
knowledge_base["document_content"].append(content)
|
324 |
|
|
|
|
|
325 |
for i in range(0, len(audios), 3):
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
"
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
for i in range(0, len(social_urls), 3):
|
335 |
-
social_url, social_name, social_context = social_urls[i:i+3]
|
336 |
-
if social_url:
|
337 |
-
social_content = process_social_content(social_url)
|
338 |
-
if social_content:
|
339 |
-
knowledge_base["social_content"].append({
|
340 |
-
"url": social_url,
|
341 |
-
"name": social_name,
|
342 |
-
"context": social_context,
|
343 |
-
"text": social_content["text"],
|
344 |
-
"video": social_content["video"]
|
345 |
})
|
346 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
347 |
transcriptions_text = ""
|
348 |
raw_transcriptions = ""
|
349 |
|
|
|
|
|
350 |
for idx, data in enumerate(knowledge_base["audio_data"]):
|
351 |
if data["audio"] is not None:
|
352 |
transcription = transcribe_audio(data["audio"])
|
353 |
-
if not transcription.startswith("Error"):
|
354 |
-
transcriptions_text += f'"{transcription}" - {data["name"]}, {data["position"]}\n'
|
355 |
raw_transcriptions += f'[Audio/Video {idx + 1}]: "{transcription}" - {data["name"]}, {data["position"]}\n\n'
|
356 |
|
357 |
-
|
|
|
358 |
if data["text"] and not str(data["text"]).startswith("Error"):
|
359 |
-
|
360 |
-
|
|
|
|
|
|
|
|
|
361 |
if data["video"] and not str(data["video"]).startswith("Error"):
|
362 |
-
video_transcription = f'[Social media video]: "{data["video"]}" - {data["name"]}, {data["context"]}\n'
|
363 |
transcriptions_text += video_transcription
|
364 |
-
raw_transcriptions += video_transcription
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
369 |
|
370 |
-
|
|
|
371 |
|
372 |
Instructions: {knowledge_base["instructions"]}
|
|
|
373 |
Facts: {knowledge_base["facts"]}
|
374 |
-
|
375 |
-
Additional content from
|
|
|
|
|
|
|
|
|
376 |
|
377 |
Use these transcriptions as direct and indirect quotes:
|
378 |
{transcriptions_text}
|
@@ -380,7 +477,7 @@ Use these transcriptions as direct and indirect quotes:
|
|
380 |
Follow these requirements:
|
381 |
- Write a title
|
382 |
- Write a 15-word hook that complements the title
|
383 |
-
- Write the body with {size} words
|
384 |
- Use a {tone} tone
|
385 |
- Answer the 5 Ws (Who, What, When, Where, Why) in the first paragraph
|
386 |
- Use at least 80% direct quotes (in quotation marks)
|
@@ -388,27 +485,41 @@ Follow these requirements:
|
|
388 |
- Do not invent information
|
389 |
- Be rigorous with the provided facts [/INST]"""
|
390 |
|
391 |
-
# Optimize
|
392 |
-
|
393 |
|
394 |
-
# Generate
|
395 |
with torch.inference_mode():
|
396 |
try:
|
397 |
-
|
398 |
-
|
399 |
-
|
400 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
401 |
do_sample=True,
|
402 |
temperature=0.7,
|
403 |
top_p=0.95,
|
404 |
repetition_penalty=1.2,
|
405 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
406 |
)
|
407 |
|
408 |
-
#
|
409 |
-
|
410 |
-
|
411 |
-
news_article = news_article.replace('[INST]', '').replace('[/INST]', '').strip()
|
412 |
|
413 |
except Exception as gen_error:
|
414 |
logger.error(f"Error in text generation: {str(gen_error)}")
|
@@ -419,122 +530,173 @@ Follow these requirements:
|
|
419 |
except Exception as e:
|
420 |
logger.error(f"Error generating news: {str(e)}")
|
421 |
try:
|
422 |
-
#
|
423 |
-
model_manager.reset_models()
|
424 |
-
|
425 |
-
logger.
|
426 |
-
|
427 |
-
logger.error(f"Failed to reinitialize models: {str(reinit_error)}")
|
428 |
-
return f"Error generating the news article: {str(e)}", ""
|
429 |
|
430 |
def create_demo():
|
431 |
-
with gr.Blocks() as demo:
|
432 |
-
gr.Markdown("
|
|
|
433 |
|
434 |
with gr.Row():
|
435 |
with gr.Column(scale=2):
|
436 |
instrucciones = gr.Textbox(
|
437 |
label="Instrucciones para la noticia",
|
|
|
438 |
lines=2
|
439 |
)
|
440 |
hechos = gr.Textbox(
|
441 |
-
label="
|
|
|
442 |
lines=4
|
443 |
)
|
444 |
-
|
445 |
-
|
446 |
-
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
|
451 |
-
|
452 |
-
|
|
|
|
|
|
|
|
|
|
|
453 |
|
454 |
with gr.Column(scale=3):
|
455 |
inputs_list = [instrucciones, hechos, tamaño, tono]
|
456 |
|
457 |
with gr.Tabs():
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
-
|
466 |
-
|
467 |
-
|
468 |
-
|
469 |
-
|
470 |
-
|
471 |
-
|
472 |
-
|
473 |
-
|
474 |
-
|
475 |
-
|
476 |
-
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
-
|
486 |
-
|
487 |
-
|
488 |
-
|
489 |
-
|
490 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
491 |
url = gr.Textbox(
|
492 |
label=f"URL {i}",
|
493 |
placeholder="https://..."
|
494 |
)
|
495 |
inputs_list.append(url)
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
|
500 |
-
|
501 |
-
|
502 |
-
|
503 |
-
)
|
504 |
-
|
505 |
-
|
506 |
-
|
507 |
-
|
508 |
-
|
509 |
-
|
510 |
-
|
511 |
-
|
512 |
-
|
513 |
-
|
514 |
-
|
515 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
516 |
|
517 |
with gr.Row():
|
518 |
-
generar = gr.Button("Generar
|
519 |
-
|
520 |
-
|
521 |
-
|
522 |
-
|
523 |
-
|
524 |
-
|
525 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
526 |
|
|
|
527 |
generar.click(
|
528 |
fn=generate_news,
|
529 |
inputs=inputs_list,
|
530 |
outputs=[noticia_output, transcripciones_output]
|
531 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
532 |
|
533 |
return demo
|
534 |
|
535 |
if __name__ == "__main__":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
536 |
demo = create_demo()
|
537 |
-
demo.queue()
|
538 |
demo.launch(
|
539 |
share=True,
|
540 |
server_name="0.0.0.0",
|
|
|
6 |
import pandas as pd
|
7 |
import requests
|
8 |
from bs4 import BeautifulSoup
|
|
|
9 |
import torch
|
10 |
import whisper
|
11 |
from moviepy.editor import VideoFileClip
|
|
|
15 |
import yt_dlp
|
16 |
from functools import lru_cache
|
17 |
import gc
|
18 |
+
import time
|
19 |
+
from huggingface_hub import login
|
20 |
+
from transformers import AutoTokenizer, BitsAndBytesConfig
|
21 |
+
from unsloth import FastLanguageModel
|
22 |
+
import tqdm
|
23 |
|
24 |
# Configure logging
|
25 |
logging.basicConfig(
|
|
|
28 |
)
|
29 |
logger = logging.getLogger(__name__)
|
30 |
|
31 |
+
# Login to Hugging Face Hub if token is available
|
32 |
+
HUGGINGFACE_TOKEN = os.environ.get('HUGGINGFACE_TOKEN')
|
33 |
+
if HUGGINGFACE_TOKEN:
|
34 |
+
login(token=HUGGINGFACE_TOKEN)
|
35 |
+
|
36 |
class ModelManager:
|
37 |
_instance = None
|
38 |
|
|
|
46 |
if not self._initialized:
|
47 |
self.tokenizer = None
|
48 |
self.model = None
|
|
|
49 |
self.whisper_model = None
|
50 |
self._initialized = True
|
51 |
+
self.last_used = time.time()
|
52 |
|
53 |
+
@spaces.GPU()
|
54 |
+
def initialize_llm(self):
|
55 |
+
"""Initialize LLM model with unsloth optimization"""
|
56 |
try:
|
57 |
+
MODEL_NAME = "meta-llama/Llama-2-7b-chat-hf"
|
|
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
logger.info("Loading tokenizer...")
|
60 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
61 |
+
MODEL_NAME,
|
62 |
token=HUGGINGFACE_TOKEN,
|
63 |
use_fast=True,
|
|
|
64 |
)
|
65 |
self.tokenizer.pad_token = self.tokenizer.eos_token
|
66 |
+
|
67 |
+
# Configure 4-bit quantization
|
68 |
+
bnb_config = BitsAndBytesConfig(
|
69 |
+
load_in_4bit=True,
|
70 |
+
bnb_4bit_quant_type="nf4",
|
71 |
+
bnb_4bit_compute_dtype=torch.float16,
|
72 |
+
bnb_4bit_use_double_quant=True
|
73 |
+
)
|
74 |
+
|
75 |
+
logger.info("Loading and optimizing model with unsloth...")
|
76 |
+
# Use unsloth to load and optimize the model
|
77 |
+
self.model, self.tokenizer = FastLanguageModel.from_pretrained(
|
78 |
+
model_name=MODEL_NAME,
|
79 |
token=HUGGINGFACE_TOKEN,
|
80 |
+
quantization_config=bnb_config,
|
81 |
+
max_seq_length=2048,
|
82 |
+
device_map="auto"
|
|
|
|
|
|
|
|
|
|
|
83 |
)
|
84 |
+
|
85 |
+
# Optimize with unsloth
|
86 |
+
self.model = FastLanguageModel.get_peft_model(
|
87 |
+
self.model,
|
88 |
+
r=16,
|
89 |
+
target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
|
90 |
+
"gate_proj", "up_proj", "down_proj"],
|
91 |
+
lora_alpha=16,
|
92 |
+
lora_dropout=0,
|
93 |
+
bias="none",
|
94 |
+
use_gradient_checkpointing=True,
|
95 |
+
random_state=3407
|
|
|
|
|
|
|
|
|
|
|
96 |
)
|
97 |
+
|
98 |
+
logger.info("LLM initialized successfully")
|
99 |
+
self.last_used = time.time()
|
100 |
+
return True
|
101 |
+
|
102 |
+
except Exception as e:
|
103 |
+
logger.error(f"Error initializing LLM: {str(e)}")
|
104 |
+
raise
|
105 |
|
106 |
+
@spaces.GPU()
|
107 |
+
def initialize_whisper(self):
|
108 |
+
"""Initialize Whisper model for audio transcription"""
|
109 |
+
try:
|
110 |
logger.info("Loading Whisper model...")
|
111 |
+
# Using tiny model for efficiency but can be changed based on needs
|
112 |
self.whisper_model = whisper.load_model(
|
113 |
"tiny",
|
114 |
device="cuda" if torch.cuda.is_available() else "cpu",
|
115 |
download_root="/tmp/whisper"
|
116 |
)
|
117 |
+
logger.info("Whisper model initialized successfully")
|
118 |
+
self.last_used = time.time()
|
119 |
return True
|
|
|
120 |
except Exception as e:
|
121 |
+
logger.error(f"Error initializing Whisper: {str(e)}")
|
|
|
122 |
raise
|
123 |
|
124 |
+
def check_llm_initialized(self):
|
125 |
+
"""Check if LLM is initialized and initialize if needed"""
|
126 |
+
if self.tokenizer is None or self.model is None:
|
127 |
+
logger.info("LLM not initialized, initializing...")
|
128 |
+
self.initialize_llm()
|
129 |
+
self.last_used = time.time()
|
130 |
+
|
131 |
+
def check_whisper_initialized(self):
|
132 |
+
"""Check if Whisper model is initialized and initialize if needed"""
|
133 |
+
if self.whisper_model is None:
|
134 |
+
logger.info("Whisper model not initialized, initializing...")
|
135 |
+
self.initialize_whisper()
|
136 |
+
self.last_used = time.time()
|
137 |
+
|
138 |
+
def reset_models(self, force=False):
|
139 |
+
"""Reset models to free memory if they haven't been used recently"""
|
140 |
+
current_time = time.time()
|
141 |
+
# Only reset if forced or models haven't been used for 10 minutes
|
142 |
+
if force or (current_time - self.last_used > 600):
|
143 |
+
try:
|
144 |
+
logger.info("Resetting models to free memory...")
|
145 |
|
146 |
+
if hasattr(self, 'model') and self.model is not None:
|
147 |
+
del self.model
|
148 |
+
|
149 |
+
if hasattr(self, 'tokenizer') and self.tokenizer is not None:
|
150 |
+
del self.tokenizer
|
151 |
+
|
152 |
+
if hasattr(self, 'whisper_model') and self.whisper_model is not None:
|
153 |
+
del self.whisper_model
|
154 |
|
155 |
+
self.tokenizer = None
|
156 |
+
self.model = None
|
157 |
+
self.whisper_model = None
|
158 |
+
|
159 |
+
if torch.cuda.is_available():
|
160 |
+
torch.cuda.empty_cache()
|
161 |
+
torch.cuda.synchronize()
|
162 |
+
|
163 |
+
gc.collect()
|
164 |
+
logger.info("Models reset successfully")
|
165 |
+
|
166 |
+
except Exception as e:
|
167 |
+
logger.error(f"Error resetting models: {str(e)}")
|
168 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
169 |
# Create global model manager instance
|
170 |
model_manager = ModelManager()
|
171 |
|
|
|
196 |
try:
|
197 |
video = VideoFileClip(video_file)
|
198 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as temp_file:
|
199 |
+
video.audio.write_audiofile(temp_file.name, verbose=False, logger=None)
|
200 |
logger.info(f"Video converted to audio: {temp_file.name}")
|
201 |
return temp_file.name
|
202 |
except Exception as e:
|
|
|
216 |
logger.error(f"Error preprocessing audio: {str(e)}")
|
217 |
raise
|
218 |
|
219 |
+
@spaces.GPU()
|
220 |
def transcribe_audio(file):
|
221 |
"""Transcribe an audio or video file."""
|
222 |
try:
|
223 |
+
model_manager.check_whisper_initialized()
|
224 |
|
225 |
if isinstance(file, str) and file.startswith('http'):
|
226 |
file_path = download_social_media_video(file)
|
227 |
elif isinstance(file, str) and file.lower().endswith(('.mp4', '.avi', '.mov', '.mkv')):
|
228 |
file_path = convert_video_to_audio(file)
|
229 |
+
elif file is not None: # Handle file object from Gradio
|
230 |
+
file_path = preprocess_audio(file.name)
|
231 |
else:
|
232 |
+
return "" # Return empty string for None input
|
233 |
|
234 |
logger.info(f"Transcribing audio: {file_path}")
|
235 |
if not os.path.exists(file_path):
|
236 |
raise FileNotFoundError(f"Audio file not found: {file_path}")
|
237 |
|
238 |
with torch.inference_mode():
|
239 |
+
result = model_manager.whisper_model.transcribe(file_path)
|
240 |
if not result:
|
241 |
raise RuntimeError("Transcription failed to produce results")
|
242 |
|
243 |
transcription = result.get("text", "Error in transcription")
|
244 |
logger.info(f"Transcription completed: {transcription[:50]}...")
|
245 |
+
|
246 |
+
# Clean up temp file
|
247 |
+
try:
|
248 |
+
if os.path.exists(file_path):
|
249 |
+
os.remove(file_path)
|
250 |
+
except Exception as e:
|
251 |
+
logger.warning(f"Could not remove temp file {file_path}: {str(e)}")
|
252 |
+
|
253 |
return transcription
|
254 |
except Exception as e:
|
255 |
logger.error(f"Error transcribing: {str(e)}")
|
|
|
265 |
elif document_path.endswith(".docx"):
|
266 |
doc = docx.Document(document_path)
|
267 |
return "\n".join([paragraph.text for paragraph in doc.paragraphs])
|
268 |
+
elif document_path.endswith((".xlsx", ".xls")):
|
269 |
return pd.read_excel(document_path).to_string()
|
270 |
elif document_path.endswith(".csv"):
|
271 |
return pd.read_csv(document_path).to_string()
|
|
|
278 |
@lru_cache(maxsize=32)
|
279 |
def read_url(url):
|
280 |
"""Read the content of a URL."""
|
281 |
+
if not url or url.strip() == "":
|
282 |
+
return ""
|
283 |
+
|
284 |
try:
|
285 |
+
headers = {
|
286 |
+
'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36'
|
287 |
+
}
|
288 |
+
response = requests.get(url, headers=headers, timeout=15)
|
289 |
response.raise_for_status()
|
290 |
soup = BeautifulSoup(response.content, 'html.parser')
|
291 |
+
|
292 |
+
# Remove non-content elements
|
293 |
+
for element in soup(["script", "style", "meta", "noscript", "iframe", "header", "footer", "nav"]):
|
294 |
+
element.extract()
|
295 |
+
|
296 |
+
# Extract main content
|
297 |
+
main_content = soup.find("main") or soup.find("article") or soup.find("div", class_=["content", "main", "article"])
|
298 |
+
if main_content:
|
299 |
+
text = main_content.get_text(separator='\n', strip=True)
|
300 |
+
else:
|
301 |
+
text = soup.get_text(separator='\n', strip=True)
|
302 |
+
|
303 |
+
# Clean up whitespace
|
304 |
+
lines = [line.strip() for line in text.split('\n') if line.strip()]
|
305 |
+
text = '\n'.join(lines)
|
306 |
+
|
307 |
+
return text[:10000] # Limit to 10k chars to avoid huge inputs
|
308 |
except Exception as e:
|
309 |
logger.error(f"Error reading URL: {str(e)}")
|
310 |
return f"Error reading URL: {str(e)}"
|
311 |
|
312 |
def process_social_content(url):
|
313 |
"""Process social media content."""
|
314 |
+
if not url or url.strip() == "":
|
315 |
+
return None
|
316 |
+
|
317 |
try:
|
318 |
text_content = read_url(url)
|
319 |
try:
|
|
|
330 |
logger.error(f"Error processing social content: {str(e)}")
|
331 |
return None
|
332 |
|
333 |
+
@spaces.GPU()
|
334 |
def generate_news(instructions, facts, size, tone, *args):
|
335 |
+
"""Generate a news article based on provided data"""
|
336 |
try:
|
337 |
+
# Ensure size is integer
|
338 |
+
if isinstance(size, float):
|
339 |
+
size = int(size)
|
340 |
+
elif not isinstance(size, int):
|
341 |
+
size = 250 # Default size
|
342 |
+
|
343 |
+
# Check if models are initialized
|
344 |
+
model_manager.check_llm_initialized()
|
345 |
|
346 |
+
# Prepare data structure for inputs
|
347 |
knowledge_base = {
|
348 |
"instructions": instructions,
|
349 |
"facts": facts,
|
|
|
353 |
"social_content": []
|
354 |
}
|
355 |
|
356 |
+
# Define the indices for parsing args
|
357 |
num_audios = 5 * 3
|
358 |
num_social_urls = 3 * 3
|
359 |
num_urls = 5
|
360 |
|
361 |
+
# Parse arguments
|
362 |
audios = args[:num_audios]
|
363 |
social_urls = args[num_audios:num_audios+num_social_urls]
|
364 |
urls = args[num_audios+num_social_urls:num_audios+num_social_urls+num_urls]
|
365 |
documents = args[num_audios+num_social_urls+num_urls:]
|
366 |
|
367 |
+
# Process URLs with progress reporting
|
368 |
+
logger.info("Processing URLs...")
|
369 |
for url in urls:
|
370 |
+
if url and isinstance(url, str) and url.strip():
|
371 |
content = read_url(url)
|
372 |
if content and not content.startswith("Error"):
|
373 |
knowledge_base["url_content"].append(content)
|
374 |
|
375 |
+
# Process documents
|
376 |
+
logger.info("Processing documents...")
|
377 |
for document in documents:
|
378 |
if document is not None:
|
379 |
content = read_document(document.name)
|
380 |
if content and not content.startswith("Error"):
|
381 |
knowledge_base["document_content"].append(content)
|
382 |
|
383 |
+
# Process audio/video files
|
384 |
+
logger.info("Processing audio/video files...")
|
385 |
for i in range(0, len(audios), 3):
|
386 |
+
if i+2 < len(audios): # Ensure we have complete set of 3 elements
|
387 |
+
audio_file, name, position = audios[i:i+3]
|
388 |
+
if audio_file is not None:
|
389 |
+
knowledge_base["audio_data"].append({
|
390 |
+
"audio": audio_file,
|
391 |
+
"name": name or "Unknown",
|
392 |
+
"position": position or "Not specified"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
393 |
})
|
394 |
|
395 |
+
# Process social media content
|
396 |
+
logger.info("Processing social media content...")
|
397 |
+
for i in range(0, len(social_urls), 3):
|
398 |
+
if i+2 < len(social_urls): # Ensure we have complete set of 3 elements
|
399 |
+
social_url, social_name, social_context = social_urls[i:i+3]
|
400 |
+
if social_url and isinstance(social_url, str) and social_url.strip():
|
401 |
+
social_content = process_social_content(social_url)
|
402 |
+
if social_content:
|
403 |
+
knowledge_base["social_content"].append({
|
404 |
+
"url": social_url,
|
405 |
+
"name": social_name or "Unknown",
|
406 |
+
"context": social_context or "Not specified",
|
407 |
+
"text": social_content.get("text", ""),
|
408 |
+
"video": social_content.get("video", "")
|
409 |
+
})
|
410 |
+
|
411 |
+
# Prepare transcriptions text
|
412 |
transcriptions_text = ""
|
413 |
raw_transcriptions = ""
|
414 |
|
415 |
+
# Process audio data transcriptions
|
416 |
+
logger.info("Transcribing audio...")
|
417 |
for idx, data in enumerate(knowledge_base["audio_data"]):
|
418 |
if data["audio"] is not None:
|
419 |
transcription = transcribe_audio(data["audio"])
|
420 |
+
if transcription and not transcription.startswith("Error"):
|
421 |
+
transcriptions_text += f'"{transcription}" - {data["name"]}, {data["position"]}\n\n'
|
422 |
raw_transcriptions += f'[Audio/Video {idx + 1}]: "{transcription}" - {data["name"]}, {data["position"]}\n\n'
|
423 |
|
424 |
+
# Process social media content transcriptions
|
425 |
+
for idx, data in enumerate(knowledge_base["social_content"]):
|
426 |
if data["text"] and not str(data["text"]).startswith("Error"):
|
427 |
+
# Truncate long texts for the prompt
|
428 |
+
text_excerpt = data["text"][:500] + "..." if len(data["text"]) > 500 else data["text"]
|
429 |
+
social_text = f'[Social media {idx+1} - text]: "{text_excerpt}" - {data["name"]}, {data["context"]}\n\n'
|
430 |
+
transcriptions_text += social_text
|
431 |
+
raw_transcriptions += social_text
|
432 |
+
|
433 |
if data["video"] and not str(data["video"]).startswith("Error"):
|
434 |
+
video_transcription = f'[Social media {idx+1} - video]: "{data["video"]}" - {data["name"]}, {data["context"]}\n\n'
|
435 |
transcriptions_text += video_transcription
|
436 |
+
raw_transcriptions += video_transcription
|
437 |
+
|
438 |
+
# Combine document content and URL content (with truncation for very long content)
|
439 |
+
document_summaries = []
|
440 |
+
for idx, doc in enumerate(knowledge_base["document_content"]):
|
441 |
+
# Truncate long documents
|
442 |
+
if len(doc) > 1000:
|
443 |
+
doc_excerpt = doc[:1000] + "... [document continues]"
|
444 |
+
else:
|
445 |
+
doc_excerpt = doc
|
446 |
+
document_summaries.append(f"[Document {idx+1}]: {doc_excerpt}")
|
447 |
+
|
448 |
+
document_content = "\n\n".join(document_summaries)
|
449 |
+
|
450 |
+
url_summaries = []
|
451 |
+
for idx, url_content in enumerate(knowledge_base["url_content"]):
|
452 |
+
# Truncate long URL content
|
453 |
+
if len(url_content) > 1000:
|
454 |
+
url_excerpt = url_content[:1000] + "... [content continues]"
|
455 |
+
else:
|
456 |
+
url_excerpt = url_content
|
457 |
+
url_summaries.append(f"[URL {idx+1}]: {url_excerpt}")
|
458 |
+
|
459 |
+
url_content = "\n\n".join(url_summaries)
|
460 |
|
461 |
+
# Create prompt for the model
|
462 |
+
prompt = f"""<s>[INST] You are a professional news writer. Write a news article based on the following information:
|
463 |
|
464 |
Instructions: {knowledge_base["instructions"]}
|
465 |
+
|
466 |
Facts: {knowledge_base["facts"]}
|
467 |
+
|
468 |
+
Additional content from documents:
|
469 |
+
{document_content}
|
470 |
+
|
471 |
+
Additional content from URLs:
|
472 |
+
{url_content}
|
473 |
|
474 |
Use these transcriptions as direct and indirect quotes:
|
475 |
{transcriptions_text}
|
|
|
477 |
Follow these requirements:
|
478 |
- Write a title
|
479 |
- Write a 15-word hook that complements the title
|
480 |
+
- Write the body with approximately {size} words
|
481 |
- Use a {tone} tone
|
482 |
- Answer the 5 Ws (Who, What, When, Where, Why) in the first paragraph
|
483 |
- Use at least 80% direct quotes (in quotation marks)
|
|
|
485 |
- Do not invent information
|
486 |
- Be rigorous with the provided facts [/INST]"""
|
487 |
|
488 |
+
# Optimize for requested size
|
489 |
+
max_new_tokens = min(int(size * 2.5), 1024) # Increased limit for better quality
|
490 |
|
491 |
+
# Generate response using optimized unsloth model
|
492 |
with torch.inference_mode():
|
493 |
try:
|
494 |
+
logger.info("Generating news article...")
|
495 |
+
# Use unsloth's optimized generate method
|
496 |
+
inputs = model_manager.tokenizer(
|
497 |
+
prompt,
|
498 |
+
return_tensors="pt",
|
499 |
+
add_special_tokens=False
|
500 |
+
).to(model_manager.model.device)
|
501 |
+
|
502 |
+
# Generate with optimized settings
|
503 |
+
outputs = model_manager.model.generate(
|
504 |
+
**inputs,
|
505 |
+
max_new_tokens=max_new_tokens,
|
506 |
do_sample=True,
|
507 |
temperature=0.7,
|
508 |
top_p=0.95,
|
509 |
repetition_penalty=1.2,
|
510 |
+
pad_token_id=model_manager.tokenizer.eos_token_id,
|
511 |
+
use_cache=True
|
512 |
+
)
|
513 |
+
|
514 |
+
# Decode the generated text
|
515 |
+
generated_text = model_manager.tokenizer.decode(
|
516 |
+
outputs[0][inputs.input_ids.shape[1]:],
|
517 |
+
skip_special_tokens=True
|
518 |
)
|
519 |
|
520 |
+
# Clean up the generated text
|
521 |
+
news_article = generated_text.strip()
|
522 |
+
logger.info(f"News generation completed: {len(news_article)} chars")
|
|
|
523 |
|
524 |
except Exception as gen_error:
|
525 |
logger.error(f"Error in text generation: {str(gen_error)}")
|
|
|
530 |
except Exception as e:
|
531 |
logger.error(f"Error generating news: {str(e)}")
|
532 |
try:
|
533 |
+
# Reset models to recover from errors
|
534 |
+
model_manager.reset_models(force=True)
|
535 |
+
except Exception as reset_error:
|
536 |
+
logger.error(f"Failed to reset models: {str(reset_error)}")
|
537 |
+
return f"Error generando la noticia: {str(e)}", "Error procesando las transcripciones."
|
|
|
|
|
538 |
|
539 |
def create_demo():
|
540 |
+
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
541 |
+
gr.Markdown("# 📰 NewsIA - Generador de Noticias IA")
|
542 |
+
gr.Markdown("Crea noticias profesionales a partir de múltiples fuentes de información.")
|
543 |
|
544 |
with gr.Row():
|
545 |
with gr.Column(scale=2):
|
546 |
instrucciones = gr.Textbox(
|
547 |
label="Instrucciones para la noticia",
|
548 |
+
placeholder="Escribe instrucciones específicas para la generación de tu noticia",
|
549 |
lines=2
|
550 |
)
|
551 |
hechos = gr.Textbox(
|
552 |
+
label="Hechos principales",
|
553 |
+
placeholder="Describe los hechos más importantes que debe incluir la noticia",
|
554 |
lines=4
|
555 |
)
|
556 |
+
|
557 |
+
with gr.Row():
|
558 |
+
tamaño = gr.Slider(
|
559 |
+
label="Longitud aproximada (palabras)",
|
560 |
+
minimum=100,
|
561 |
+
maximum=500,
|
562 |
+
value=250,
|
563 |
+
step=50
|
564 |
+
)
|
565 |
+
tono = gr.Dropdown(
|
566 |
+
label="Tono de la noticia",
|
567 |
+
choices=["serio", "neutral", "divertido", "formal", "informal", "urgente"],
|
568 |
+
value="neutral"
|
569 |
+
)
|
570 |
|
571 |
with gr.Column(scale=3):
|
572 |
inputs_list = [instrucciones, hechos, tamaño, tono]
|
573 |
|
574 |
with gr.Tabs():
|
575 |
+
with gr.TabItem("📝 Documentos"):
|
576 |
+
for i in range(1, 4): # Reduced to 3 for better UX
|
577 |
+
with gr.Row():
|
578 |
+
documento = gr.File(
|
579 |
+
label=f"Documento {i}",
|
580 |
+
file_types=["pdf", "docx", "xlsx", "csv"],
|
581 |
+
file_count="single"
|
582 |
+
)
|
583 |
+
inputs_list.append(documento)
|
584 |
+
|
585 |
+
# Add empty inputs to match the original expected array length
|
586 |
+
for i in range(4, 6):
|
587 |
+
inputs_list.append(None)
|
588 |
+
|
589 |
+
with gr.TabItem("🔊 Audio/Video"):
|
590 |
+
for i in range(1, 4): # Reduced to 3 for better UX
|
591 |
+
with gr.Group():
|
592 |
+
gr.Markdown(f"**Fuente {i}**")
|
593 |
+
file = gr.File(
|
594 |
+
label=f"Audio/Video {i}",
|
595 |
+
file_types=["audio", "video"]
|
596 |
+
)
|
597 |
+
with gr.Row():
|
598 |
+
nombre = gr.Textbox(
|
599 |
+
label="Nombre",
|
600 |
+
placeholder="Nombre del entrevistado"
|
601 |
+
)
|
602 |
+
cargo = gr.Textbox(
|
603 |
+
label="Cargo/Rol",
|
604 |
+
placeholder="Cargo o rol"
|
605 |
+
)
|
606 |
+
inputs_list.extend([file, nombre, cargo])
|
607 |
+
|
608 |
+
# Add empty inputs to match the original expected array length
|
609 |
+
for i in range(4, 6):
|
610 |
+
inputs_list.extend([None, None, None])
|
611 |
+
|
612 |
+
with gr.TabItem("🌐 URLs"):
|
613 |
+
for i in range(1, 4): # Reduced to 3 for better UX
|
614 |
url = gr.Textbox(
|
615 |
label=f"URL {i}",
|
616 |
placeholder="https://..."
|
617 |
)
|
618 |
inputs_list.append(url)
|
619 |
+
|
620 |
+
# Add empty inputs to match the original expected array length
|
621 |
+
for i in range(4, 6):
|
622 |
+
inputs_list.append(None)
|
623 |
+
|
624 |
+
with gr.TabItem("📱 Redes Sociales"):
|
625 |
+
for i in range(1, 3): # Reduced to 2 for better UX
|
626 |
+
with gr.Group():
|
627 |
+
gr.Markdown(f"**Red Social {i}**")
|
628 |
+
social_url = gr.Textbox(
|
629 |
+
label=f"URL",
|
630 |
+
placeholder="https://..."
|
631 |
+
)
|
632 |
+
with gr.Row():
|
633 |
+
social_nombre = gr.Textbox(
|
634 |
+
label=f"Nombre/Cuenta",
|
635 |
+
placeholder="Nombre de la persona o cuenta"
|
636 |
+
)
|
637 |
+
social_contexto = gr.Textbox(
|
638 |
+
label=f"Contexto",
|
639 |
+
placeholder="Contexto relevante"
|
640 |
+
)
|
641 |
+
inputs_list.extend([social_url, social_nombre, social_contexto])
|
642 |
+
|
643 |
+
# Add empty inputs to match the original expected array length
|
644 |
+
for i in range(3, 4):
|
645 |
+
inputs_list.extend([None, None, None])
|
646 |
|
647 |
with gr.Row():
|
648 |
+
generar = gr.Button("✨ Generar Noticia", variant="primary")
|
649 |
+
reset = gr.Button("🔄 Limpiar Todo")
|
650 |
+
|
651 |
+
with gr.Tabs():
|
652 |
+
with gr.TabItem("📄 Noticia Generada"):
|
653 |
+
noticia_output = gr.Textbox(
|
654 |
+
label="Borrador de la noticia",
|
655 |
+
lines=15,
|
656 |
+
show_copy_button=True
|
657 |
+
)
|
658 |
+
|
659 |
+
with gr.TabItem("🎙️ Transcripciones"):
|
660 |
+
transcripciones_output = gr.Textbox(
|
661 |
+
label="Transcripciones de fuentes",
|
662 |
+
lines=10,
|
663 |
+
show_copy_button=True
|
664 |
+
)
|
665 |
|
666 |
+
# Set up event handlers
|
667 |
generar.click(
|
668 |
fn=generate_news,
|
669 |
inputs=inputs_list,
|
670 |
outputs=[noticia_output, transcripciones_output]
|
671 |
)
|
672 |
+
|
673 |
+
# Reset functionality to clear all inputs
|
674 |
+
def reset_all():
|
675 |
+
output = []
|
676 |
+
for _ in range(len(inputs_list)):
|
677 |
+
output.append(None)
|
678 |
+
output.append("")
|
679 |
+
output.append("")
|
680 |
+
return output
|
681 |
+
|
682 |
+
reset.click(
|
683 |
+
fn=reset_all,
|
684 |
+
inputs=[],
|
685 |
+
outputs=inputs_list + [noticia_output, transcripciones_output]
|
686 |
+
)
|
687 |
|
688 |
return demo
|
689 |
|
690 |
if __name__ == "__main__":
|
691 |
+
# Initialize models on startup to reduce first request latency
|
692 |
+
try:
|
693 |
+
model_manager.initialize_whisper()
|
694 |
+
model_manager.initialize_llm()
|
695 |
+
except Exception as e:
|
696 |
+
logger.warning(f"Initial model loading failed: {str(e)}")
|
697 |
+
|
698 |
demo = create_demo()
|
699 |
+
demo.queue(concurrency_count=1, max_size=5)
|
700 |
demo.launch(
|
701 |
share=True,
|
702 |
server_name="0.0.0.0",
|