Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -94,31 +94,45 @@
|
|
94 |
|
95 |
# return Response("No audio generated", status_code=400)
|
96 |
|
97 |
-
from fastapi import FastAPI, Response, HTTPException, Request
|
98 |
-
from fastapi.responses import JSONResponse
|
99 |
-
from fastapi.staticfiles import StaticFiles
|
100 |
-
from kokoro import KPipeline
|
101 |
import os
|
102 |
-
import
|
103 |
-
import torch
|
104 |
-
from huggingface_hub import InferenceClient
|
105 |
-
from pydantic import BaseModel
|
106 |
import base64
|
107 |
import logging
|
|
|
|
|
|
|
|
|
108 |
from typing import Optional, ClassVar, List
|
109 |
-
import
|
|
|
|
|
|
|
110 |
|
111 |
# Set up logging
|
112 |
logging.basicConfig(level=logging.INFO)
|
113 |
logger = logging.getLogger(__name__)
|
114 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
class TextImageRequest(BaseModel):
|
116 |
text: Optional[str] = None
|
117 |
image_base64: Optional[str] = None
|
118 |
-
voice: str = "af_heart" # Default voice
|
119 |
speed: float = 1.0
|
120 |
|
121 |
-
#
|
122 |
AVAILABLE_VOICES: ClassVar[List[str]] = ["af_heart"]
|
123 |
|
124 |
def validate_voice(self):
|
@@ -126,6 +140,7 @@ class TextImageRequest(BaseModel):
|
|
126 |
return "af_heart"
|
127 |
return self.voice
|
128 |
|
|
|
129 |
class AudioResponse(BaseModel):
|
130 |
status: str
|
131 |
message: str
|
@@ -134,107 +149,72 @@ class ErrorResponse(BaseModel):
|
|
134 |
error: str
|
135 |
detail: Optional[str] = None
|
136 |
|
137 |
-
#
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
)
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
|
|
|
|
152 |
try:
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
with open(image_path, "wb") as f:
|
176 |
-
f.write(image_data)
|
177 |
-
# Construct image URL (assumes BASE_URL environment variable or defaults to localhost)
|
178 |
-
base_url = os.getenv("BASE_URL", "http://localhost:8000")
|
179 |
-
image_url = f"{base_url}/static/{filename}"
|
180 |
-
prompt = text if text else "Describe this image in one sentence."
|
181 |
-
# Construct message exactly as in the reference
|
182 |
-
messages = [
|
183 |
-
{
|
184 |
-
"role": "user",
|
185 |
-
"content": [
|
186 |
-
{"type": "text", "text": prompt},
|
187 |
-
{"type": "image_url", "image_url": {"url": image_url}}
|
188 |
-
]
|
189 |
-
}
|
190 |
-
]
|
191 |
-
else:
|
192 |
-
logger.info("Processing text-only request")
|
193 |
-
messages = [
|
194 |
-
{
|
195 |
-
"role": "user",
|
196 |
-
"content": text + " Describe in one line only."
|
197 |
-
}
|
198 |
]
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
|
|
203 |
completion = client.chat.completions.create(
|
204 |
model="meta-llama/Llama-3.2-11B-Vision-Instruct",
|
205 |
messages=messages,
|
206 |
max_tokens=500
|
207 |
)
|
208 |
-
|
209 |
-
logger.info("
|
210 |
-
|
211 |
-
|
212 |
-
try:
|
213 |
-
response = completion.choices[0].message.content
|
214 |
-
logger.info(f"Extracted response content: {response}")
|
215 |
-
return response
|
216 |
-
except Exception as e:
|
217 |
-
logger.error(f"Error extracting message content: {str(e)}")
|
218 |
-
try:
|
219 |
-
if hasattr(completion.choices[0], "message") and hasattr(completion.choices[0].message, "content"):
|
220 |
-
return completion.choices[0].message.content
|
221 |
-
return completion.choices[0]["message"]["content"]
|
222 |
-
except Exception as e2:
|
223 |
-
logger.error(f"All extraction methods failed: {str(e2)}")
|
224 |
-
return "I couldn't process that input. Please try again with a different query."
|
225 |
-
|
226 |
except Exception as e:
|
227 |
-
logger.error(f"Error
|
228 |
raise HTTPException(status_code=500, detail=str(e))
|
229 |
|
230 |
-
# Initialize
|
231 |
try:
|
232 |
logger.info("Initializing KPipeline...")
|
233 |
pipeline = KPipeline(lang_code='a')
|
234 |
logger.info("KPipeline initialized successfully")
|
235 |
except Exception as e:
|
236 |
logger.error(f"Failed to initialize KPipeline: {str(e)}")
|
237 |
-
# The
|
238 |
|
239 |
@app.post("/generate", responses={
|
240 |
200: {"content": {"application/octet-stream": {}}},
|
@@ -243,95 +223,69 @@ except Exception as e:
|
|
243 |
})
|
244 |
async def generate_audio(request: TextImageRequest):
|
245 |
"""
|
246 |
-
Generate audio from
|
247 |
-
|
248 |
-
- If text is provided, it is used as input.
|
249 |
-
- If an image is provided (base64), it is saved and a URL is generated for processing.
|
250 |
-
- The LLM response is then converted to speech.
|
251 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
try:
|
253 |
-
|
254 |
-
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
validated_voice = request.validate_voice()
|
269 |
-
if validated_voice != request.voice:
|
270 |
-
logger.warning(f"Requested voice '{request.voice}' not available, using '{validated_voice}' instead")
|
271 |
-
|
272 |
-
logger.info(f"Generating audio using voice={validated_voice}, speed={request.speed}")
|
273 |
-
try:
|
274 |
-
generator = pipeline(
|
275 |
-
text_reply,
|
276 |
-
voice=validated_voice,
|
277 |
-
speed=request.speed,
|
278 |
-
split_pattern=r'\n+'
|
279 |
-
)
|
280 |
-
|
281 |
-
for i, (gs, ps, audio) in enumerate(generator):
|
282 |
-
logger.info(f"Audio generated successfully: segment {i}")
|
283 |
-
# Convert PyTorch tensor to NumPy array
|
284 |
-
audio_numpy = audio.cpu().numpy()
|
285 |
-
# Clip values to range [-1, 1] and convert to 16-bit PCM
|
286 |
-
audio_numpy = np.clip(audio_numpy, -1, 1)
|
287 |
-
pcm_data = (audio_numpy * 32767).astype(np.int16)
|
288 |
-
raw_audio = pcm_data.tobytes()
|
289 |
-
|
290 |
-
return Response(
|
291 |
-
content=raw_audio,
|
292 |
-
media_type="application/octet-stream",
|
293 |
-
headers={
|
294 |
-
"Content-Disposition": 'attachment; filename="output.pcm"',
|
295 |
-
"X-Sample-Rate": "24000",
|
296 |
-
"X-Bits-Per-Sample": "16",
|
297 |
-
"X-Endianness": "little"
|
298 |
-
}
|
299 |
-
)
|
300 |
-
|
301 |
-
logger.error("No audio segments generated")
|
302 |
-
return JSONResponse(
|
303 |
-
status_code=400,
|
304 |
-
content={"error": "No audio generated", "detail": "The pipeline did not produce any audio"}
|
305 |
-
)
|
306 |
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
|
|
|
|
|
|
|
|
312 |
)
|
313 |
-
|
314 |
except Exception as e:
|
315 |
-
logger.error(f"
|
316 |
-
|
317 |
-
status_code=500,
|
318 |
-
content={"error": "Internal server error", "detail": str(e)}
|
319 |
-
)
|
320 |
|
321 |
@app.get("/")
|
322 |
async def root():
|
323 |
-
return {"message": "Welcome
|
324 |
|
325 |
@app.exception_handler(404)
|
326 |
async def not_found_handler(request: Request, exc):
|
327 |
-
return JSONResponse(
|
328 |
-
status_code=404,
|
329 |
-
content={"error": "Endpoint not found. Please use POST /generate for queries."}
|
330 |
-
)
|
331 |
|
332 |
@app.exception_handler(405)
|
333 |
async def method_not_allowed_handler(request: Request, exc):
|
334 |
-
return JSONResponse(
|
335 |
-
|
336 |
-
content={"error": "Method not allowed. Please check the API documentation."}
|
337 |
-
)
|
|
|
94 |
|
95 |
# return Response("No audio generated", status_code=400)
|
96 |
|
|
|
|
|
|
|
|
|
97 |
import os
|
98 |
+
import uuid
|
|
|
|
|
|
|
99 |
import base64
|
100 |
import logging
|
101 |
+
from fastapi import FastAPI, HTTPException, Response, Request
|
102 |
+
from fastapi.responses import JSONResponse
|
103 |
+
from fastapi.staticfiles import StaticFiles
|
104 |
+
from pydantic import BaseModel
|
105 |
from typing import Optional, ClassVar, List
|
106 |
+
from huggingface_hub import InferenceClient
|
107 |
+
import numpy as np
|
108 |
+
import torch
|
109 |
+
from kokoro import KPipeline # Assuming you have this pipeline for audio generation
|
110 |
|
111 |
# Set up logging
|
112 |
logging.basicConfig(level=logging.INFO)
|
113 |
logger = logging.getLogger(__name__)
|
114 |
|
115 |
+
# Create FastAPI app
|
116 |
+
app = FastAPI(
|
117 |
+
title="Text-to-Speech API with Vision Support",
|
118 |
+
description="This API uses meta-llama/Llama-3.2-11B-Vision-Instruct, which requires an image input.",
|
119 |
+
version="1.0.0"
|
120 |
+
)
|
121 |
+
|
122 |
+
# Mount a static directory for serving saved images
|
123 |
+
STATIC_DIR = "static_images"
|
124 |
+
if not os.path.exists(STATIC_DIR):
|
125 |
+
os.makedirs(STATIC_DIR)
|
126 |
+
app.mount("/static", StaticFiles(directory=STATIC_DIR), name="static")
|
127 |
+
|
128 |
+
# Pydantic model for request
|
129 |
class TextImageRequest(BaseModel):
|
130 |
text: Optional[str] = None
|
131 |
image_base64: Optional[str] = None
|
132 |
+
voice: str = "af_heart" # Default voice
|
133 |
speed: float = 1.0
|
134 |
|
135 |
+
# Use ClassVar so that Pydantic doesn't treat this as a model field.
|
136 |
AVAILABLE_VOICES: ClassVar[List[str]] = ["af_heart"]
|
137 |
|
138 |
def validate_voice(self):
|
|
|
140 |
return "af_heart"
|
141 |
return self.voice
|
142 |
|
143 |
+
# (Optional) Pydantic models for responses
|
144 |
class AudioResponse(BaseModel):
|
145 |
status: str
|
146 |
message: str
|
|
|
149 |
error: str
|
150 |
detail: Optional[str] = None
|
151 |
|
152 |
+
# Function to call the LLM model following the reference code exactly
|
153 |
+
def llm_chat_response(text: str, image_base64: str) -> str:
|
154 |
+
HF_TOKEN = os.getenv("HF_TOKEN")
|
155 |
+
logger.info("Checking HF_TOKEN...")
|
156 |
+
if not HF_TOKEN:
|
157 |
+
logger.error("HF_TOKEN not configured")
|
158 |
+
raise HTTPException(status_code=500, detail="HF_TOKEN not configured")
|
159 |
+
|
160 |
+
logger.info("Initializing InferenceClient...")
|
161 |
+
client = InferenceClient(
|
162 |
+
provider="hf-inference",
|
163 |
+
api_key=HF_TOKEN
|
164 |
+
)
|
165 |
+
|
166 |
+
# Save the base64-encoded image locally so it is accessible via a URL
|
167 |
+
filename = f"{uuid.uuid4()}.jpg"
|
168 |
+
image_path = os.path.join(STATIC_DIR, filename)
|
169 |
try:
|
170 |
+
image_data = base64.b64decode(image_base64)
|
171 |
+
except Exception as e:
|
172 |
+
logger.error(f"Error decoding image: {str(e)}")
|
173 |
+
raise HTTPException(status_code=400, detail="Invalid base64 image data")
|
174 |
+
|
175 |
+
with open(image_path, "wb") as f:
|
176 |
+
f.write(image_data)
|
177 |
+
|
178 |
+
# Construct the public URL for the saved image.
|
179 |
+
# BASE_URL should be set to your public URL if not running locally.
|
180 |
+
base_url = os.getenv("BASE_URL", "http://localhost:8000")
|
181 |
+
image_url = f"{base_url}/static/{filename}"
|
182 |
+
|
183 |
+
# Build the message exactly as in the reference code.
|
184 |
+
# This model requires a list with two items: one for text and one for the image.
|
185 |
+
prompt = text if text else "Describe this image in one sentence."
|
186 |
+
messages = [
|
187 |
+
{
|
188 |
+
"role": "user",
|
189 |
+
"content": [
|
190 |
+
{"type": "text", "text": prompt},
|
191 |
+
{"type": "image_url", "image_url": {"url": image_url}}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
]
|
193 |
+
}
|
194 |
+
]
|
195 |
+
logger.info(f"Message structure: {messages}")
|
196 |
+
|
197 |
+
try:
|
198 |
completion = client.chat.completions.create(
|
199 |
model="meta-llama/Llama-3.2-11B-Vision-Instruct",
|
200 |
messages=messages,
|
201 |
max_tokens=500
|
202 |
)
|
203 |
+
response = completion.choices[0].message.content
|
204 |
+
logger.info(f"Extracted response: {response}")
|
205 |
+
return response
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
206 |
except Exception as e:
|
207 |
+
logger.error(f"Error during model inference: {str(e)}")
|
208 |
raise HTTPException(status_code=500, detail=str(e))
|
209 |
|
210 |
+
# Initialize audio generation pipeline (your audio conversion pipeline)
|
211 |
try:
|
212 |
logger.info("Initializing KPipeline...")
|
213 |
pipeline = KPipeline(lang_code='a')
|
214 |
logger.info("KPipeline initialized successfully")
|
215 |
except Exception as e:
|
216 |
logger.error(f"Failed to initialize KPipeline: {str(e)}")
|
217 |
+
# The API can still run, but audio generation will fail.
|
218 |
|
219 |
@app.post("/generate", responses={
|
220 |
200: {"content": {"application/octet-stream": {}}},
|
|
|
223 |
})
|
224 |
async def generate_audio(request: TextImageRequest):
|
225 |
"""
|
226 |
+
Generate audio from a multimodal (text+image) input.
|
227 |
+
This model does not support text-only inputs.
|
|
|
|
|
|
|
228 |
"""
|
229 |
+
logger.info("Received generation request")
|
230 |
+
# Ensure an image is provided because the model is multimodal.
|
231 |
+
if not request.image_base64:
|
232 |
+
raise HTTPException(status_code=400, detail="This model requires an image input.")
|
233 |
+
|
234 |
+
# Get the text prompt. If none is provided, use a default.
|
235 |
+
user_text = request.text if request.text else "Describe this image in one sentence."
|
236 |
+
|
237 |
+
# Get the LLM's response
|
238 |
+
logger.info("Calling the LLM model")
|
239 |
+
text_reply = llm_chat_response(user_text, request.image_base64)
|
240 |
+
logger.info(f"LLM response: {text_reply}")
|
241 |
+
|
242 |
+
# Validate voice parameter (if needed for audio generation)
|
243 |
+
validated_voice = request.validate_voice()
|
244 |
+
if validated_voice != request.voice:
|
245 |
+
logger.warning(f"Voice '{request.voice}' not available; using '{validated_voice}' instead")
|
246 |
+
|
247 |
+
# Convert the text reply to audio using your audio pipeline
|
248 |
+
logger.info(f"Generating audio using voice={validated_voice}, speed={request.speed}")
|
249 |
try:
|
250 |
+
# Generate audio segments (assumes pipeline yields segments)
|
251 |
+
generator = pipeline(
|
252 |
+
text_reply,
|
253 |
+
voice=validated_voice,
|
254 |
+
speed=request.speed,
|
255 |
+
split_pattern=r'\n+'
|
256 |
+
)
|
257 |
+
for i, (gs, ps, audio) in enumerate(generator):
|
258 |
+
logger.info(f"Audio generated, segment {i}")
|
259 |
+
# Convert audio tensor to 16-bit PCM bytes
|
260 |
+
audio_numpy = audio.cpu().numpy()
|
261 |
+
audio_numpy = np.clip(audio_numpy, -1, 1)
|
262 |
+
pcm_data = (audio_numpy * 32767).astype(np.int16)
|
263 |
+
raw_audio = pcm_data.tobytes()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
|
265 |
+
return Response(
|
266 |
+
content=raw_audio,
|
267 |
+
media_type="application/octet-stream",
|
268 |
+
headers={
|
269 |
+
"Content-Disposition": 'attachment; filename="output.pcm"',
|
270 |
+
"X-Sample-Rate": "24000",
|
271 |
+
"X-Bits-Per-Sample": "16",
|
272 |
+
"X-Endianness": "little"
|
273 |
+
}
|
274 |
)
|
275 |
+
raise HTTPException(status_code=400, detail="No audio segments generated.")
|
276 |
except Exception as e:
|
277 |
+
logger.error(f"Error generating audio: {str(e)}")
|
278 |
+
raise HTTPException(status_code=500, detail=str(e))
|
|
|
|
|
|
|
279 |
|
280 |
@app.get("/")
|
281 |
async def root():
|
282 |
+
return {"message": "Welcome! Use POST /generate with text and image_base64."}
|
283 |
|
284 |
@app.exception_handler(404)
|
285 |
async def not_found_handler(request: Request, exc):
|
286 |
+
return JSONResponse(status_code=404, content={"error": "Endpoint not found."})
|
|
|
|
|
|
|
287 |
|
288 |
@app.exception_handler(405)
|
289 |
async def method_not_allowed_handler(request: Request, exc):
|
290 |
+
return JSONResponse(status_code=405, content={"error": "Method not allowed."})
|
291 |
+
|
|
|
|