feast/sdk/python/feast/embedder.py at revert-6838-fix_minio_image · feast-dev/feast · GitHub

Repository navigation

History

359 lines (279 loc) · 11.4 KB

Raw

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

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

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

from abc import ABC, abstractmethod

from dataclasses import dataclass

from typing import (

TYPE_CHECKING,

Any,

Callable,

List,

Optional,

Protocol,

runtime_checkable,

)

import pandas as pd

if TYPE_CHECKING:

import numpy as np

from feast.repo_config import EmbeddingModelConfig

@runtime_checkable

class EmbeddingProvider(Protocol):

"""Protocol for query-time embedding providers.

Implement this to plug in any embedding backend (OpenAI, Cohere,

SentenceTransformers, a local model, etc.) for use in

``openai_search`` and the OpenAI-compatible

feature server endpoints.

Example::

class MyEmbeddingProvider:

def embed(self, texts: List[str]) -> List[List[float]]:

return my_model.encode(texts)

async def aembed(self, texts: List[str]) -> List[List[float]]:

return await my_model.aencode(texts)

"""

def embed(self, texts: List[str]) -> List[List[float]]:

"""Return one embedding vector per input text."""

...

async def aembed(self, texts: List[str]) -> List[List[float]]:

"""Async variant of :meth:`embed`."""

...

class SentenceTransformersEmbeddingProvider:

"""EmbeddingProvider backed by Sentence Transformers (local inference).

Runs entirely on-device — no external API key required. Ideal for

air-gapped deployments, cost-sensitive workloads, or environments where

outbound API calls are restricted.

Requires the ``sentence-transformers`` package::

pip install sentence-transformers

Configuration in ``feature_store.yaml``::

embedding_model:

provider: sentence_transformers

model: all-MiniLM-L6-v2 # any HuggingFace sentence-transformers model

The model is loaded lazily on the first call to :meth:`embed` or

:meth:`aembed` so that import cost is paid only when the provider is

actually used.

"""

def __init__(self, model: str = "all-MiniLM-L6-v2"):

self.model_name = model

self._model = None

@classmethod

def from_config(

cls, config: "EmbeddingModelConfig"

) -> "SentenceTransformersEmbeddingProvider":

"""Build a provider from a :class:`~feast.repo_config.EmbeddingModelConfig`."""

return cls(model=config.model)

def _load_model(self):

if self._model is None:

try:

from sentence_transformers import SentenceTransformer

except ImportError:

raise ImportError(

"sentence-transformers is required for the "

"'sentence_transformers' embedding provider. "

"Install with: pip install sentence-transformers"

)

self._model = SentenceTransformer(self.model_name)

return self._model

def embed(self, texts: List[str]) -> List[List[float]]:

model = self._load_model()

embeddings = model.encode(texts, convert_to_numpy=True)

return [emb.tolist() for emb in embeddings]

async def aembed(self, texts: List[str]) -> List[List[float]]:

import asyncio

loop = asyncio.get_event_loop()

return await loop.run_in_executor(None, self.embed, texts)

def get_embedding_provider(

config: "EmbeddingModelConfig",

) -> "EmbeddingProvider":

"""Factory that returns the appropriate :class:`EmbeddingProvider` for *config*.

Dispatches on ``config.provider``:

* ``"sentence_transformers"`` (default) → :class:`SentenceTransformersEmbeddingProvider`

Raises:

ValueError: If ``config.provider`` is not a recognised value.

"""

provider = (config.provider or "sentence_transformers").lower()

if provider == "sentence_transformers":

return SentenceTransformersEmbeddingProvider.from_config(config)

raise ValueError(

f"Unknown embedding provider: '{config.provider}'. "

"Supported values: 'sentence_transformers'."

)

@dataclass

class EmbeddingConfig:

batch_size: int = 64

show_progress: bool = True

class BaseEmbedder(ABC):

"""

Abstract base class for embedding generation.

Supports multiple modalities via routing.

Users can register custom modality handlers.

"""

def __init__(self, config: Optional[EmbeddingConfig] = None):

self.config = config or EmbeddingConfig()

# Registry: modality -> embedding function

self._modality_handlers: dict[str, Callable[[List[Any]], "np.ndarray"]] = {}

# Register default modalities (subclass can override)

self._register_default_modalities()

def _register_default_modalities(self) -> None:

"""Override in subclass to register default modality handlers."""

pass

def register_modality(

self,

modality: str,

handler: Callable[[List[Any]], "np.ndarray"],

) -> None:

"""

Args:

modality: Name of modality ("text", "image", "video", etc.)

handler: Function that takes list of inputs and returns embeddings.

"""

self._modality_handlers[modality] = handler

@property

def supported_modalities(self) -> List[str]:

"""Return list of supported modalities."""

return list(self._modality_handlers.keys())

def get_embedding_dim(self, modality: str) -> Optional[int]:

"""

Return the embedding dimension for a given modality.

Subclasses should override this to return the actual dimension

so that auto-generated FeatureView schemas use the correct vector_length.

Args:

modality: The modality to query (e.g. "text", "image").

Returns:

The embedding dimension, or None if unknown.

"""

return None

@abstractmethod

def embed(self, inputs: List[Any], modality: str) -> "np.ndarray":

"""

Generate embeddings for inputs of a given modality.

Args:

inputs: List of inputs.

modality: Type of content ("text", "image", "video", etc.)

Returns:

numpy array of shape (len(inputs), embedding_dim)

"""

pass

def embed_dataframe(

self,

df: pd.DataFrame,

column_mapping: dict[str, tuple[str, str]],

) -> pd.DataFrame:

"""

Add embeddings for multiple columns with modality routing.

Args:

df: Input DataFrame.

column_mapping: Dict mapping source_column -> (modality, output_column).

Example: {

"text": ("text", "text_embedding"),

"image_path": ("image", "image_embedding"),

"video_path": ("video", "video_embedding"),

}

"""

df = df.copy()

for source_column, (modality, output_column) in column_mapping.items():

inputs = df[source_column].tolist()

embeddings = self.embed(inputs, modality)

df[output_column] = pd.Series(

[emb.tolist() for emb in embeddings], dtype=object, index=df.index

)

return df

class MultiModalEmbedder(BaseEmbedder):

"""

Multi-modal embedder with built-in support for common modalities.

Supports: text, image, video (extensible)

"""

def __init__(

self,

text_model: str = "all-MiniLM-L6-v2",

image_model: str = "openai/clip-vit-base-patch32",

config: Optional[EmbeddingConfig] = None,

self.text_model_name = text_model

self.image_model_name = image_model

# Lazy-loaded models

self._text_model = None

self._image_model = None

self._image_processor = None

super().__init__(config)

def _register_default_modalities(self) -> None:

"""Register built-in modality handlers."""

self.register_modality("text", self._embed_text)

self.register_modality("image", self._embed_image)

# Future: add more as needed

# self.register_modality("video", self._embed_video)

# self.register_modality("audio", self._embed_audio)

def embed(self, inputs: List[Any], modality: str) -> "np.ndarray":

"""Route to appropriate handler based on modality."""

if modality not in self._modality_handlers:

raise ValueError(

f"Unsupported modality: '{modality}'. "

f"Supported: {self.supported_modalities}"

)

handler = self._modality_handlers[modality]

return handler(inputs)

def get_embedding_dim(self, modality: str) -> Optional[int]:

"""

Return the embedding dimension for a given modality.

For "text", this queries the SentenceTransformer model's dimension

(which triggers lazy model loading).

Args:

modality: The modality to query (e.g. "text", "image").

Returns:

The embedding dimension, or None if unknown.

"""

if modality == "text":

return self.text_model.get_sentence_embedding_dimension()

elif modality == "image":

return self.image_model.config.vision_config.hidden_size

return None

# Text Embedding

@property

def text_model(self):

if self._text_model is None:

from sentence_transformers import SentenceTransformer

self._text_model = SentenceTransformer(self.text_model_name)

return self._text_model

def _embed_text(self, inputs: List[str]) -> "np.ndarray":

return self.text_model.encode(

inputs,

batch_size=self.config.batch_size,

show_progress_bar=self.config.show_progress,

)

# Image Embedding

@property

def image_model(self):

if self._image_model is None:

from transformers import CLIPModel

self._image_model = CLIPModel.from_pretrained(self.image_model_name)

return self._image_model

@property

def image_processor(self):

if self._image_processor is None:

from transformers import CLIPProcessor

self._image_processor = CLIPProcessor.from_pretrained(self.image_model_name)

return self._image_processor

def _embed_image(self, inputs: List[Any]) -> "np.ndarray":

from pathlib import Path

import numpy as np

from PIL import Image

all_embeddings: List["np.ndarray"] = []

batch_size = self.config.batch_size

for start in range(0, len(inputs), batch_size):

batch = inputs[start : start + batch_size]

images = []

opened: List[Image.Image] = []

try:

for inp in batch:

if isinstance(

inp, (str, Path)

): # If the input string path is too large that It gives error and we could not open the image.

img = Image.open(inp)

opened.append(img)

images.append(img)

else:

images.append(inp)

processed = self.image_processor(images=images, return_tensors="pt")

finally:

for opened_img in opened:

opened_img.close()

embeddings = self.image_model.get_image_features(**processed)

embeddings = embeddings / embeddings.norm(p=2, dim=-1, keepdim=True)

all_embeddings.append(embeddings.detach().numpy())

return np.concatenate(all_embeddings, axis=0)

feast/sdk/python/feast/embedder.py at revert-6838-fix_minio_image · feast-dev/feast · GitHub

Repository navigation

Expand file tree

Breadcrumbs

embedder.py

Latest commit

History

Breadcrumbs

embedder.py

File metadata and controls