File size: 4,101 Bytes
6835659 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 | from __future__ import annotations
from typing import Optional
import numpy as np
from src.embeddings.audio_embedder import AudioEmbedder
from src.embeddings.image_embedder import ImageEmbedder
from src.embeddings.projection import ProjectionHead
from src.embeddings.text_embedder import TextEmbedder
from src.utils.cache import EmbeddingCache
class AlignedEmbedder:
"""
Cross-modal embedding with correct space alignment.
Two pre-trained shared spaces are used:
- CLIP: text ↔ image (both from openai/clip-vit-base-patch32, 512-d)
- CLAP: text ↔ audio (both from laion/clap-htsat-unfused, 512-d)
For text-image similarity: use embed_text() and embed_image()
For text-audio similarity: use embed_text_for_audio() and embed_audio()
For image-audio similarity: these are cross-space (CLIP vs CLAP) —
no meaningful direct comparison without a trained bridge.
ProjectionHead is identity when in_dim == out_dim (preserving pre-trained
alignment). Only applies a linear transformation when dimensions differ.
"""
def __init__(
self,
target_dim: int = 512,
enable_cache: bool = True,
cache_dir: str = ".cache/embeddings",
):
self.text = TextEmbedder() # CLIP text encoder
self.image = ImageEmbedder() # CLIP image encoder
self.audio = AudioEmbedder() # CLAP audio encoder (also has text)
# Identity projections when dims match (512 → 512)
self.text_proj = ProjectionHead(512, target_dim)
self.image_proj = ProjectionHead(512, target_dim)
self.audio_proj = ProjectionHead(512, target_dim)
self.cache: Optional[EmbeddingCache] = None
if enable_cache:
self.cache = EmbeddingCache(cache_dir=cache_dir)
def embed_text(self, text: str) -> np.ndarray:
"""CLIP text embedding — use for text-image comparison."""
if self.cache:
cached = self.cache.get(text, "text")
if cached is not None:
return cached
emb = self.text.embed(text)
projected = self.text_proj.project(emb)
if self.cache:
self.cache.set(text, "text", projected)
return projected
def embed_text_for_audio(self, text: str) -> np.ndarray:
"""CLAP text embedding — use for text-audio comparison."""
if self.cache:
cached = self.cache.get(text, "text_clap")
if cached is not None:
return cached
emb = self.audio.embed_text(text)
projected = self.audio_proj.project(emb)
if self.cache:
self.cache.set(text, "text_clap", projected)
return projected
def embed_image(self, path: str) -> np.ndarray:
"""CLIP image embedding — use for text-image comparison."""
if self.cache:
cached = self.cache.get(path, "image")
if cached is not None:
return cached
emb = self.image.embed(path)
projected = self.image_proj.project(emb)
if self.cache:
self.cache.set(path, "image", projected)
return projected
def embed_audio(self, path: str) -> np.ndarray:
"""CLAP audio embedding — use for text-audio comparison."""
if self.cache:
cached = self.cache.get(path, "audio")
if cached is not None:
return cached
emb = self.audio.embed(path)
projected = self.audio_proj.project(emb)
if self.cache:
self.cache.set(path, "audio", projected)
return projected
@staticmethod
def shared(
target_dim: int = 512,
enable_cache: bool = True,
cache_dir: str = ".cache/embeddings",
) -> "AlignedEmbedder":
"""
Get the thread-safe shared singleton instance.
Use this in parallel experiments to avoid loading CLIP+CLAP per thread.
"""
from src.embeddings.shared_embedder import get_shared_embedder
return get_shared_embedder(target_dim, enable_cache, cache_dir)
|