imagebind / main.py
fcastrovilli's picture
refactor: computeSimilarities
63b0848
import os
import torch
from imagebind import data
from imagebind.models import imagebind_model
from imagebind.models.imagebind_model import ModalityType
from pydub import AudioSegment
from fastapi import FastAPI, UploadFile, File, Form
from typing import List, Dict
import tempfile
from pydantic import BaseModel
import uvicorn
import numpy as np
from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
from fastapi import Depends, HTTPException, status
app = FastAPI()
# Add these lines after the app initialization
security = HTTPBearer()
API_TOKEN = os.getenv("API_TOKEN", "your-default-token-here") # Set a default token or use environment variable
# Add this function for token verification
async def verify_token(credentials: HTTPAuthorizationCredentials = Depends(security)):
if credentials.credentials != API_TOKEN:
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Invalid authentication token",
headers={"WWW-Authenticate": "Bearer"},
)
return credentials.credentials
def convert_audio_to_wav(audio_path: str) -> str:
"""Convert MP3 to WAV if necessary."""
if audio_path.lower().endswith('.mp3'):
wav_path = audio_path.rsplit('.', 1)[0] + '.wav'
if not os.path.exists(wav_path):
audio = AudioSegment.from_mp3(audio_path)
audio.export(wav_path, format='wav')
return wav_path
return audio_path
class EmbeddingManager:
def __init__(self):
self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
self.model = imagebind_model.imagebind_huge(pretrained=True)
self.model.eval()
self.model.to(self.device)
def compute_embeddings(self,
images: List[str] = None,
audio_files: List[str] = None,
texts: List[str] = None) -> dict:
"""Compute embeddings for provided modalities only."""
with torch.no_grad():
inputs = {}
if texts:
inputs[ModalityType.TEXT] = data.load_and_transform_text(texts, self.device)
if images:
inputs[ModalityType.VISION] = data.load_and_transform_vision_data(images, self.device)
if audio_files:
inputs[ModalityType.AUDIO] = data.load_and_transform_audio_data(audio_files, self.device)
if not inputs:
return {}
embeddings = self.model(inputs)
result = {}
if ModalityType.VISION in inputs:
result['vision'] = embeddings[ModalityType.VISION].cpu().numpy().tolist()
if ModalityType.AUDIO in inputs:
result['audio'] = embeddings[ModalityType.AUDIO].cpu().numpy().tolist()
if ModalityType.TEXT in inputs:
result['text'] = embeddings[ModalityType.TEXT].cpu().numpy().tolist()
return result
@staticmethod
def compute_similarities(embeddings: Dict[str, List[List[float]]]) -> dict:
"""Compute similarities between available embeddings."""
similarities = {}
# Convert available embeddings to tensors
tensors = {
k: torch.tensor(v) for k, v in embeddings.items()
if isinstance(v, (list, np.ndarray)) and len(v) > 0
}
# Compute cross-modal similarities
modality_pairs = [
('vision', 'audio', 'vision_audio'),
('vision', 'text', 'vision_text'),
('audio', 'text', 'audio_text')
]
for mod1, mod2, key in modality_pairs:
if mod1 in tensors and mod2 in tensors:
similarities[key] = torch.softmax(
tensors[mod1] @ tensors[mod2].T,
dim=-1
).numpy().tolist()
# Compute same-modality similarities
for modality in ['vision', 'audio', 'text']:
if modality in tensors:
key = f'{modality}_{modality}'
similarities[key] = torch.softmax(
tensors[modality] @ tensors[modality].T,
dim=-1
).numpy().tolist()
return similarities
# Initialize the embedding manager
embedding_manager = EmbeddingManager()
class EmbeddingResponse(BaseModel):
embeddings: dict
file_names: dict
class SimilarityRequest(BaseModel):
embeddings: Dict[str, List[List[float]]]
threshold: float = 0.5
top_k: int | None = None
include_self_similarity: bool = False
normalize_scores: bool = True
class SimilarityMatch(BaseModel):
index_a: int
index_b: int
score: float
modality_a: str
modality_b: str
item_a: str # Original item identifier (filename or text)
item_b: str # Original item identifier (filename or text)
class SimilarityResponse(BaseModel):
matches: List[SimilarityMatch]
statistics: Dict[str, float] # Contains avg_score, max_score, etc.
modality_pairs: List[str] # Lists which modality comparisons were performed
class ModalityPair:
def __init__(self, mod1: str, mod2: str):
self.mod1 = min(mod1, mod2) # Ensure consistent ordering
self.mod2 = max(mod1, mod2)
def __str__(self):
return f"{self.mod1}_to_{self.mod2}"
def compute_similarity_matrix(tensor1: torch.Tensor, tensor2: torch.Tensor, normalize: bool = True) -> torch.Tensor:
"""Compute cosine similarity between two sets of embeddings."""
# Normalize embeddings if requested
if normalize:
tensor1 = torch.nn.functional.normalize(tensor1, dim=1)
tensor2 = torch.nn.functional.normalize(tensor2, dim=1)
# Compute similarity matrix
similarity = torch.matmul(tensor1, tensor2.T)
return similarity
def get_top_k_matches(similarity_matrix: torch.Tensor, top_k: int | None = None) -> List[tuple]:
"""Get top-k matches from a similarity matrix."""
if top_k is None:
top_k = similarity_matrix.numel()
# Flatten and get top-k indices
flat_sim = similarity_matrix.flatten()
top_k = min(top_k, flat_sim.numel())
values, indices = torch.topk(flat_sim, k=top_k)
# Convert flat indices to 2D indices
rows = indices // similarity_matrix.size(1)
cols = indices % similarity_matrix.size(1)
return [(r.item(), c.item(), v.item()) for r, c, v in zip(rows, cols, values)]
@app.post("/compute_embeddings", response_model=EmbeddingResponse)
async def generate_embeddings(
credentials: HTTPAuthorizationCredentials = Depends(verify_token),
texts: str | None = Form(None),
images: List[UploadFile] | None = File(default=None),
audio_files: List[UploadFile] | None = File(default=None)
):
"""Generate embeddings for any provided files and texts."""
temp_files = []
try:
image_paths = []
image_names = []
audio_paths = []
audio_names = []
text_list = []
# Process images if provided
if images:
for img in images:
with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(img.filename)[1]) as tmp:
content = await img.read()
tmp.write(content)
image_paths.append(tmp.name)
image_names.append(img.filename)
temp_files.append(tmp.name)
# Process audio files if provided
if audio_files:
for audio in audio_files:
with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(audio.filename)[1]) as tmp:
content = await audio.read()
tmp.write(content)
audio_path = convert_audio_to_wav(tmp.name)
audio_paths.append(audio_path)
audio_names.append(audio.filename)
temp_files.append(tmp.name)
if audio_path != tmp.name:
temp_files.append(audio_path)
# Process texts if provided
if texts:
text_list = [text.strip() for text in texts.split('\n') if text.strip()]
# Compute embeddings only if we have any input
if not any([image_paths, audio_paths, text_list]):
return EmbeddingResponse(
embeddings={},
file_names={}
)
embeddings = embedding_manager.compute_embeddings(
image_paths if image_paths else None,
audio_paths if audio_paths else None,
text_list if text_list else None
)
file_names = {}
if image_names:
file_names['images'] = image_names
if audio_names:
file_names['audio'] = audio_names
if text_list:
file_names['texts'] = text_list
return EmbeddingResponse(
embeddings=embeddings,
file_names=file_names
)
finally:
# Clean up temporary files
for temp_file in temp_files:
try:
os.unlink(temp_file)
except:
pass
@app.post("/compute_similarities", response_model=SimilarityResponse)
async def compute_similarities(
request: SimilarityRequest,
file_names: Dict[str, List[str]], # Maps modality to list of file/text names
credentials: HTTPAuthorizationCredentials = Depends(verify_token)
):
"""
Compute cross-modal similarities with advanced filtering and matching options.
Parameters:
- embeddings: Dict mapping modality to embedding tensors
- threshold: Minimum similarity score to include in results
- top_k: Maximum number of matches to return (per modality pair)
- include_self_similarity: Whether to include same-item comparisons
- normalize_scores: Whether to normalize embeddings before comparison
- file_names: Dict mapping modality to list of original file/text names
"""
matches = []
statistics = {
"avg_score": 0.0,
"max_score": 0.0,
"min_score": 1.0,
"total_comparisons": 0
}
# Convert embeddings to tensors
tensors = {
k: torch.tensor(v) for k, v in request.embeddings.items()
if isinstance(v, (list, np.ndarray)) and len(v) > 0
}
modality_pairs = []
all_scores = []
# Get all possible modality pairs
modalities = list(tensors.keys())
for i, mod1 in enumerate(modalities):
for mod2 in modalities[i:]: # Include self-comparisons if requested
if mod1 == mod2 and not request.include_self_similarity:
continue
pair = ModalityPair(mod1, mod2)
modality_pairs.append(str(pair))
# Compute similarity matrix
sim_matrix = compute_similarity_matrix(
tensors[mod1],
tensors[mod2],
normalize=request.normalize_scores
)
# Get top matches
top_matches = get_top_k_matches(sim_matrix, request.top_k)
# Filter by threshold and create match objects
for idx_a, idx_b, score in top_matches:
if score < request.threshold:
continue
# Skip self-matches if not requested
if mod1 == mod2 and idx_a == idx_b and not request.include_self_similarity:
continue
matches.append(SimilarityMatch(
index_a=idx_a,
index_b=idx_b,
score=float(score),
modality_a=mod1,
modality_b=mod2,
item_a=file_names[mod1][idx_a],
item_b=file_names[mod2][idx_b]
))
all_scores.append(score)
# Compute statistics
if all_scores:
statistics.update({
"avg_score": float(np.mean(all_scores)),
"max_score": float(np.max(all_scores)),
"min_score": float(np.min(all_scores)),
"total_comparisons": len(all_scores)
})
# Sort matches by score in descending order
matches.sort(key=lambda x: x.score, reverse=True)
return SimilarityResponse(
matches=matches,
statistics=statistics,
modality_pairs=modality_pairs
)
@app.get("/health")
async def health_check(
credentials: HTTPAuthorizationCredentials = Depends(verify_token)
):
"""Basic healthcheck endpoint that returns the status of the service."""
return {
"status": "healthy",
"model_device": embedding_manager.device
}
if __name__ == "__main__":
uvicorn.run(app, host="0.0.0.0", port=7860)