|
import torch |
|
from typing import Dict, Any, List |
|
from PIL import Image |
|
import base64 |
|
from io import BytesIO |
|
import logging |
|
from transformers import AutoImageProcessor, AutoModel |
|
import os |
|
from dataclasses import dataclass |
|
|
|
|
|
|
|
@dataclass |
|
class ImageEncodingResult: |
|
image_encoded: List[List[float]] |
|
image_encoded_average: List[float] |
|
|
|
|
|
class EndpointHandler: |
|
""" |
|
A handler class for processing images and generating embeddings using a pre-trained model. |
|
Attributes: |
|
processor: The pre-trained image processor. |
|
model: The pre-trained model for generating embeddings. |
|
device: The device (CPU or CUDA) used to run model inference. |
|
""" |
|
|
|
def __init__(self, path: str = ""): |
|
""" |
|
Initializes the EndpointHandler with the model and processor from the current directory. |
|
""" |
|
|
|
logging.basicConfig(level=logging.INFO) |
|
self.logger = logging.getLogger(__name__) |
|
|
|
|
|
self.device = "cuda" if torch.cuda.is_available() else "cpu" |
|
self.logger.info(f"Using device: {self.device}") |
|
|
|
|
|
self.logger.info("Loading model and processor from the current directory.") |
|
try: |
|
self.processor = AutoImageProcessor.from_pretrained(path) |
|
self.model = AutoModel.from_pretrained(path, trust_remote_code=True).to( |
|
self.device |
|
) |
|
self.logger.info("Model and processor loaded successfully.") |
|
except Exception as e: |
|
self.logger.error(f"Failed to load model or processor: {e}") |
|
raise |
|
|
|
def _resize_image_if_large( |
|
self, image: Image.Image, max_size: int = 1080 |
|
) -> Image.Image: |
|
""" |
|
Resizes an image if its dimensions exceed the specified maximum size. |
|
Args: |
|
image (Image.Image): Input image. |
|
max_size (int): Maximum size for the image dimensions. |
|
Returns: |
|
Image.Image: Resized image. |
|
""" |
|
width, height = image.size |
|
if width > max_size or height > max_size: |
|
scale = max_size / max(width, height) |
|
new_width = int(width * scale) |
|
new_height = int(height * scale) |
|
image = image.resize((new_width, new_height), resample=Image.BILINEAR) |
|
return image |
|
|
|
def _encode_image(self, image: Image.Image) -> ImageEncodingResult: |
|
""" |
|
Encodes an image into embeddings using the model. |
|
Args: |
|
image (Image.Image): Input image. |
|
Returns: |
|
ImageEncodingResult: Dataclass containing the encoded embeddings and their average. |
|
""" |
|
try: |
|
|
|
image = self._resize_image_if_large(image) |
|
|
|
|
|
inputs = self.processor(image, return_tensors="pt").to(self.device) |
|
with torch.inference_mode(): |
|
outputs = self.model(**inputs) |
|
last_hidden_state = outputs.last_hidden_state |
|
image_encoded = last_hidden_state.squeeze().tolist() |
|
image_encoded_average = last_hidden_state.mean(dim=1).squeeze().tolist() |
|
|
|
return ImageEncodingResult( |
|
image_encoded=image_encoded, |
|
image_encoded_average=image_encoded_average, |
|
) |
|
except Exception as e: |
|
self.logger.error(f"Error encoding image: {e}") |
|
raise |
|
|
|
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]: |
|
""" |
|
Processes input data containing base64-encoded images and generates embeddings. |
|
Args: |
|
data (Dict[str, Any]): Dictionary containing input images. |
|
Returns: |
|
Dict[str, Any]: Dictionary containing encoded embeddings or error messages. |
|
""" |
|
images_data = data.get("inputs", []) |
|
|
|
if not images_data: |
|
return {"error": "No image data provided."} |
|
|
|
results = [] |
|
for img_data in images_data: |
|
if isinstance(img_data, str): |
|
try: |
|
|
|
image_bytes = base64.b64decode(img_data) |
|
image = Image.open(BytesIO(image_bytes)).convert("RGB") |
|
|
|
|
|
encoded_image = self._encode_image(image) |
|
results.append(encoded_image) |
|
except Exception as e: |
|
self.logger.error(f"Invalid image data: {e}") |
|
return {"error": f"Invalid image data: {e}"} |
|
else: |
|
self.logger.error("Images should be base64-encoded strings.") |
|
return {"error": "Images should be base64-encoded strings."} |
|
|
|
|
|
return {"results": [result.__dict__ for result in results]} |
|
|