app.py

"""
SigLIP 2 Text & Image Encoder -- HuggingFace Space
Encodes text or image queries to 768-dim vectors for the Epstein photo search.

Model: google/siglip2-base-patch16-224
"""

import gradio as gr
import torch
import torch.nn.functional as F
from PIL import Image
from transformers import AutoModel, AutoTokenizer, AutoProcessor

MODEL_NAME = "google/siglip2-base-patch16-224"

print(f"Loading {MODEL_NAME}...")
model = AutoModel.from_pretrained(MODEL_NAME).eval()
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
processor = AutoProcessor.from_pretrained(MODEL_NAME)
print(f"Model loaded. Text hidden size: {model.config.text_config.hidden_size}")

def encode(text: str) -> list:
    inputs = tokenizer([text], return_tensors="pt", padding="max_length", max_length=64, truncation=True)
    with torch.no_grad():
        feats = model.text_model(**inputs).pooler_output
        feats = F.normalize(feats, dim=-1)
    return feats[0].tolist()

def encode_image(image) -> list:
    if image is None:
        raise gr.Error("No image provided")
    # Gradio 6.x base64 shortcut returns RGBA — SigLIP needs RGB
    if isinstance(image, Image.Image):
        image = image.convert("RGB")
    elif isinstance(image, str):
        image = Image.open(image).convert("RGB")
    else:
        raise gr.Error(f"Unexpected image type: {type(image)}")
    inputs = processor(images=[image], return_tensors="pt")
    with torch.no_grad():
        feats = model.get_image_features(pixel_values=inputs["pixel_values"])
        if not isinstance(feats, torch.Tensor):
            feats = feats.pooler_output
        feats = F.normalize(feats, dim=-1)
    return feats[0].tolist()

with gr.Blocks(title="SigLIP 2 Encoder") as demo:
    gr.Markdown("# SigLIP 2 Encoder\nEncodes text or images to 768-dim normalized vectors using google/siglip2-base-patch16-224")

    with gr.Tab("Text"):
        text_input = gr.Textbox(label="Text")
        text_output = gr.JSON(label="Embedding (768-dim)")
        text_btn = gr.Button("Encode Text")
        text_btn.click(fn=encode, inputs=text_input, outputs=text_output, api_name="encode")

    with gr.Tab("Image"):
        image_input = gr.Image(type="pil", label="Image")
        image_output = gr.JSON(label="Embedding (768-dim)")
        image_btn = gr.Button("Encode Image")
        image_btn.click(fn=encode_image, inputs=image_input, outputs=image_output, api_name="encode_image")

demo.launch()