app.py

import torch
import numpy as np
import gradio as gr
from faiss import read_index
from PIL import Image, ImageOps
from datasets import load_dataset
import torchvision.transforms as T
from torchvision.models import resnet50

from model import DINO

transforms = T.Compose(
    [T.ToTensor(), T.Resize(244), T.CenterCrop(224), T.Normalize([0.5], [0.5])]
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

datset = load_dataset("ethz/food101")

model = DINO(batch_size_per_device=32, num_classes=1000).to(device)
model.load_state_dict(torch.load("./bin/model.ckpt", map_location=device)["state_dict"])


def augment(img, transforms=transforms) -> torch.Tensor:
    img = Image.fromarray(img)
    if img.mode == "L":
        # Convert grayscale image to RGB by duplicating the single channel three times
        img = ImageOps.colorize(img, black="black", white="white")
    return transforms(img).unsqueeze(0)


def search_index(input_image, k: int):
    with torch.no_grad():
        embedding = model(augment(input_image))
        index = read_index("./bin/dino.index")
        _, I = index.search(np.array(embedding[0].reshape(1, -1)), k)
        indices = I[0]
        answer = ""
        for i, index in enumerate(indices[:3]):
            answer += index
            # retrieved_img = dataset["train"][int(index)]["image"]
    return answer


app = gr.Interface(
    search_index,
    inputs=[gr.Image(), gr.Slider(value=3, minimum=1, step=1)],
    outputs="text",
)

if __name__ == "__main__":
    app.launch()