app.py

from typing import Any, Dict

import spaces

import cv2
import gradio as gr
import numpy as np
import torch
from gradio_image_annotation import image_annotator
from sam2 import load_model
from sam2.sam2_image_predictor import SAM2ImagePredictor

from src.plot_utils import export_mask


@spaces.GPU()
def predict(model_choice, annotations: Dict[str, Any]):
    # device = "cuda" if torch.cuda.is_available() else "cpu"
    torch.autocast(device_type="cuda", dtype=torch.bfloat16).__enter__()
    sam2_model = load_model(
        variant=model_choice,
        ckpt_path=f"assets/checkpoints/sam2_hiera_{model_choice}.pt",
        device="cuda",
    )
    predictor = SAM2ImagePredictor(sam2_model)  # type:ignore
    predictor.set_image(annotations["image"])
    coordinates = []
    for i in range(len(annotations["boxes"])):
        coordinate = [
            int(annotations["boxes"][i]["xmin"]),
            int(annotations["boxes"][i]["ymin"]),
            int(annotations["boxes"][i]["xmax"]),
            int(annotations["boxes"][i]["ymax"]),
        ]
        coordinates.append(coordinate)

    masks, scores, _ = predictor.predict(
        point_coords=None,
        point_labels=None,
        box=np.array(coordinates),
        multimask_output=False,
    )

    if masks.shape[0] == 1:
        # handle single mask cases
        masks = np.expand_dims(masks, axis=0)

    return export_mask(masks)


with gr.Blocks(delete_cache=(30, 30)) as demo:
    gr.Markdown(
        """
        # 1. Choose Model Checkpoint
        """
    )
    with gr.Row():
        model = gr.Dropdown(
            choices=["tiny", "small", "base_plus", "large"],
            value="tiny",
            label="Model Checkpoint",
            info="Which model checkpoint to load?",
        )

    gr.Markdown(
        """
        # 2. Upload your Image and draw bounding box(es)
        """
    )

    annotator = image_annotator(
        value={"image": cv2.imread("assets/example.png")},
        disable_edit_boxes=True,
        label="Draw a bounding box",
    )
    btn = gr.Button("Get Segmentation Mask(s)")
    btn.click(
        fn=predict, inputs=[model, annotator], outputs=[gr.Image(label="Mask(s)")]
    )

demo.launch()