app.py

import gradio as gr
from transformers import SegformerFeatureExtractor, SegformerForSemanticSegmentation
from PIL import Image
import numpy as np
import torch

# 모델과 feature extractor 로드
model_name = "nvidia/segformer-b0-finetuned-ade-512-512"
model = SegformerForSemanticSegmentation.from_pretrained(model_name)
feature_extractor = SegformerFeatureExtractor.from_pretrained(model_name)

def create_color_map(num_classes):
    """ 임의의 색상 매핑 생성 """
    np.random.seed(42)  # 재현성을 위한 시드 설정
    return {i: np.random.randint(0, 256, 3) for i in range(num_classes)}

def segment_image(image):
    # 이미지 처리
    inputs = feature_extractor(images=image, return_tensors="pt")
    with torch.no_grad():
        outputs = model(**inputs)

    # 마스크 생성
    upsampled_logits = torch.nn.functional.interpolate(
        outputs.logits, size=image.size[::-1], mode="bilinear", align_corners=False
    )
    upsampled_predictions = upsampled_logits.argmax(dim=1)
    mask = upsampled_predictions.squeeze().numpy()

    # 색상 매핑
    color_map = create_color_map(150)  # ADE20K에는 약 150개의 클래스가 있음
    colored_mask = np.array([color_map[class_id] for class_id in mask.flatten()]).reshape(mask.shape + (3,))

    # 결과 반환
    return Image.fromarray(colored_mask.astype(np.uint8))

# 예시 이미지 경로
example_images = ["path/to/image1.jpg", "path/to/image2.jpg", "path/to/image3.jpg"]

# Gradio 인터페이스 설정
iface = gr.Interface(
    fn=segment_image,
    inputs=gr.inputs.Image(type="pil"),
    outputs="image",
    title="Image Segmentation with SegFormer",
    description="Upload an image to segment it using SegFormer model.",
    examples=example_images
)

# 인터페이스 실행
iface.launch()