app.py

import gradio as gr
from transformers import DPTFeatureExtractor, DPTForDepthEstimation
import torch
import numpy as np
from PIL import Image

# Load model và feature extractor
feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large")
model.eval()

def process_image(image):
    # Đảm bảo ảnh là RGB
    if image.mode != "RGB":
        image = image.convert("RGB")

    # Encode
    encoding = feature_extractor(image, return_tensors="pt")
    
    # Dự đoán depth
    with torch.no_grad():
        outputs = model(**encoding)
        predicted_depth = outputs.predicted_depth
    
    # Resize về kích thước ảnh gốc
    prediction = torch.nn.functional.interpolate(
        predicted_depth.unsqueeze(1),
        size=image.size[::-1],
        mode="bicubic",
        align_corners=False
    ).squeeze()
    
    # Chuẩn hóa và chuyển về ảnh uint8
    output = prediction.cpu().numpy()
    output = (output - np.min(output)) / (np.max(output) - np.min(output))  # normalize
    formatted = (output * 255).astype("uint8")
    depth_img = Image.fromarray(formatted)

    return depth_img

title = "Demo: zero-shot depth estimation with DPT"
description = "Demo for Intel's DPT, a Dense Prediction Transformer for state-of-the-art dense prediction tasks such as semantic segmentation and depth estimation."

iface = gr.Interface(
    fn=process_image,
    inputs=gr.inputs.Image(type="pil", label="Input Image"),
    outputs=gr.outputs.Image(type="pil", label="Predicted Depth"),
    title=title,
    description=description,
)

iface.launch(debug=True)