app.py

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

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

def process_image(image):
    # Chuẩn hóa ảnh đầu vào
    encoding = feature_extractor(image, return_tensors="pt")
    
    # Forward qua model
    with torch.no_grad():
        outputs = model(**encoding)
        predicted_depth = outputs.predicted_depth
    
    # Resize output về đúng kích thước ảnh gốc
    prediction = torch.nn.functional.interpolate(
        predicted_depth.unsqueeze(1),
        size=image.size[::-1],  # (H, W)
        mode="bicubic",
        align_corners=False
    ).squeeze()
    
    # Chuyển thành ảnh uint8
    output = prediction.cpu().numpy()
    formatted = (output * 255 / np.max(output)).astype('uint8')
    img = Image.fromarray(formatted)
    
    return img

# Interface
title = "Demo: Zero-shot Depth Estimation with DPT"
description = "Intel's DPT: Dense Prediction Transformer for depth estimation from a single image."

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,
    allow_flagging="never"
)

iface.launch(debug=True)