app.py

import os
import cv2
import torch
import numpy as np
import gradio as gr
from contextlib import nullcontext

# DA² utilities
from da2.utils.base import load_config
from da2.utils.model import load_model
from da2.utils.io import read_cv2_image, torch_transform, tensorize

# 引入真正的 Accelerator
from accelerate import Accelerator
from accelerate.logging import get_logger

# ============================================================
# Global Initialization (Run once at startup)
# ============================================================
def initialize_app(config_path="configs/infer.json"):
    # 1. 初始化 Accelerator (解決 logging 報錯)
    accelerator = Accelerator()
    
    # 2. 載入 Config
    config = load_config(config_path)

    # 設定 Logger
    logger = get_logger(__name__, log_level="INFO")
    config.setdefault("env", {})
    config["env"]["logger"] = logger
    config["env"].setdefault("seed", 42)

    accelerator.print(f"Running on device: {accelerator.device}")
    
    # 3. 載入模型 (Global Load，避免每次推論重載)
    model = load_model(config, accelerator)
    model = model.to(accelerator.device)
    model.eval()

    return config, accelerator, model

# 初始化全局變數
try:
    CONFIG, ACCELERATOR, MODEL = initialize_app()
    print("Model loaded successfully!")
except Exception as e:
    print(f"Error loading model: {e}")
    CONFIG, ACCELERATOR, MODEL = None, None, None

# ============================================================
# Mask loader
# ============================================================
def read_mask_demo(mask_path, img_shape):
    if mask_path is None:
        if len(img_shape) == 3:
            return np.ones((img_shape[1], img_shape[2]), dtype=bool)
        return np.ones(img_shape[:2], dtype=bool)

    mask = cv2.imread(mask_path, cv2.IMREAD_GRAYSCALE)
    if mask is None:
        if len(img_shape) == 3:
            return np.ones((img_shape[1], img_shape[2]), dtype=bool)
        return np.ones(img_shape[:2], dtype=bool)

    # 確保 Mask 尺寸也跟隨 Tensor (如果需要的話)，但通常 Mask 是跟原圖
    # 這裡簡單處理，如果尺寸不對稍微 resize 一下避免崩潰
    if mask.shape[:2] != img_shape[-2:]:
         mask = cv2.resize(mask, (img_shape[-1], img_shape[-2]), interpolation=cv2.INTER_NEAREST)

    return mask > 0

# ============================================================
# Core inference function
# ============================================================
def run_inference_and_save_depth(image_path, mask_path=None):
    if MODEL is None:
        return None, "Error: Model not loaded."

    device = ACCELERATOR.device

    # 1. 讀取原始圖片
    cv2_img = read_cv2_image(image_path)
    if cv2_img is None:
        print(f"Error reading image: {image_path}")
        return None, None

    # 【關鍵修復】獲取原始圖片的寬高 (Height, Width)
    # cv2 shape 是 (H, W, C)
    original_h, original_w = cv2_img.shape[:2]

    # 2. 轉換為 Tensor (這步通常會 resize 成 518x518 或其他模型輸入尺寸)
    img_tensor = torch_transform(cv2_img)  # CxHxW tensor
    
    # 處理 Mask
    mask = read_mask_demo(mask_path, img_tensor.shape)

    # 準備輸入
    model_dtype = CONFIG.get("spherevit", {}).get("dtype", "float32")
    input_tensor = tensorize(img_tensor, model_dtype, device)

    # Autocast
    use_autocast = (device.type == "cuda")
    autocast_ctx = torch.autocast(device_type="cuda") if use_autocast else nullcontext()

    # 3. 推論
    with autocast_ctx, torch.no_grad():
        pred = MODEL(input_tensor)

        if isinstance(pred, (tuple, list)):
            pred = pred[0]

        # 轉回 Numpy float32
        depth = pred.float().cpu().numpy()

    # 4. 後處理
    depth = np.squeeze(depth)

    # 【關鍵修復】將深度圖 Resize 回原始尺寸
    # cv2.resize 接受的參數是 (Width, Height)
    if (depth.shape[0] != original_h) or (depth.shape[1] != original_w):
        depth = cv2.resize(depth, (original_w, original_h), interpolation=cv2.INTER_CUBIC)

    # 5. 正規化 (Normalization) -> 8-bit
    dmin, dmax = float(np.nanmin(depth)), float(np.nanmax(depth))

    if dmax - dmin > 1e-6:
        depth_norm = (depth - dmin) / (dmax - dmin)
    else:
        depth_norm = np.zeros_like(depth, dtype=np.float32)

    depth_8bit = (depth_norm * 255).astype(np.uint8)

    # 6. 儲存
    os.makedirs("outputs", exist_ok=True)
    base = os.path.splitext(os.path.basename(image_path))[0]
    out_path = f"outputs/{base}_depth.png"
    cv2.imwrite(out_path, depth_8bit)

    return depth_8bit, out_path

# ============================================================
# Gradio UI
# ============================================================
def gradio_fn(image, mask):
    if image is None:
        return None, None
    depth_img, out_path = run_inference_and_save_depth(image, mask)
    return depth_img, out_path

demo = gr.Interface(
    fn=gradio_fn,
    inputs=[
        gr.Image(label="Input Image", type="filepath"),
        gr.Image(label="Optional Mask", type="filepath"),
    ],
    outputs=[
        gr.Image(label="Depth (8-bit Grayscale)", type="numpy"),
        gr.File(label="Download Depth PNG"),
    ],
    title="DA² — Minimal Depth Demo",
    description="Upload an image (and optional mask) -> outputs an 8-bit grayscale depth PNG (Resized to Original).",
    allow_flagging="never",
)

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