import torch
from torch import nn
import torch.nn.functional as F
import torchvision.transforms.functional as TF

from torch import Tensor
import spaces

import numpy as np
from PIL import Image
import gradio as gr
from matplotlib import cm
from huggingface_hub import hf_hub_download
from warnings import warn

from models import get_model


mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
alpha = 0.8
EPS = 1e-8
loaded_model = None
current_model_config = {"variant": None, "dataset": None, "metric": None}

pretrained_models = [
    "ZIP-B @ ShanghaiTech A @ MAE", "ZIP-B @ ShanghaiTech A @ NAE",
    "ZIP-B @ ShanghaiTech B @ MAE", "ZIP-B @ ShanghaiTech B @ NAE",
    "ZIP-B @ UCF-QNRF @ MAE", "ZIP-B @ UCF-QNRF @ NAE",
    "ZIP-B @ NWPU-Crowd @ MAE", "ZIP-B @ NWPU-Crowd @ NAE",
    "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━",
    "ZIP-S @ ShanghaiTech A @ MAE", "ZIP-S @ ShanghaiTech A @ NAE",
    "ZIP-S @ ShanghaiTech B @ MAE", "ZIP-S @ ShanghaiTech B @ NAE",
    "ZIP-S @ UCF-QNRF @ MAE", "ZIP-S @ UCF-QNRF @ NAE",
    "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━",
    "ZIP-T @ ShanghaiTech A @ MAE", "ZIP-T @ ShanghaiTech A @ NAE",
    "ZIP-T @ ShanghaiTech B @ MAE", "ZIP-T @ ShanghaiTech B @ NAE",
    "ZIP-T @ UCF-QNRF @ MAE", "ZIP-T @ UCF-QNRF @ NAE",
    "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━",
    "ZIP-N @ ShanghaiTech A @ MAE", "ZIP-N @ ShanghaiTech A @ NAE",
    "ZIP-N @ ShanghaiTech B @ MAE", "ZIP-N @ ShanghaiTech B @ NAE",
    "ZIP-N @ UCF-QNRF @ MAE", "ZIP-N @ UCF-QNRF @ NAE",
    "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━",
    "ZIP-P @ ShanghaiTech A @ MAE", "ZIP-P @ ShanghaiTech A @ NAE",
    "ZIP-P @ ShanghaiTech B @ MAE", "ZIP-P @ ShanghaiTech B @ NAE",
    "ZIP-P @ UCF-QNRF @ MAE", "ZIP-P @ UCF-QNRF @ NAE",
]

# -----------------------------
# Model management functions
# -----------------------------
def update_model_if_needed(variant_dataset_metric: str):
    """
    Load a new model only if the configuration has changed.
    """
    global loaded_model, current_model_config
    
    # 如果是分割线，则跳过
    if "━━━━━━" in variant_dataset_metric:
        return "Please select a valid model configuration"
    
    parts = variant_dataset_metric.split(" @ ")
    if len(parts) != 3:
        return "Invalid model configuration format"
        
    variant, dataset, metric = parts[0], parts[1], parts[2].lower()

    if dataset == "ShanghaiTech A":
        dataset_name = "sha"
    elif dataset == "ShanghaiTech B":
        dataset_name = "shb"
    elif dataset == "UCF-QNRF":
        dataset_name = "qnrf"
    elif dataset == "NWPU-Crowd":
        dataset_name = "nwpu"
    else:
        return f"Unknown dataset: {dataset}"

    # 只更新配置，不在主进程中加载模型
    if (current_model_config["variant"] != variant or 
        current_model_config["dataset"] != dataset_name or 
        current_model_config["metric"] != metric):
        
        print(f"Model configuration updated: {variant} @ {dataset} with {metric} metric")
        current_model_config = {"variant": variant, "dataset": dataset_name, "metric": metric}
        loaded_model = None  # 重置模型，将在GPU进程中重新加载
        return f"Model configuration set: {variant} @ {dataset} ({metric})"
    else:
        print(f"Model configuration unchanged: {variant} @ {dataset} with {metric} metric")
        return f"Model configuration: {variant} @ {dataset} ({metric})"


# -----------------------------
# Define the model architecture
# -----------------------------
def load_model(variant: str, dataset: str = "ShanghaiTech B", metric: str = "mae"):
    """ Load the model weights from the Hugging Face Hub."""
    # global loaded_model
    # Build model

    model_info_path = hf_hub_download(
        repo_id=f"Yiming-M/{variant}",
        filename=f"checkpoints/{dataset}/best_{metric}.pth",
    )

    model = get_model(model_info_path=model_info_path)
    model.eval()
    # loaded_model = model
    return model


def _calc_size(
    img_w: int,
    img_h: int,
    min_size: int,
    max_size: int,
    base: int = 32
):
    """
    This function generates a new size for an image while keeping the aspect ratio. The new size should be within the given range (min_size, max_size).

    Args:
        img_w (int): The width of the image.
        img_h (int): The height of the image.
        min_size (int): The minimum size of the edges of the image.
        max_size (int): The maximum size of the edges of the image.
        # base (int): The base number to which the new size should be a multiple of.
    """
    assert min_size % base == 0, f"min_size ({min_size}) must be a multiple of {base}"
    if max_size != float("inf"):
        assert max_size % base == 0, f"max_size ({max_size}) must be a multiple of {base} if provided"

    assert min_size <= max_size, f"min_size ({min_size}) must be less than or equal to max_size ({max_size})"

    aspect_ratios = (img_w / img_h, img_h / img_w)
    if min_size / max_size <= min(aspect_ratios) <= max(aspect_ratios) <= max_size / min_size:  # possible to resize and preserve the aspect ratio
        if min_size <= min(img_w, img_h) <= max(img_w, img_h) <= max_size:  # already within the range, no need to resize
            ratio = 1.
        elif min(img_w, img_h) < min_size:  # smaller than the minimum size, resize to the minimum size
            ratio = min_size / min(img_w, img_h)
        else:  # larger than the maximum size, resize to the maximum size
            ratio = max_size / max(img_w, img_h)

        new_w, new_h = int(round(img_w * ratio / base) * base), int(round(img_h * ratio / base) * base)
        new_w = max(min_size, min(max_size, new_w))
        new_h = max(min_size, min(max_size, new_h))
        return new_w, new_h

    else:  # impossible to resize and preserve the aspect ratio
        msg = f"Impossible to resize {img_w}x{img_h} image while preserving the aspect ratio to a size within the range ({min_size}, {max_size}). Will not limit the maximum size."
        warn(msg)
        return _calc_size(img_w, img_h, min_size, float("inf"), base)
    

# -----------------------------
# Preprocessing function
# -----------------------------
# Adjust the image transforms to match what your model expects.
def transform(image: Image.Image, dataset_name: str) -> Tensor:
    assert isinstance(image, Image.Image), "Input must be a PIL Image"
    image_tensor = TF.to_tensor(image)

    if dataset_name == "sha":
        min_size = 448
        max_size = float("inf")
    elif dataset_name == "shb":
        min_size = 448
        max_size = float("inf")
    elif dataset_name == "qnrf":
        min_size = 448
        max_size = 2048
    elif dataset_name == "nwpu":
        min_size = 448
        max_size = 3072

    image_height, image_width = image_tensor.shape[-2:]
    new_width, new_height = _calc_size(
        img_w=image_width,
        img_h=image_height,
        min_size=min_size,
        max_size=max_size,
        base=32
    )
    if new_height != image_height or new_width != image_width:
        image_tensor = TF.resize(image_tensor, size=(new_height, new_width), interpolation=TF.InterpolationMode.BICUBIC, antialias=True)

    image_tensor = TF.normalize(image_tensor, mean=mean, std=std)
    return image_tensor.unsqueeze(0)  # Add batch dimension


def _sliding_window_predict(
    model: nn.Module,
    image: Tensor,
    window_size: int,
    stride: int, 
    max_num_windows: int = 256
):
    assert len(image.shape) == 4, f"Image must be a 4D tensor (1, c, h, w), got {image.shape}"
    window_size = (int(window_size), int(window_size)) if isinstance(window_size, (int, float)) else window_size
    stride = (int(stride), int(stride)) if isinstance(stride, (int, float)) else stride
    window_size = tuple(window_size)
    stride = tuple(stride)
    assert isinstance(window_size, tuple) and len(window_size) == 2 and window_size[0] > 0 and window_size[1] > 0, f"Window size must be a positive integer tuple (h, w), got {window_size}"
    assert isinstance(stride, tuple) and len(stride) == 2 and stride[0] > 0 and stride[1] > 0, f"Stride must be a positive integer tuple (h, w), got {stride}"
    assert stride[0] <= window_size[0] and stride[1] <= window_size[1], f"Stride must be smaller than window size, got {stride} and {window_size}"

    image_height, image_width = image.shape[-2:]
    window_height, window_width = window_size
    assert image_height >= window_height and image_width >= window_width, f"Image size must be larger than window size, got image size {image.shape} and window size {window_size}"
    stride_height, stride_width = stride

    num_rows = int(np.ceil((image_height - window_height) / stride_height) + 1)
    num_cols = int(np.ceil((image_width - window_width) / stride_width) + 1)

    if hasattr(model, "block_size"):
        block_size = model.block_size
    elif hasattr(model, "module") and hasattr(model.module, "block_size"):
        block_size = model.module.block_size
    else:
        raise ValueError("Model must have block_size attribute")
    assert window_height % block_size == 0 and window_width % block_size == 0, f"Window size must be divisible by block size, got {window_size} and {block_size}"

    windows = []
    for i in range(num_rows):
        for j in range(num_cols):
            x_start, y_start = i * stride_height, j * stride_width
            x_end, y_end = x_start + window_height, y_start + window_width
            if x_end > image_height:
                x_start, x_end = image_height - window_height, image_height
            if y_end > image_width:
                y_start, y_end = image_width - window_width, image_width

            window = image[:, :, x_start:x_end, y_start:y_end]
            windows.append(window)

    windows = torch.cat(windows, dim=0).to(image.device)  # batched windows, shape: (num_windows, c, h, w)
    
    model.eval()
    pi_maps, lambda_maps = [], []
    for i in range(0, len(windows), max_num_windows):
        with torch.no_grad():
            image_feats = model.backbone(windows[i: min(i + max_num_windows, len(windows))])
            pi_image_feats, lambda_image_feats = model.pi_head(image_feats), model.lambda_head(image_feats)
            pi_image_feats = F.normalize(pi_image_feats.permute(0, 2, 3, 1), p=2, dim=-1)  # shape (B, H, W, C)
            lambda_image_feats = F.normalize(lambda_image_feats.permute(0, 2, 3, 1), p=2, dim=-1)  # shape (B, H, W, C)

            pi_text_feats, lambda_text_feats = model.pi_text_feats, model.lambda_text_feats
            pi_logit_scale, lambda_logit_scale = model.pi_logit_scale.exp(), model.lambda_logit_scale.exp()

            pi_logit_map = pi_logit_scale * pi_image_feats @ pi_text_feats.t()  # (B, H, W, 2), logits per image
            lambda_logit_map = lambda_logit_scale * lambda_image_feats @ lambda_text_feats.t()  # (B, H, W, N - 1), logits per image

            pi_logit_map =  pi_logit_map.permute(0, 3, 1, 2)  # (B, 2, H, W)
            lambda_logit_map = lambda_logit_map.permute(0, 3, 1, 2)  # (B, N - 1, H, W)

            lambda_map = (lambda_logit_map.softmax(dim=1) * model.bin_centers[:, 1:]).sum(dim=1, keepdim=True)  # (B, 1, H, W)
            pi_map = pi_logit_map.softmax(dim=1)[:, 0:1]  # (B, 1, H, W)

            pi_maps.append(pi_map.cpu().numpy())
            lambda_maps.append(lambda_map.cpu().numpy())

    # assemble the density map
    pi_maps = np.concatenate(pi_maps, axis=0)  # shape: (num_windows, 1, H, W)
    lambda_maps = np.concatenate(lambda_maps, axis=0)  # shape: (num_windows, 1, H, W)
    assert pi_maps.shape == lambda_maps.shape, f"pi_maps and lambda_maps must have the same shape, got {pi_maps.shape} and {lambda_maps.shape}"

    pi_map = np.zeros((pi_maps.shape[1], image_height // block_size, image_width // block_size), dtype=np.float32)
    lambda_map = np.zeros((lambda_maps.shape[1], image_height // block_size, image_width // block_size), dtype=np.float32)
    count_map = np.zeros((pi_maps.shape[1], image_height // block_size, image_width // block_size), dtype=np.float32)
    idx = 0
    for i in range(num_rows):
        for j in range(num_cols):
            x_start, y_start = i * stride_height, j * stride_width
            x_end, y_end = x_start + window_height, y_start + window_width
            if x_end > image_height:
                x_start, x_end = image_height - window_height, image_height
            if y_end > image_width:
                y_start, y_end = image_width - window_width, image_width

            pi_map[:, (x_start // block_size): (x_end // block_size), (y_start // block_size): (y_end // block_size)] += pi_maps[idx, :, :, :]
            lambda_map[:, (x_start // block_size): (x_end // block_size), (y_start // block_size): (y_end // block_size)] += lambda_maps[idx, :, :, :]
            count_map[:, (x_start // block_size): (x_end // block_size), (y_start // block_size): (y_end // block_size)] += 1.
            idx += 1

    # average the density map
    pi_map /= count_map
    lambda_map /= count_map
    
    # convert to Tensor and reshape
    pi_map = torch.from_numpy(pi_map).unsqueeze(0)  # shape: (1, 1, H // block_size, W // block_size)
    lambda_map = torch.from_numpy(lambda_map).unsqueeze(0)  # shape: (1, 1, H // block_size, W // block_size)
    return pi_map, lambda_map


# -----------------------------
# Inference function
# -----------------------------
@spaces.GPU(duration=120)
def predict(image: Image.Image, variant_dataset_metric: str):
    """
    Given an input image, preprocess it, run the model to obtain a density map,
    compute the total crowd count, and prepare the density map for display.
    """
    global loaded_model, current_model_config
    
    # 在GPU进程中定义device
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    # 如果选择的是分割线，返回错误信息
    if "━━━━━━" in variant_dataset_metric:
        return image, None, None, "⚠️ Please select a valid model configuration", None, None, None
    
    parts = variant_dataset_metric.split(" @ ")
    if len(parts) != 3:
        return image, None, None, "❌ Invalid model configuration format", None, None, None
        
    variant, dataset, metric = parts[0], parts[1], parts[2].lower()

    if dataset == "ShanghaiTech A":
        dataset_name = "sha"
    elif dataset == "ShanghaiTech B":
        dataset_name = "shb"
    elif dataset == "UCF-QNRF":
        dataset_name = "qnrf"
    elif dataset == "NWPU-Crowd":
        dataset_name = "nwpu"
    else:
        return image, None, None, f"❌ Unknown dataset: {dataset}", None, None, None

    # 在GPU进程中加载模型（如果需要）
    if (loaded_model is None or 
        current_model_config["variant"] != variant or 
        current_model_config["dataset"] != dataset_name or 
        current_model_config["metric"] != metric):
        
        print(f"Loading model in GPU process: {variant} @ {dataset} with {metric} metric")
        loaded_model = load_model(variant=variant, dataset=dataset_name, metric=metric)
        current_model_config = {"variant": variant, "dataset": dataset_name, "metric": metric}

    if not hasattr(loaded_model, "input_size"):
        if dataset_name == "sha":
            loaded_model.input_size = 224
        elif dataset_name == "shb":
            loaded_model.input_size = 448
        elif dataset_name == "qnrf":
            loaded_model.input_size = 672
        elif dataset_name == "nwpu":
            loaded_model.input_size = 672
    elif isinstance(loaded_model.input_size, (list, tuple)):
        loaded_model.input_size = loaded_model.input_size[0]  # Use the first element if it's a list or tuple
    else:
        assert isinstance(loaded_model.input_size, (int, float)), f"input_size must be an int or float, got {type(loaded_model.input_size)}"

    loaded_model.to(device)

    # Preprocess the image
    input_width, input_height = image.size
    image_tensor = transform(image, dataset_name).to(device)  # shape: (1, 3, H, W)

    input_size = loaded_model.input_size
    image_height, image_width = image_tensor.shape[-2:]
    aspect_ratio = image_width / image_height
    if image_height < input_size:
        new_height = input_size
        new_width = int(new_height * aspect_ratio)
        image_tensor = F.interpolate(image_tensor, size=(new_height, new_width), mode="bicubic", align_corners=False, antialias=True)
        image_height, image_width = new_height, new_width
    if image_width < input_size:
        new_width = input_size
        new_height = int(new_width / aspect_ratio)
        image_tensor = F.interpolate(image_tensor, size=(new_height, new_width), mode="bicubic", align_corners=False, antialias=True)
        image_height, image_width = new_height, new_width
    
    with torch.no_grad():
        if hasattr(loaded_model, "num_vpt") and loaded_model.num_vpt is not None and loaded_model.num_vpt > 0:  # For ViT models, use sliding window prediction
            # For ViT models with VPT
            pi_map, lambda_map = _sliding_window_predict(
                model=loaded_model,
                image=image_tensor,
                window_size=input_size,
                stride=input_size
            )
        
        elif hasattr(loaded_model, "pi_text_feats") and hasattr(loaded_model, "lambda_text_feats") and loaded_model.pi_text_feats is not None and loaded_model.lambda_text_feats is not None:  # For other CLIP-based models
            image_feats = loaded_model.backbone(image_tensor)
            # image_feats = F.normalize(image_feats.permute(0, 2, 3, 1), p=2, dim=-1)  # shape (B, H, W, C)
            pi_image_feats, lambda_image_feats = loaded_model.pi_head(image_feats), loaded_model.lambda_head(image_feats)
            pi_image_feats = F.normalize(pi_image_feats.permute(0, 2, 3, 1), p=2, dim=-1)  # shape (B, H, W, C)
            lambda_image_feats = F.normalize(lambda_image_feats.permute(0, 2, 3, 1), p=2, dim=-1)  # shape (B, H, W, C)

            pi_text_feats, lambda_text_feats = loaded_model.pi_text_feats, loaded_model.lambda_text_feats
            pi_logit_scale, lambda_logit_scale = loaded_model.pi_logit_scale.exp(), loaded_model.lambda_logit_scale.exp()

            pi_logit_map = pi_logit_scale * pi_image_feats @ pi_text_feats.t()  # (B, H, W, 2), logits per image
            lambda_logit_map = lambda_logit_scale * lambda_image_feats @ lambda_text_feats.t()  # (B, H, W, N - 1), logits per image

            pi_logit_map =  pi_logit_map.permute(0, 3, 1, 2)  # (B, 2, H, W)
            lambda_logit_map = lambda_logit_map.permute(0, 3, 1, 2)  # (B, N - 1, H, W)

            lambda_map = (lambda_logit_map.softmax(dim=1) * loaded_model.bin_centers[:, 1:]).sum(dim=1, keepdim=True)  # (B, 1, H, W)
            pi_map = pi_logit_map.softmax(dim=1)[:, 0:1]  # (B, 1, H, W)
        
        else: # For non-CLIP models
            x = loaded_model.backbone(image_tensor)
            logit_pi_map = loaded_model.pi_head(x)  # shape: (B, 2, H, W)
            logit_map = loaded_model.bin_head(x)  # shape: (B, C, H, W)
            lambda_map= (logit_map.softmax(dim=1) * loaded_model.bin_centers[:, 1:]).sum(dim=1, keepdim=True)  # shape: (B, 1, H, W)
            pi_map = logit_pi_map.softmax(dim=1)[:, 0:1]  # shape: (B, 1, H, W)

        
        den_map = (1.0 - pi_map) * lambda_map  # shape: (B, 1, H, W)
        count = den_map.sum().item()

        strucrual_zero_map = F.interpolate(
            pi_map, size=(input_height, input_width), mode="bilinear", align_corners=False, antialias=True
        ).cpu().squeeze().numpy()

        lambda_map = F.interpolate(
            lambda_map, size=(input_height, input_width), mode="bilinear", align_corners=False, antialias=True
        ).cpu().squeeze().numpy()

        den_map = F.interpolate(
            den_map, size=(input_height, input_width), mode="bilinear", align_corners=False, antialias=True
        ).cpu().squeeze().numpy()
    
    sampling_zero_map = (1.0 - strucrual_zero_map) * np.exp(-lambda_map)
    complete_zero_map = strucrual_zero_map + sampling_zero_map

    # Normalize maps for display purposes
    def normalize_map(x: np.ndarray) -> np.ndarray:
        """ Normalize the map to [0, 1] range for visualization. """
        x_min = np.min(x)
        x_max = np.max(x)
        if x_max - x_min < EPS:
            return np.zeros_like(x)
        return (x - x_min) / (x_max - x_min + EPS)
    
    # strucrual_zero_map = normalize_map(strucrual_zero_map)
    # sampling_zero_map = normalize_map(sampling_zero_map)
    # lambda_map = normalize_map(lambda_map)
    # den_map = normalize_map(den_map)
    # complete_zero_map = normalize_map(complete_zero_map)
    
    # Apply a colormap for better visualization
    # Options: 'viridis', 'plasma', 'hot', 'inferno', 'jet' (recommended)
    colormap = cm.get_cmap("jet")

    # The colormap returns values in [0,1]. Scale to [0,255] and convert to uint8.
    den_map = (colormap(den_map) * 255).astype(np.uint8)
    strucrual_zero_map = (colormap(strucrual_zero_map) * 255).astype(np.uint8)
    sampling_zero_map = (colormap(sampling_zero_map) * 255).astype(np.uint8)
    lambda_map = (colormap(lambda_map) * 255).astype(np.uint8)
    complete_zero_map = (colormap(complete_zero_map) * 255).astype(np.uint8)

    # Convert to PIL images
    den_map = Image.fromarray(den_map).convert("RGBA")
    strucrual_zero_map = Image.fromarray(strucrual_zero_map).convert("RGBA")
    sampling_zero_map = Image.fromarray(sampling_zero_map).convert("RGBA")
    lambda_map = Image.fromarray(lambda_map).convert("RGBA")
    complete_zero_map = Image.fromarray(complete_zero_map).convert("RGBA")
    
    # Ensure the original image is in RGBA format.
    image_rgba = image.convert("RGBA")

    den_map = Image.blend(image_rgba, den_map, alpha=alpha)
    strucrual_zero_map = Image.blend(image_rgba, strucrual_zero_map, alpha=alpha)
    sampling_zero_map = Image.blend(image_rgba, sampling_zero_map, alpha=alpha)
    lambda_map = Image.blend(image_rgba, lambda_map, alpha=alpha)
    complete_zero_map = Image.blend(image_rgba, complete_zero_map, alpha=alpha)
    
    # 格式化计数显示
    count_display = f"👥 {round(count, 2)} people detected"
    if count < 1:
        count_display = "👤 Less than 1 person detected"
    elif count == 1:
        count_display = "👤 1 person detected"
    elif count < 10:
        count_display = f"👥 {round(count, 1)} people detected"
    else:
        count_display = f"👥 {round(count)} people detected"
    
    return image, den_map, lambda_map, count_display, strucrual_zero_map, sampling_zero_map, complete_zero_map


# -----------------------------
# Build Gradio Interface using Blocks for a two-column layout
# -----------------------------
css = """
/* 分割线样式 - 灰色不可选择 */
.dropdown select option[value*="━━━━━━"] {
    color: #999 !important;
    background-color: #f0f0f0 !important;
    font-style: italic !important;
    text-align: center !important;
    pointer-events: none !important;
    cursor: not-allowed !important;
    border: none !important;
}

/* Gradio下拉菜单中的分割线样式 */
.gr-dropdown .choices__item[data-value*="━━━━━━"] {
    color: #999 !important;
    background-color: #f0f0f0 !important;
    font-style: italic !important;
    text-align: center !important;
    pointer-events: none !important;
    cursor: not-allowed !important;
    user-select: none !important;
    opacity: 0.6 !important;
}

/* 悬停时保持灰色 */
.gr-dropdown .choices__item[data-value*="━━━━━━"]:hover {
    background-color: #f0f0f0 !important;
    color: #999 !important;
    cursor: not-allowed !important;
}

/* 通用的分割线样式 */
option:disabled {
    color: #999 !important;
    background-color: #f0f0f0 !important;
    font-style: italic !important;
}

/* 为包含分割线字符的选项添加样式 */
option[value*="━━━━━━"], 
select option[value*="━━━━━━"] {
    color: #999 !important;
    background-color: #f0f0f0 !important;
    cursor: not-allowed !important;
    pointer-events: none !important;
    text-align: center !important;
    opacity: 0.6 !important;
}

/* 整体主题美化 */
.gradio-container {
    max-width: 1600px !important;
    margin: 0 auto !important;
    font-family: 'Inter', -apple-system, BlinkMacSystemFont, sans-serif !important;
    background: linear-gradient(135deg, #f5f7fa 0%, #c3cfe2 100%) !important;
    min-height: 100vh !important;
    padding: 20px !important;
}

/* 响应式布局 - 自动调整列宽 */
@media (max-width: 1400px) {
    .gradio-container {
        max-width: 1200px !important;
        padding: 18px !important;
    }
}

@media (max-width: 1200px) {
    .gradio-container {
        max-width: 100% !important;
        padding: 16px !important;
    }
    
    /* 在中等屏幕上，将第二行改为垂直布局 */
    .gr-row:nth-of-type(2) {
        flex-direction: column !important;
    }
    
    .gr-row:nth-of-type(2) .gr-column {
        width: 100% !important;
        margin-bottom: 20px !important;
    }
    
    /* 重置中等屏幕上的组件高度 */
    .gr-row:nth-of-type(2) .gr-group,
    .gr-row:nth-of-type(3) .gr-group {
        height: auto !important;
        min-height: auto !important;
        position: static !important;
    }
    
    .gr-row:nth-of-type(3) .gr-column:first-child .gr-group {
        position: static !important;
    }
    
    .gr-row:nth-of-type(3) .gr-column:first-child .gr-button {
        position: static !important;
        bottom: auto !important;
        left: auto !important;
        right: auto !important;
        width: auto !important;
        margin-top: 16px !important;
    }
}

@media (max-width: 900px) {
    /* 在小屏幕上，将第三行也改为垂直布局 */
    .gr-row:nth-of-type(3) {
        flex-direction: column !important;
    }
    
    .gr-row:nth-of-type(3) .gr-column {
        width: 100% !important;
        margin-bottom: 20px !important;
    }
    
    /* Zero Analysis 在小屏幕上也改为垂直布局 */
    .gr-group .gr-row {
        flex-direction: column !important;
    }
    
    .gr-group .gr-row .gr-column {
        width: 100% !important;
        margin-bottom: 16px !important;
    }
    
    /* 重置小屏幕上的组件高度 */
    .gr-row:nth-of-type(2) .gr-group,
    .gr-row:nth-of-type(3) .gr-group {
        height: auto !important;
        min-height: auto !important;
        position: static !important;
    }
    
    .gr-row:nth-of-type(3) .gr-column:first-child .gr-group {
        position: static !important;
    }
    
    .gr-row:nth-of-type(3) .gr-column:first-child .gr-button {
        position: static !important;
        bottom: auto !important;
        left: auto !important;
        right: auto !important;
        width: auto !important;
        margin-top: 16px !important;
    }
}

@media (max-width: 768px) {
    .gradio-container {
        padding: 12px !important;
    }
    
    .gr-column {
        margin-bottom: 16px !important;
        padding: 0 4px !important;
    }
    
    .gr-markdown h1 {
        font-size: 2rem !important;
    }
    
    .gr-group {
        padding: 16px !important;
    }
    
    .gr-button {
        padding: 12px 24px !important;
        font-size: 1rem !important;
    }
    
    /* 图像高度在小屏幕上调整 */
    .gr-image {
        height: 300px !important;
    }
    
    .zero-analysis-image {
        height: 300px !important;
    }
    
    /* 小屏幕上重置组件高度 */
    .gr-row:nth-of-type(2) .gr-group,
    .gr-row:nth-of-type(3) .gr-group {
        height: auto !important;
        min-height: auto !important;
        position: static !important;
    }
    
    .gr-row:nth-of-type(3) .gr-column:first-child .gr-group {
        position: static !important;
    }
    
    .gr-row:nth-of-type(3) .gr-column:first-child .gr-button {
        position: static !important;
        bottom: auto !important;
        left: auto !important;
        right: auto !important;
        width: auto !important;
        margin-top: 16px !important;
    }
    
    .gr-row:nth-of-type(2) .gr-textbox,
    .gr-row:nth-of-type(2) .gr-dropdown {
        min-height: auto !important;
        max-height: none !important;
    }
    
    .gr-row:nth-of-type(3) .gr-image {
        height: 300px !important;
        min-height: auto !important;
        max-height: none !important;
        flex: none !important;
    }
}

/* 超宽屏幕优化 */
@media (min-width: 1600px) {
    .gradio-container {
        max-width: 1800px !important;
        padding: 24px !important;
    }
    
    .gr-image {
        height: 450px !important;
    }
    
    .zero-analysis-image {
        height: 450px !important;
    }
}

/* 标题样式 */
.gr-markdown h1 {
    text-align: center !important;
    color: #2563eb !important;
    font-weight: 700 !important;
    font-size: 3rem !important;
    margin-bottom: 0.5rem !important;
    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
    -webkit-background-clip: text !important;
    -webkit-text-fill-color: transparent !important;
    text-shadow: 0 4px 8px rgba(0,0,0,0.1) !important;
}

/* 副标题样式 */
.gr-markdown p {
    text-align: center !important;
    color: #6b7280 !important;
    font-size: 1.2rem !important;
    margin-bottom: 2rem !important;
    font-weight: 500 !important;
}

/* 主要布局组美化 */
.gr-group {
    background: rgba(255, 255, 255, 0.9) !important;
    backdrop-filter: blur(10px) !important;
    border-radius: 20px !important;
    padding: 24px !important;
    margin: 16px 0 !important;
    box-shadow: 0 8px 32px rgba(0, 0, 0, 0.1) !important;
    border: 1px solid rgba(255, 255, 255, 0.2) !important;
    transition: all 0.3s ease !important;
}

.gr-group:hover {
    transform: translateY(-4px) !important;
    box-shadow: 0 12px 40px rgba(0, 0, 0, 0.15) !important;
}

/* 按钮美化 */
.gr-button {
    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
    border: none !important;
    border-radius: 12px !important;
    color: white !important;
    font-weight: 600 !important;
    font-size: 1.1rem !important;
    padding: 16px 32px !important;
    transition: all 0.3s ease !important;
    box-shadow: 0 6px 20px rgba(102, 126, 234, 0.3) !important;
    text-transform: uppercase !important;
    letter-spacing: 0.5px !important;
}

.gr-button:hover {
    transform: translateY(-3px) !important;
    box-shadow: 0 10px 30px rgba(102, 126, 234, 0.4) !important;
    background: linear-gradient(135deg, #5a67d8 0%, #6b46c1 100%) !important;
}

/* 输入框样式 */
.gr-textbox, .gr-dropdown {
    border-radius: 12px !important;
    border: 2px solid #e5e7eb !important;
    transition: all 0.3s ease !important;
    background: rgba(255, 255, 255, 0.8) !important;
    font-size: 1rem !important;
    padding: 12px 16px !important;
}

.gr-textbox:focus, .gr-dropdown:focus {
    border-color: #667eea !important;
    box-shadow: 0 0 0 4px rgba(102, 126, 234, 0.1) !important;
    background: rgba(255, 255, 255, 1) !important;
}

/* 图像容器美化 - 统一尺寸 */
.gr-image {
    border-radius: 16px !important;
    overflow: hidden !important;
    box-shadow: 0 8px 25px rgba(0, 0, 0, 0.15) !important;
    transition: all 0.3s ease !important;
    background: white !important;
    height: 400px !important;
    width: 100% !important;
}

.gr-image:hover {
    box-shadow: 0 15px 35px rgba(0, 0, 0, 0.2) !important;
    transform: translateY(-2px) !important;
}

/* 确保第二行组件等高 - 重新调整为三列等宽 */
.gr-row:nth-of-type(2) .gr-group {
    min-height: 180px !important;
    display: flex !important;
    flex-direction: column !important;
    justify-content: center !important;
}

.gr-row:nth-of-type(2) .gr-group > * {
    flex: 1 !important;
}

/* 确保第二行的文本框具有相同的高度 */
.gr-row:nth-of-type(2) .gr-textbox {
    min-height: 100px !important;
    max-height: 100px !important;
    display: flex !important;
    align-items: center !important;
}

/* 确保第二行下拉菜单区域等高 */
.gr-row:nth-of-type(2) .gr-dropdown {
    min-height: 80px !important;
}

/* 确保下拉框可以正常展开 */
.gr-dropdown {
    position: relative !important;
    z-index: 1000 !important;
}

.gr-dropdown .choices {
    position: absolute !important;
    z-index: 1001 !important;
    width: 100% !important;
    max-height: 300px !important;
    overflow-y: auto !important;
}

/* 确保第三行组件等高 - 重新调整高度 */
.gr-row:nth-of-type(3) .gr-group {
    height: 520px !important;
    min-height: 520px !important;
    display: flex !important;
    flex-direction: column !important;
    justify-content: flex-start !important;
}

/* 第三行的图像容器统一高度 */
.gr-row:nth-of-type(3) .gr-image {
    height: 400px !important;
    min-height: 400px !important;
    max-height: 400px !important;
    flex: 0 0 400px !important;
}

/* 第三行的按钮固定在底部 - 只对第一列（Image Input）应用 */
.gr-row:nth-of-type(3) .gr-column:first-child .gr-group {
    position: relative !important;
}

.gr-row:nth-of-type(3) .gr-column:first-child .gr-button {
    position: absolute !important;
    bottom: 24px !important;
    left: 24px !important;
    right: 24px !important;
    margin: 0 !important;
    width: calc(100% - 48px) !important;
}

/* 列间距优化 */
.gr-column {
    padding: 0 8px !important;
    margin-bottom: 16px !important;
}

/* 第二行特殊布局调整 - 移除，现在是三列等宽 */
.gr-row:nth-of-type(2) .gr-column {
    padding: 0 8px !important;
}

/* 标签美化 */
.gr-label {
    font-weight: 700 !important;
    color: #374151 !important;
    margin-bottom: 12px !important;
    font-size: 1.1rem !important;
    text-transform: uppercase !important;
    letter-spacing: 0.5px !important;
}

/* 模型状态框特殊样式 */
.gr-textbox[data-testid*="model-status"] {
    background: linear-gradient(135deg, #ecfdf5 0%, #d1fae5 100%) !important;
    font-family: 'Monaco', 'Menlo', monospace !important;
    font-size: 0.95rem !important;
    font-weight: 600 !important;
    border: 2px solid #10b981 !important;
}

/* Zero Analysis 特殊布局 */
.gr-row:has(.gr-image[label*="Zero"]) {
    background: linear-gradient(135deg, rgba(255,255,255,0.95) 0%, rgba(248,250,252,0.95) 100%) !important;
    border-radius: 20px !important;
    padding: 24px !important;
    margin: 20px 0 !important;
    box-shadow: 0 10px 30px rgba(0, 0, 0, 0.1) !important;
}

/* Zero Analysis 图像特殊样式 - 统一尺寸 */
.zero-analysis-image {
    border: 3px solid transparent !important;
    background: linear-gradient(white, white) padding-box, 
                linear-gradient(135deg, #667eea, #764ba2) border-box !important;
    border-radius: 16px !important;
    transition: all 0.3s ease !important;
    height: 400px !important;
    width: 100% !important;
}

.zero-analysis-image:hover {
    transform: scale(1.02) !important;
    box-shadow: 0 12px 35px rgba(102, 126, 234, 0.2) !important;
}

/* 确保所有行的组件等高 */
.gr-row .gr-group {
    min-height: 100% !important;
    display: flex !important;
    flex-direction: column !important;
}

.gr-row .gr-column {
    height: 100% !important;
}

/* 第二行内部子行等高处理 - 移除，现在不需要内部子行 */

/* 统计信息卡片美化 */
.gr-textbox[label*="Count"] {
    background: linear-gradient(135deg, #ecfcff 0%, #cffafe 100%) !important;
    border: 2px solid #06b6d4 !important;
    font-size: 1.2rem !important;
    font-weight: 700 !important;
    text-align: center !important;
    color: #0e7490 !important;
}

/* 示例区域美化 */
.gr-examples {
    background: linear-gradient(135deg, rgba(255,255,255,0.9) 0%, rgba(248,250,252,0.9) 100%) !important;
    backdrop-filter: blur(10px) !important;
    border-radius: 20px !important;
    padding: 30px !important;
    margin-top: 30px !important;
    border: 1px solid rgba(255, 255, 255, 0.2) !important;
    box-shadow: 0 10px 30px rgba(0, 0, 0, 0.1) !important;
}

/* Accordion 美化 */
.gr-accordion {
    background: rgba(255, 255, 255, 0.8) !important;
    border-radius: 16px !important;
    margin: 16px 0 !important;
    border: 1px solid rgba(255, 255, 255, 0.2) !important;
    box-shadow: 0 6px 20px rgba(0, 0, 0, 0.08) !important;
}

/* 响应式设计 - 移除旧的媒体查询，已在上方重新定义 */

/* 加载动画 */
@keyframes pulse {
    0%, 100% { opacity: 1; }
    50% { opacity: 0.5; }
}

.gr-loading .gr-image {
    animation: pulse 2s cubic-bezier(0.4, 0, 0.6, 1) infinite !important;
}

/* 成功状态指示 */
.status-success {
    color: #059669 !important;
    background-color: #d1fae5 !important;
    border: 1px solid #a7f3d0 !important;
}

/* 错误状态指示 */
.status-error {
    color: #dc2626 !important;
    background-color: #fee2e2 !important;
    border: 1px solid #fecaca !important;
}
"""

with gr.Blocks(css=css, theme=gr.themes.Soft(), title="ZIP Crowd Counting") as demo:
    gr.Markdown("""
    # 🎯 Crowd Counting by ZIP
    ### Upload an image and get precise crowd density predictions with ZIP models!
    """)
    
    # 添加信息面板
    with gr.Accordion("ℹ️ About ZIP", open=False):
        gr.Markdown("""
        **ZIP (Zero-Inflated Poisson)** is a framework designed for crowd counting, a task where the goal is to estimate how many people are present in an image. It was introduced in the paper [ZIP: Scalable Crowd Counting via Zero-Inflated Poisson Modeling](https://arxiv.org/abs/2506.19955).
        ZIP is based on a simple idea: not all empty areas in an image mean the same thing. Some regions are empty because there are truly no people there (like walls or sky), while others are places where people could appear but just happen not to in this particular image. ZIP separates these two cases using two prediction heads:
        - **Structural Zeros**: These are regions that naturally never contain people (e.g., the background or torso areas). These are handled by the π head.
        - **Sampling Zeros**: These are regions where people could appear but don't in this image. These are modeled by the λ head.
        
        By separating *where* people are likely to be from *how many* are present, ZIP produces more accurate and interpretable crowd estimates, especially in scenes with large empty spaces or varied crowd densities.
        
        Choose from different model variants: **ZIP-B** (Base), **ZIP-S** (Small), **ZIP-T** (Tiny), **ZIP-N** (Nano), **ZIP-P** (Pico)
        """)

    # 第二行：模型配置、状态和预测结果（三列等宽）
    with gr.Row():
        with gr.Column(scale=1):
            with gr.Group():
                model_dropdown = gr.Dropdown(
                    choices=pretrained_models,
                    value="ZIP-B @ NWPU-Crowd @ MAE",
                    label="🎛️ Select Model & Dataset",
                    info="Choose model variant, dataset, and evaluation metric"
                )
        
        with gr.Column(scale=1):
            with gr.Group():
                model_status = gr.Textbox(
                    label="📊 Model Status",
                    value="🔄 No model loaded",
                    interactive=False,
                    elem_classes=["status-display"],
                    lines=3
                )
        
        with gr.Column(scale=1):
            with gr.Group():
                output_text = gr.Textbox(
                    label="🧙 Predicted Count",
                    value="",
                    interactive=False,
                    info="Total number of people detected",
                    lines=3
                )

    # 第三行：主要图像（输入图像、密度图、Lambda图）
    with gr.Row():
        with gr.Column(scale=1):
            with gr.Group():
                input_img = gr.Image(
                    label="📸 Upload Image", 
                    sources=["upload", "clipboard"], 
                    type="pil",
                    height=400
                )
                submit_btn = gr.Button(
                    "🚀 Analyze Crowd", 
                    variant="primary",
                    size="lg"
                )
        
        with gr.Column(scale=1):
            with gr.Group():
                output_den_map = gr.Image(
                    label="🎯 Predicted Density Map", 
                    type="pil", 
                    height=400
                )
        
        with gr.Column(scale=1):
            with gr.Group():
                output_lambda_map = gr.Image(
                    label="📈 Lambda Map", 
                    type="pil", 
                    height=400
                )
    
    # 第四行：Zero Analysis - 全宽，内部三列等宽
    with gr.Group():
        gr.Markdown("### 🔍 Zero Analysis")
        gr.Markdown("*Explore different types of zero predictions in crowd analysis*")
        with gr.Row():
            with gr.Column(scale=1):
                output_structural_zero_map = gr.Image(
                    label="🏗️ Structural Zero Map", 
                    type="pil", 
                    height=400,
                    elem_classes=["zero-analysis-image"]
                )
            
            with gr.Column(scale=1):
                output_sampling_zero_map = gr.Image(
                    label="📊 Sampling Zero Map", 
                    type="pil", 
                    height=400,
                    elem_classes=["zero-analysis-image"]
                )
            
            with gr.Column(scale=1):
                output_complete_zero_map = gr.Image(
                    label="👺 Complete Zero Map", 
                    type="pil", 
                    height=400,
                    elem_classes=["zero-analysis-image"]
                )

    # 当模型变化时，自动更新模型
    def on_model_change(variant_dataset_metric):
        # 如果选择的是分割线，保持当前选择不变
        if "━━━━━━" in variant_dataset_metric:
            return "⚠️ Please select a valid model configuration"
        result = update_model_if_needed(variant_dataset_metric)
        if "Model loaded:" in result:
            return f"✅ {result}"
        elif "Model already loaded:" in result:
            return f"🔄 {result}"
        else:
            return f"❌ {result}"
    
    model_dropdown.change(
        fn=on_model_change,
        inputs=[model_dropdown],
        outputs=[model_status]
    )

    # 页面加载时设置默认模型配置（不在主进程中加载模型）
    demo.load(
        fn=lambda: f"✅ {update_model_if_needed('ZIP-B @ NWPU-Crowd @ MAE')}",
        outputs=[model_status]
    )

    submit_btn.click(
        fn=predict,
        inputs=[input_img, model_dropdown],
        outputs=[input_img, output_den_map, output_lambda_map, output_text, output_structural_zero_map, output_sampling_zero_map, output_complete_zero_map]
    )

    # 美化示例区域
    with gr.Accordion("🖼️ Try Example Images", open=True):
        gr.Markdown("**Click on any example below to test the model:**")
        gr.Examples(
            examples=[
                ["example1.jpg"], ["example2.jpg"], ["example3.jpg"], ["example4.jpg"],
                ["example5.jpg"], ["example6.jpg"], ["example7.jpg"], ["example8.jpg"],
                ["example9.jpg"], ["example10.jpg"], ["example11.jpg"], ["example12.jpg"]
            ],
            inputs=input_img,
            label="📚 Example Gallery",
            examples_per_page=12
        )
    
    # 添加使用说明
    with gr.Accordion("📖 How to Use", open=False):
        gr.Markdown("""
        ### Step-by-step Guide:
        
        1. **🎛️ Select Model**: Choose your preferred model variant, pre-trained dataset, and evaluation metric from the dropdown
        2. **📸 Upload Image**: Click the image area to upload your crowd photo or use clipboard
        3. **🚀 Analyze**: Click the "Analyze Crowd" button to start processing
        4. **📊 View Results**: Examine the density maps and crowd count in the output panels
        
        ### Understanding the Outputs:
        
        **📊 Main Results:**
        - **🎯 Density Map**: Shows where people are located with color intensity, modeled by (1-π) * λ
        - **🧙 Predicted Count**: Total number of people detected in the image
        
        **🔍 Zero Analysis:**
        - **🏗️ Structural Zero Map**: Indicates regions that structurally cannot contain head annotations (e.g., walls, sky, torso, or background). These are governed by the π head, which estimates the probability that a region never contains people.
        - **📊 Sampling Zero Map**: Shows areas where people could be present but happen not to appear in the current image. These zeros are modeled by (1-π) * exp(-λ), where the expected count λ is near zero.
        - **👺 Complete Zero Map**: A combined visualization of zero probabilities, capturing both structural and sampling zeros. This map reflects overall non-crowd likelihood per region.
        
        **🔥 Hotspots:**
        - **📈 Lambda Map**: Highlights areas with high expected crowd density. Each value represents the expected number of people in that region, modeled by the Poisson intensity (λ). This map focuses on *how many* people are likely to be present, **WITHOUT** assuming people could appear there. ⚠️ Lambda Map **NEEDS** to be combined with Structural Zero Map by (1-π) * λ to produce the final density map.
        """)
    
    # 添加技术信息
    with gr.Accordion("🔬 Technical Details", open=False):
        gr.Markdown("""
        ### Model Variants:
        - **ZIP-B**: Base model with best performance
        - **ZIP-S**: Small model for faster inference
        - **ZIP-T**: Tiny model for resource-constrained environments
        - **ZIP-N**: Nano model for mobile applications
        - **ZIP-P**: Pico model for edge devices
        
        ### Datasets:
        - **ShanghaiTech A**: Dense, low-resolution crowd scenes
        - **ShanghaiTech B**: Sparse, high-resolution crowd scenes
        - **UCF-QNRF**: Dense, ultra high-resolution crowd images
        - **NWPU-Crowd**: Largest ultra high-resolution crowd counting dataset
        
        ### Metrics:
        - **MAE**: Mean Absolute Error - average counting error
        - **NAE**: Normalized Absolute Error - relative counting error
        """)

demo.launch(
    server_name="0.0.0.0",
    server_port=7860,
    show_api=False,
    share=False
)