# 四种提示类型示例
prompts = {
    'appearance': "light brown hair girl with blue and white striped shirt",
    'position': "person at point [0.52, 0.48]",
    'action': "person holding a phone",
    'pose': "person facing left"
}

"""
GazeAnywhere 推理示例

环境: Python 3.12+, PyTorch 2.7+, CUDA 12.6+
"""

# 安装
# pip3 install torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
# pip3 install -r requirements.txt
# pip3 install "git+https://github.com/facebookresearch/detectron2.git@017abbfa5f2c2a2afa045200c2af9ccf2fc6227f#egg=detectron2" --no-build-isolation

# 推理命令
"""
python tools/inference.py \
    --config_file configs/gazeanywhere_config.py \
    --model_weights /path/to/GazeAnywhere/model.safetensors \
    --image_path assets/example.jpg \
    --text "light brown hair girl with blue and white striped shirt" \
    --save_path visualization.jpg \
    --use_dark_inference
"""

# Python API 调用
import torch
from gazeanywhere.config import get_cfg
from gazeanywhere.model import GazeAnywhereModel
from gazeanywhere.utils import load_image, visualize_gaze

# 加载模型
cfg = get_cfg()
cfg.merge_from_file("configs/gazeanywhere_config.py")
model = GazeAnywhereModel(cfg)
model.load_weights("path/to/model.safetensors")
model.eval().cuda()

# 推理
image = load_image("driver_cabin.jpg")
prompt = "driver in the seat"

with torch.no_grad():
    output = model.inference(image, prompt)
    # output包含:
    # - person_location: 人物位置
    # - in_frame_prob: 框内/框外概率
    # - gaze_heatmap: 64×64 视线热图

visualize_gaze(image, output, save_path="gaze_result.jpg")

"""
AnyGaze Agent: 视线估计 + LLM 推理

示例交互:
Q: "驾驶员在看哪里？是否分心？"
A: "驾驶员视线偏向右侧中控区域，持续3秒，可能正在操作中控屏幕，
    属于视觉分心。建议触发一级警告。"
"""

# AnyGaze Agent 集成示例
class AnyGazeAgent:
    def __init__(self, gaze_model, llm_client):
        self.gaze = gaze_model
        self.llm = llm_client  # Gemini API
        
    def analyze(self, image, question):
        # 1. 视线估计
        gaze_result = self.gaze.inference(image, "driver")
        
        # 2. 提取视线信息
        gaze_info = {
            'in_frame': gaze_result['in_frame_prob'],
            'heatmap_peak': gaze_result['heatmap'].max(),
            'peak_location': gaze_result['heatmap'].argmax(),
        }
        
        # 3. LLM 推理
        context = f"""
        驾驶员视线分析:
        - 框内概率: {gaze_info['in_frame']:.2f}
        - 视线焦点: ({gaze_info['peak_location'] % 64}, {gaze_info['peak_location'] // 64})
        - 热图峰值: {gaze_info['heatmap_peak']:.3f}
        """
        
        answer = self.llm.generate(question, context)
        return answer