demo_test_all_v4_resume.py

import tensorflow as tf
from android_env.proto.a11y import android_accessibility_forest_pb2
import json
import os
from google.protobuf.json_format import MessageToJson
from tqdm import tqdm


def parse_node(node, nodes, depth=0, parent_bounds=None):
    # 提取节点信息
    class_name = node.get("className", "Unknown")
    # 获取 viewIdResourceName 并简化输出，只保留资源名部分
    view_id = node.get("viewIdResourceName", "")
    simplified_view_id = view_id.split(":")[-1] if view_id else ""
    simplified_view_id = ' '.join(simplified_view_id.split('_'))
    simplified_view_id = simplified_view_id.replace('id/','')

    bounds = node.get("boundsInScreen", {})
    bounds_str = f"[{bounds.get('left', 0)}, {bounds.get('top', 0)}, {bounds.get('right', 0)}, {bounds.get('bottom', 0)}]"

    if parent_bounds and bounds == parent_bounds:
        bounds_str = ""

    node_str = f"{'  ' * depth}└── {simplified_view_id} {f'({bounds_str})' if bounds_str else ''}"

    children_str = ""
    if "childIds" in node:
        children = [child for child in nodes if 'uniqueId' in child and child['uniqueId'] in node['childIds']]
        for child in children:
            children_str += "\n" + parse_node(child, nodes, depth + 1, bounds)

    return node_str + children_str


def parse_tree(window_data):
    # 获取树的根节点（Window信息）
    window_info = window_data['windows'][0]
    window_type = window_info.get('windowType', 'Unknown')
    root_node_str = f"Window ({window_type})"
    nodes = window_info['tree']['nodes']
    root_node_str += "\n" + parse_node(nodes[0], nodes, 1)
    return root_node_str



# 设置路径
output_dir = "output_data"
os.makedirs(output_dir, exist_ok=True)

intermediate_json_path = os.path.join("/inspire/hdd/ws-ba572160-47f8-4ca1-984e-d6bcdeb95dbb/a100-maybe/wangbaode/NIPS_2025/Agent/Dataset/android_env/intermediate_data_8k_succ.json")
intermediate_json_path_new = os.path.join("/inspire/hdd/ws-ba572160-47f8-4ca1-984e-d6bcdeb95dbb/a100-maybe/wangbaode/NIPS_2025/Agent/Dataset/android_env/intermediate_data.jsonl")



# === 恢复已处理的 episode_id 列表 ===
processed_episodes = set()
all_data = []

if os.path.exists(intermediate_json_path):
    with open(intermediate_json_path, "r", encoding="utf-8") as f:
        try:
            all_data = json.load(f)
            for record in all_data:
                processed_episodes.add(record["episode_id"])
            print(f"恢复 {len(processed_episodes)} 条已处理记录")
        except Exception as e:
            print(f"读取中间文件失败: {e}")

# === 加载 TFRecord 数据 ===
filenames = tf.io.gfile.glob('/inspire/hdd/ws-ba572160-47f8-4ca1-984e-d6bcdeb95dbb/a100-maybe/wangbaode/NIPS_2025/Agent/Dataset/AndControl/android_control*')
raw_dataset = tf.data.TFRecordDataset(filenames, compression_type='GZIP')

# === 遍历数据集 ===
record_index = len(all_data)  # 从当前进度继续

for raw_record in tqdm(raw_dataset, desc="遍历 TFRecords"):
    try:
        example = tf.train.Example.FromString(raw_record.numpy())

        episode_id = example.features.feature['episode_id'].int64_list.value[0]
        if episode_id in processed_episodes:
            continue  # 已处理过，跳过

        goal = example.features.feature['goal'].bytes_list.value[0].decode('utf-8')
        screenshots = example.features.feature['screenshots'].bytes_list.value
        accessibility_trees = [
            android_accessibility_forest_pb2.AndroidAccessibilityForest().FromString(tree)
            for tree in example.features.feature['accessibility_trees'].bytes_list.value
        ]
        screenshot_widths = list(example.features.feature['screenshot_widths'].int64_list.value)
        screenshot_heights = list(example.features.feature['screenshot_heights'].int64_list.value)
        actions = [a.decode('utf-8') for a in example.features.feature['actions'].bytes_list.value]
        step_instructions = [s.decode('utf-8') for s in example.features.feature['step_instructions'].bytes_list.value]

        # 构造数据
        data = {
            "episode_id": episode_id,
            "goal": goal,
            "screenshots": [],
            "screenshot_widths": screenshot_widths,
            "screenshot_heights": screenshot_heights,
            "actions": actions,
            "step_instructions": step_instructions,
            "accessibility_trees": []
        }

        # 保存截图
        for i, img in enumerate(screenshots):
            path = os.path.join(output_dir, f"screenshot_{episode_id}_{i}.png")
            with open(path, "wb") as f:
                f.write(img)
            data["screenshots"].append(path)

        # 转换可达性树
        for tree in accessibility_trees:
            json_dict = json.loads(MessageToJson(tree))
            tree_output = parse_tree(json_dict)
            data["accessibility_trees"].append(tree_output)

        # 加入列表
        all_data.append(data)
        processed_episodes.add(episode_id)

        # 覆盖写入中间 JSON 文件（可断点续跑）
        # with open(intermediate_json_path_new, "a+", encoding="utf-8") as f:
        #     json.dump(all_data, f, indent=2, ensure_ascii=False)

        # 逐条追加写入 JSONL
        with open(intermediate_json_path_new, "a+", encoding="utf-8") as fout:
            fout.write(json.dumps(data, ensure_ascii=False) + "\n")

        print(f"✅ 已处理 episode_id={episode_id}")

    except Exception as e:
        print(f"❌ 处理失败，跳过该记录：{e}")
        continue