import gradio as gr import os import random import time from datetime import datetime from functools import partial import json import io from huggingface_hub import HfApi from huggingface_hub.hf_api import HfHubHTTPError import traceback from itertools import combinations, product # ==== 全局配置 ==== # ---- 测试模式开关 ---- REPEAT_SINGLE_TARGET_FOR_TESTING = False # 设置为 True 以启用“重复单一目标图”测试模式 NUM_REPEATED_TRIALS_FOR_TESTING = 5 # 在该测试模式下，单个目标图片重复的次数 (原为20，改为5方便测试) # ---- 常规配置 ---- BASE_IMAGE_DIR = "/data/images/images" TARGET_DIR_BASENAME = "gt" TARGET_DIR = os.path.join(BASE_IMAGE_DIR, TARGET_DIR_BASENAME) METHOD_ROOTS = [] if os.path.exists(BASE_IMAGE_DIR): try: METHOD_ROOTS = [ os.path.join(BASE_IMAGE_DIR, d) for d in os.listdir(BASE_IMAGE_DIR) if os.path.isdir(os.path.join(BASE_IMAGE_DIR, d)) and \ d != TARGET_DIR_BASENAME and \ not d.startswith('.') ] if not METHOD_ROOTS: print(f"警告：在 '{BASE_IMAGE_DIR}' 中没有找到有效的方法目录 (除了 '{TARGET_DIR_BASENAME}')。") else: print(f"已识别的方法根目录 (原始): {METHOD_ROOTS}") except Exception as e: print(f"错误：在扫描 '{BASE_IMAGE_DIR}' 时发生错误: {e}"); METHOD_ROOTS = [] else: print(f"警告：基础目录 '{BASE_IMAGE_DIR}' 不存在。将无法加载候选图片。") SUBJECTS = ["subj01", "subj02", "subj05", "subj07"] # 修正了 "subj05," 的拼写 SENTINEL_TRIAL_INTERVAL = 20 NUM_TRIALS_PER_RUN = 100 # 正常运行时的每轮试验数 LOG_BATCH_SIZE = 20 DATASET_REPO_ID = "YanmHa/image-aligned-experiment-data" BATCH_LOG_FOLDER = "run_logs_batch" CSS = ".gr-block {margin-top: 4px !important; margin-bottom: 4px !important;} .compact_button { padding: 4px 8px; min-width: auto; }" # ---- 测试模式的列表缩减逻辑 (仅当 REPEAT_SINGLE_TARGET_FOR_TESTING 为 True 时生效) ---- if REPEAT_SINGLE_TARGET_FOR_TESTING: print(f"--- 特殊测试模式 (重复单一目标图) 已激活 ---") NUM_TRIALS_PER_RUN = NUM_REPEATED_TRIALS_FOR_TESTING # 确保UI显示的数字和实际测试一致 print(f"测试模式：NUM_TRIALS_PER_RUN 已被设置为: {NUM_TRIALS_PER_RUN}") if METHOD_ROOTS: original_method_roots_count = len(METHOD_ROOTS) METHOD_ROOTS = [METHOD_ROOTS[0]] print(f"测试模式：METHOD_ROOTS 已从 {original_method_roots_count} 个缩减为仅包含第一个方法: {METHOD_ROOTS}") else: print("测试模式警告：METHOD_ROOTS 为空，无法缩减。") if len(METHOD_ROOTS) == 1: # 只有一种方法时，确保至少有两个subject以形成对比 if len(SUBJECTS) >= 2: original_subjects_count = len(SUBJECTS) SUBJECTS = [SUBJECTS[0], SUBJECTS[1]] print(f"测试模式：由于方法仅1种，SUBJECTS 已从 {original_subjects_count} 个缩减为前两个: {SUBJECTS}") elif SUBJECTS: print(f"测试模式：SUBJECTS 只有一个元素 ({SUBJECTS})，方法也只有一种。注意：可能无法形成候选对。") else: print("测试模式警告：SUBJECTS 为空。") print(f"--- 特殊测试模式配置结束 ---") else: print(f"正常模式：使用完整配置。每轮目标试验数: {NUM_TRIALS_PER_RUN}") print(f"方法根目录: {METHOD_ROOTS}") print(f"Subjects: {SUBJECTS}") # ==== 全局持久化历史记录 ==== PERSISTENT_STORAGE_BASE = "/data" DATA_SUBDIR_NAME = "my_user_study_persistent_history" if not os.path.exists(PERSISTENT_STORAGE_BASE): try: os.makedirs(PERSISTENT_STORAGE_BASE, exist_ok=True) print(f"信息：基础持久化目录 '{PERSISTENT_STORAGE_BASE}' 尝试确保其存在。") except Exception as e: print(f"警告：操作基础持久化目录 '{PERSISTENT_STORAGE_BASE}' 时出现问题: {e}。") full_subdir_path = os.path.join(PERSISTENT_STORAGE_BASE, DATA_SUBDIR_NAME) if not os.path.exists(full_subdir_path): try: os.makedirs(full_subdir_path) print(f"成功创建持久化子目录: {full_subdir_path}") except Exception as e: print(f"错误：创建持久化子目录 '{full_subdir_path}' 失败: {e}") else: print(f"信息：持久化子目录 '{full_subdir_path}' 已存在。") GLOBAL_HISTORY_FILE = os.path.join(full_subdir_path, "global_experiment_shown_pairs.json") if not (os.path.isdir(full_subdir_path) and os.access(full_subdir_path, os.W_OK)): print(f"严重警告：持久化子目录 '{full_subdir_path}' 无效或不可写。") print(f"全局历史文件将被加载/保存到: {GLOBAL_HISTORY_FILE}") global_shown_pairs_cache = {} global_history_has_unsaved_changes = False exhausted_target_images = set() def load_global_shown_pairs(): global global_shown_pairs_cache, global_history_has_unsaved_changes, exhausted_target_images exhausted_target_images = set() if not GLOBAL_HISTORY_FILE or not os.path.exists(GLOBAL_HISTORY_FILE): print(f"信息：全局历史文件 '{GLOBAL_HISTORY_FILE}' 未找到或路径无效。将创建新的空历史记录。") global_shown_pairs_cache = {} global_history_has_unsaved_changes = False return try: with open(GLOBAL_HISTORY_FILE, 'r', encoding='utf-8') as f: content = f.read() if not content.strip(): print(f"信息：全局历史文件 '{GLOBAL_HISTORY_FILE}' 为空。将使用空历史记录。") global_shown_pairs_cache = {} else: data_from_file = json.loads(content) global_shown_pairs_cache = { target_img: {frozenset(pair) for pair in pairs_list} for target_img, pairs_list in data_from_file.items() } print(f"已成功从 '{GLOBAL_HISTORY_FILE}' 加载全局已展示图片对历史。") except json.JSONDecodeError as jde: print(f"错误：加载全局历史文件 '{GLOBAL_HISTORY_FILE}' 失败 (JSON解析错误: {jde})。文件内容可能已损坏。将使用空历史记录。") global_shown_pairs_cache = {} except Exception as e: print(f"错误：加载全局历史文件 '{GLOBAL_HISTORY_FILE}' 时发生其他错误: {e}。将使用空历史记录。") global_shown_pairs_cache = {} global_history_has_unsaved_changes = False def save_global_shown_pairs(): global global_shown_pairs_cache, global_history_has_unsaved_changes if not GLOBAL_HISTORY_FILE: print("错误：GLOBAL_HISTORY_FILE 未定义。无法保存历史。") return False final_save_path = os.path.abspath(GLOBAL_HISTORY_FILE) try: parent_dir = os.path.dirname(final_save_path) if not os.path.exists(parent_dir): try: os.makedirs(parent_dir, exist_ok=True) print(f"信息: 为保存历史文件，创建了父目录 {parent_dir}") except Exception as e_mkdir: print(f"错误: 创建历史文件的父目录 {parent_dir} 失败: {e_mkdir}。保存可能失败。") return False data_to_save = { target_img: [sorted(list(pair_fset)) for pair_fset in pairs_set] for target_img, pairs_set in global_shown_pairs_cache.items() } temp_file_path = final_save_path + ".tmp" with open(temp_file_path, 'w', encoding='utf-8') as f: json.dump(data_to_save, f, ensure_ascii=False, indent=2) os.replace(temp_file_path, final_save_path) print(f"已成功将全局已展示图片对历史保存到 '{final_save_path}'。") global_history_has_unsaved_changes = False return True except Exception as e: print(f"错误：保存全局历史文件 '{final_save_path}' 失败: {e}") return False load_global_shown_pairs() # ==== 加载所有可用的目标图片 ==== master_image_list = [] if os.path.exists(TARGET_DIR): try: master_image_list = sorted( [f for f in os.listdir(TARGET_DIR) if f.lower().endswith((".jpg", ".png", ".jpeg"))], key=lambda x: int(os.path.splitext(x)[0]) ) except ValueError: master_image_list = sorted([f for f in os.listdir(TARGET_DIR) if f.lower().endswith((".jpg", ".png", ".jpeg"))]) if master_image_list: print(f"警告： '{TARGET_DIR}' 文件名非纯数字，按字母排序。") if not master_image_list: print(f"警告：在 '{TARGET_DIR}' 中无有效图片。") elif not os.path.exists(TARGET_DIR) and os.path.exists(BASE_IMAGE_DIR): print(f"错误：目标目录 '{TARGET_DIR}' 未找到。") # ---- 测试模式：缩减 master_image_list (仅当 REPEAT_SINGLE_TARGET_FOR_TESTING 为 True 时生效) ---- if REPEAT_SINGLE_TARGET_FOR_TESTING: if not master_image_list: print(f"测试模式错误：master_image_list 为空，无法进行重复单一目标图测试。") else: original_first_image = master_image_list[0] master_image_list = [original_first_image] print(f"测试模式：master_image_list 已被缩减为原列表的第一个图像: {master_image_list}") if not master_image_list: print(f"关键错误：无目标图片可用 (master_image_list为空)。实验无法进行。") # ==== 辅助函数 ==== # ############################################################################# # ############# 函数修改点：get_next_trial_info ################################ # ############################################################################# # ############################################################################# # ############# 函数修改点：get_next_trial_info ################################ # ############################################################################# def get_next_trial_info(current_trial_idx_in_run, current_run_image_list_for_trial, num_trials_in_this_run_for_trial): global TARGET_DIR, METHOD_ROOTS, SUBJECTS, SENTINEL_TRIAL_INTERVAL global global_shown_pairs_cache, global_history_has_unsaved_changes, exhausted_target_images if not current_run_image_list_for_trial or current_trial_idx_in_run >= num_trials_in_this_run_for_trial: return None, current_trial_idx_in_run img_filename_original = current_run_image_list_for_trial[current_trial_idx_in_run] target_full_path = os.path.join(TARGET_DIR, img_filename_original) trial_number_for_display = current_trial_idx_in_run + 1 # ---- MODIFICATION START: 创建两个用于特定方法对比的候选池 ---- pool_image_flited = [] pool_reconed_image_color = [] # 这个独立的池用于“哨兵试验”，它需要从“任何”方法中随机抽取一个候选图 combined_pool_for_sentinel = [] for m_root_path in METHOD_ROOTS: method_name = os.path.basename(m_root_path) subjects_for_method = SUBJECTS if method_name.lower() == "takagi": if "subj01" in SUBJECTS: subjects_for_method = ["subj01"] else: continue for s_id in subjects_for_method: base, ext = os.path.splitext(img_filename_original) reconstructed_filename = f"{base}_0{ext}" candidate_path = os.path.join(m_root_path, s_id, reconstructed_filename) if os.path.exists(candidate_path): internal_label = f"{method_name}/{s_id}/{reconstructed_filename}" candidate_tuple = (internal_label, candidate_path) # 为常规试验，将候选图放入对应的特定池中 if method_name == "image_flited": pool_image_flited.append(candidate_tuple) elif method_name == "generated_images_color": pool_reconed_image_color.append(candidate_tuple) # 将“所有”有效的候选图都添加到哨兵池中 combined_pool_for_sentinel.append(candidate_tuple) # ---- MODIFICATION END: 候选池已填充完毕 ---- trial_info = {"image_id": img_filename_original, "target_path": target_full_path, "cur_no": trial_number_for_display, "is_sentinel": False, "left_display_label": "N/A", "left_internal_label": "N/A", "left_path": None, "right_display_label": "N/A", "right_internal_label": "N/A", "right_path": None} is_potential_sentinel_trial = (trial_number_for_display > 0 and trial_number_for_display % SENTINEL_TRIAL_INTERVAL == 0) if is_potential_sentinel_trial: # 对于哨兵试验，我们从包含所有方法候选图的池中随机选择一个 if not combined_pool_for_sentinel: print(f"警告：哨兵图 '{img_filename_original}' (trial {trial_number_for_display}) 无任何候选图。") else: print(f"生成哨兵试验 for '{img_filename_original}' (trial {trial_number_for_display})") trial_info["is_sentinel"] = True sentinel_candidate_target_tuple = ("目标图像", target_full_path) random_reconstruction_candidate_tuple = random.choice(combined_pool_for_sentinel) candidates_for_sentinel = [ (("目标图像", target_full_path), sentinel_candidate_target_tuple[0]), (("重建图", random_reconstruction_candidate_tuple[1]), random_reconstruction_candidate_tuple[0]) ] random.shuffle(candidates_for_sentinel) trial_info.update({ "left_display_label": candidates_for_sentinel[0][0][0], "left_path": candidates_for_sentinel[0][0][1], "left_internal_label": candidates_for_sentinel[0][1], "right_display_label": candidates_for_sentinel[1][0][0], "right_path": candidates_for_sentinel[1][0][1], "right_internal_label": candidates_for_sentinel[1][1], }) else: # 常规试验 # ---- MODIFICATION START: 新的检查与配对逻辑 ---- # 检查两个特定的池是否都有候选图 if not pool_image_flited or not pool_reconed_image_color: print(f"警告：常规图 '{img_filename_original}' (trial {trial_number_for_display}) 候选不足以形成 'image_flited' vs 'reconed_image_color' 对。 " f"('image_flited' 找到 {len(pool_image_flited)} 个, " f"'reconed_image_color' 找到 {len(pool_reconed_image_color)} 个)。此试验无法进行。") return None, current_trial_idx_in_run target_global_history_set = global_shown_pairs_cache.setdefault(img_filename_original, set()) # 从两个特定池中各选一个，生成所有可能的配对 all_possible_pairs_in_pool = [] for c_flited, c_reconed in product(pool_image_flited, pool_reconed_image_color): pair_labels_fset = frozenset({c_flited[0], c_reconed[0]}) all_possible_pairs_in_pool.append( ((c_flited, c_reconed), pair_labels_fset) ) # ---- MODIFICATION END: 新的配对逻辑已完成 ---- unseen_globally_pairs_with_data = [ item for item in all_possible_pairs_in_pool if item[1] not in target_global_history_set ] selected_candidates_tuples = None if unseen_globally_pairs_with_data: chosen_pair_data_and_labels = random.choice(unseen_globally_pairs_with_data) selected_candidates_tuples = chosen_pair_data_and_labels[0] chosen_pair_frozenset = chosen_pair_data_and_labels[1] target_global_history_set.add(chosen_pair_frozenset) global_history_has_unsaved_changes = True else: print(f"警告：目标图 '{img_filename_original}' (trial {trial_number_for_display}): 所有 ({len(all_possible_pairs_in_pool)}) 个 'image_flited' vs 'reconed_image_color' 对均已在全局展示过。") if all_possible_pairs_in_pool: print(f"目标图 '{img_filename_original}' 将被标记为已耗尽，未来轮次中将被跳过。") exhausted_target_images.add(img_filename_original) return None, current_trial_idx_in_run display_order_candidates = list(selected_candidates_tuples) if random.random() > 0.5: display_order_candidates = display_order_candidates[::-1] trial_info.update({ "left_display_label": "候选图 1", "left_path": display_order_candidates[0][1], "left_internal_label": display_order_candidates[0][0], "right_display_label": "候选图 2", "right_path": display_order_candidates[1][1], "right_internal_label": display_order_candidates[1][0], }) return trial_info, current_trial_idx_in_run + 1 # ==== 批量保存用户选择日志函数 (保持不变) ==== def save_single_log_to_hf_dataset(log_entry, user_identifier_str): global DATASET_REPO_ID, INDIVIDUAL_LOGS_FOLDER if not isinstance(log_entry, dict): print(f"错误：单个日志条目不是字典格式，无法保存：{log_entry}") return False current_user_id = user_identifier_str if user_identifier_str else "unknown_user_session" identifier_safe = str(current_user_id).replace('.', '_').replace(':', '_').replace('/', '_') print(f"用户 {identifier_safe} - 准备保存单条日志 for image {log_entry.get('image_id', 'Unknown')}...") try: token = os.getenv("HF_TOKEN") if not token: print("错误：环境变量 HF_TOKEN 未设置。无法保存单条日志到Dataset。") return False if not DATASET_REPO_ID: print("错误：DATASET_REPO_ID 未配置。无法保存单条日志到Dataset。") return False api = HfApi(token=token) image_id_safe_for_filename = os.path.splitext(log_entry.get("image_id", "unknown_img"))[0].replace('.', '_').replace(':', '_').replace('/', '_') file_creation_timestamp_str = datetime.now().strftime('%Y%m%d_%H%M%S_%f') unique_filename = (f"run{log_entry.get('run_no', 'X')}_trial{log_entry.get('trial_sequence_in_run', 'Y')}_img{image_id_safe_for_filename}_{file_creation_timestamp_str}.json") path_in_repo = f"{INDIVIDUAL_LOGS_FOLDER}/{identifier_safe}/{unique_filename}" try: json_content = json.dumps(log_entry, ensure_ascii=False, indent=2) except Exception as json_err: print(f"错误：序列化单条日志时出错: {log_entry}. 错误: {json_err}") error_log_content = {"error": "serialization_failed_single", "original_data_keys": list(log_entry.keys()) if isinstance(log_entry, dict) else None, "timestamp": datetime.now().isoformat()} json_content = json.dumps(error_log_content, ensure_ascii=False, indent=2) log_bytes = json_content.encode('utf-8') file_like_object = io.BytesIO(log_bytes) print(f"准备上传单条日志文件: {path_in_repo} ({len(log_bytes)} bytes)") api.upload_file( path_or_fileobj=file_like_object, path_in_repo=path_in_repo, repo_id=DATASET_REPO_ID, repo_type="dataset", commit_message=(f"Log choice: img {log_entry.get('image_id', 'N/A')}, run {log_entry.get('run_no', 'N/A')}, trial {log_entry.get('trial_sequence_in_run', 'N/A')} by {identifier_safe}") ) print(f"单条日志已成功保存到 HF Dataset: {DATASET_REPO_ID}/{path_in_repo}") return True except HfHubHTTPError as hf_http_error: print(f"保存单条日志到 Hugging Face Dataset 时发生 HTTP 错误 (可能被限流或权限问题): {hf_http_error}") traceback.print_exc() return False except Exception as e: print(f"保存单条日志 (image {log_entry.get('image_id', 'Unknown')}, user {identifier_safe}) 到 Hugging Face Dataset 时发生严重错误: {e}") traceback.print_exc() return False # ==== 批量保存用户选择日志函数 (确保返回 True/False) ==== def save_collected_logs_batch(list_of_log_entries, user_identifier_str, batch_identifier): global DATASET_REPO_ID, BATCH_LOG_FOLDER if not list_of_log_entries: print("批量保存用户日志：没有累积的日志。") return True identifier_safe = str(user_identifier_str if user_identifier_str else "unknown_user_session").replace('.', '_').replace(':', '_').replace('/', '_').replace(' ', '_') print(f"用户 {identifier_safe} - 准备批量保存 {len(list_of_log_entries)} 条选择日志 (批次标识: {batch_identifier})...") try: token = os.getenv("HF_TOKEN") if not token: print("错误：HF_TOKEN 未设置。无法批量保存选择日志。") return False if not DATASET_REPO_ID: print("错误：DATASET_REPO_ID 未配置。无法批量保存选择日志。") return False api = HfApi(token=token) timestamp_str = datetime.now().strftime('%Y%m%d_%H%M%S_%f') batch_filename = f"batch_user-{identifier_safe}_id-{batch_identifier}_{timestamp_str}_logs-{len(list_of_log_entries)}.jsonl" path_in_repo = f"{BATCH_LOG_FOLDER}/{identifier_safe}/{batch_filename}" jsonl_content = "" for log_entry in list_of_log_entries: try: if isinstance(log_entry, dict): jsonl_content += json.dumps(log_entry, ensure_ascii=False) + "\n" else: print(f"警告：批量保存选择日志时，条目非字典：{log_entry}") except Exception as json_err: print(f"错误：批量保存选择日志序列化单条时出错: {log_entry}. 错误: {json_err}") jsonl_content += json.dumps({"error": "serialization_failed_in_batch_user_log", "original_data_preview": str(log_entry)[:100],"timestamp": datetime.now().isoformat()}, ensure_ascii=False) + "\n" if not jsonl_content.strip(): print(f"用户 {identifier_safe} (批次 {batch_identifier}) 无可序列化选择日志。") return True log_bytes = jsonl_content.encode('utf-8') file_like_object = io.BytesIO(log_bytes) print(f"准备批量上传选择日志文件: {path_in_repo} ({len(log_bytes)} bytes)") api.upload_file( path_or_fileobj=file_like_object, path_in_repo=path_in_repo, repo_id=DATASET_REPO_ID, repo_type="dataset", commit_message=f"Batch user choice logs for {identifier_safe}, batch_id {batch_identifier} ({len(list_of_log_entries)} entries)" ) print(f"批量选择日志已成功保存到 HF Dataset: {DATASET_REPO_ID}/{path_in_repo}") return True except HfHubHTTPError as hf_http_error: print(f"批量保存选择日志到 Hugging Face Dataset 时发生 HTTP 错误 (可能被限流或权限问题): {hf_http_error}") traceback.print_exc() return False except Exception as e: print(f"批量保存选择日志 (user {identifier_safe}, batch_id {batch_identifier}) 失败: {e}") traceback.print_exc() return False # ==== 主要的 Gradio 事件处理函数 ==== def process_experiment_step( s_trial_idx_val, s_run_no_val, s_user_logs_val, s_current_trial_data_val, s_user_session_id_val, s_current_run_image_list_val, s_num_trials_this_run_val, action_type=None, choice_value=None, request: gr.Request = None ): global master_image_list, NUM_TRIALS_PER_RUN, outputs_ui_components_definition, LOG_BATCH_SIZE global REPEAT_SINGLE_TARGET_FOR_TESTING, NUM_REPEATED_TRIALS_FOR_TESTING global exhausted_target_images, global_history_has_unsaved_changes output_s_trial_idx = s_trial_idx_val; output_s_run_no = s_run_no_val output_s_user_logs = list(s_user_logs_val); output_s_current_trial_data = dict(s_current_trial_data_val) if s_current_trial_data_val else {} output_s_user_session_id = s_user_session_id_val; output_s_current_run_image_list = list(s_current_run_image_list_val) output_s_num_trials_this_run = s_num_trials_this_run_val user_ip_fallback = request.client.host if request else "unknown_ip" user_identifier_for_logging = output_s_user_session_id if output_s_user_session_id else user_ip_fallback len_ui_outputs = len(outputs_ui_components_definition) def create_ui_error_tuple(message, progress_msg_text, stop_experiment=False): btn_start_interactive = not stop_experiment btn_choices_interactive = not stop_experiment return (gr.update(visible=False),) * 3 + \ ("", "") + \ (message, progress_msg_text) + \ (gr.update(interactive=btn_start_interactive), gr.update(interactive=btn_choices_interactive), gr.update(interactive=btn_choices_interactive)) + \ (gr.update(visible=False),) def create_no_change_tuple(): return (gr.update(),) * len_ui_outputs user_id_display_text = output_s_user_session_id if output_s_user_session_id else "用户ID待分配" if action_type == "record_choice": if output_s_current_trial_data.get("data") and output_s_current_trial_data["data"].get("left_internal_label"): chosen_internal_label = (output_s_current_trial_data["data"]["left_internal_label"] if choice_value == "left" else output_s_current_trial_data["data"]["right_internal_label"]) parsed_chosen_method, parsed_chosen_subject, parsed_chosen_filename = "N/A", "N/A", "N/A" if chosen_internal_label == "目标图像": parsed_chosen_method, parsed_chosen_subject, parsed_chosen_filename = "TARGET", "GT", output_s_current_trial_data["data"]["image_id"] else: parts = chosen_internal_label.split('/'); if len(parts) == 3: parsed_chosen_method, parsed_chosen_subject, parsed_chosen_filename = parts[0].strip(), parts[1].strip(), parts[2].strip() elif len(parts) == 2: parsed_chosen_method, parsed_chosen_subject = parts[0].strip(), parts[1].strip() elif len(parts) == 1: parsed_chosen_method = parts[0].strip() log_entry = { "timestamp": datetime.now().isoformat(), "user_identifier": user_identifier_for_logging, "run_no": output_s_run_no, "image_id": output_s_current_trial_data["data"]["image_id"], "left_internal_label": output_s_current_trial_data["data"]["left_internal_label"], "right_internal_label": output_s_current_trial_data["data"]["right_internal_label"], "chosen_side": choice_value, "chosen_internal_label": chosen_internal_label, "chosen_method": parsed_chosen_method, "chosen_subject": parsed_chosen_subject, "chosen_filename": parsed_chosen_filename, "trial_sequence_in_run": output_s_current_trial_data["data"]["cur_no"], "is_sentinel": output_s_current_trial_data["data"]["is_sentinel"] } output_s_user_logs.append(log_entry) print(f"用户 {user_identifier_for_logging} 记录选择 (img: {log_entry['image_id']})。当前批次日志数: {len(output_s_user_logs)}") if len(output_s_user_logs) >= LOG_BATCH_SIZE: print(f"累积用户选择日志达到 {LOG_BATCH_SIZE} 条，准备批量保存...") batch_id_for_filename = f"run{output_s_run_no}_trialidx{output_s_trial_idx}_logcount{len(output_s_user_logs)}" user_logs_save_success = save_collected_logs_batch(list(output_s_user_logs), user_identifier_for_logging, batch_id_for_filename) if user_logs_save_success: print("批量用户选择日志已成功（或尝试）保存，将清空累积的用户选择日志列表。") output_s_user_logs = [] else: print("严重错误：批量用户选择日志保存失败。实验无法继续。") error_message_ui = "错误：日志保存失败，可能是网络问题或API限流。实验已停止，请联系管理员。" progress_message_ui = f"用户ID: {user_id_display_text} | 实验因错误停止在第 {output_s_run_no} 轮，试验 {output_s_trial_idx+1}" error_ui_updates = create_ui_error_tuple(error_message_ui, progress_message_ui, stop_experiment=True) return output_s_trial_idx, output_s_run_no, output_s_user_logs, output_s_current_trial_data, \ output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *error_ui_updates if global_history_has_unsaved_changes: print("检测到全局图片对历史自上次保存后有更新，将一并保存...") if not save_global_shown_pairs(): print("严重错误：全局图片对历史保存失败。实验无法继续。") error_message_ui = "错误：全局历史数据保存失败。实验已停止，请联系管理员。" progress_message_ui = f"用户ID: {user_id_display_text} | 实验因错误停止在第 {output_s_run_no} 轮，试验 {output_s_trial_idx+1}" error_ui_updates = create_ui_error_tuple(error_message_ui, progress_message_ui, stop_experiment=True) return output_s_trial_idx, output_s_run_no, output_s_user_logs, output_s_current_trial_data, \ output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *error_ui_updates else: print(f"用户 {user_identifier_for_logging} 错误：记录选择时当前试验数据为空或缺少internal_label！") error_ui_updates = create_ui_error_tuple("记录选择时内部错误。", f"用户ID: {user_id_display_text} | 进度：{output_s_trial_idx}/{output_s_num_trials_this_run}", stop_experiment=False) return output_s_trial_idx, output_s_run_no, output_s_user_logs, output_s_current_trial_data, output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *error_ui_updates if action_type == "start_experiment": is_first = (output_s_num_trials_this_run == 0 and output_s_trial_idx == 0 and output_s_run_no == 1) is_completed_for_restart = (output_s_num_trials_this_run > 0 and output_s_trial_idx >= output_s_num_trials_this_run) if is_completed_for_restart: if output_s_user_logs: print(f"轮次 {output_s_run_no-1} 结束，尝试保存剩余的 {len(output_s_user_logs)} 条用户选择日志...") batch_id_for_filename = f"run{output_s_run_no-1}_final_logcount{len(output_s_user_logs)}" if not save_collected_logs_batch(list(output_s_user_logs), user_identifier_for_logging, batch_id_for_filename): print("严重错误：保存上一轮剩余用户选择日志失败。实验无法继续。") error_message_ui = "错误：日志保存失败。实验已停止，请联系管理员。" progress_message_ui = f"用户ID: {user_id_display_text} | 实验因错误停止" error_ui_updates = create_ui_error_tuple(error_message_ui, progress_message_ui, stop_experiment=True) return output_s_trial_idx, output_s_run_no-1, output_s_user_logs, output_s_current_trial_data, \ output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *error_ui_updates output_s_user_logs = [] if global_history_has_unsaved_changes: print("轮次结束，尝试保存全局图片对历史...") if not save_global_shown_pairs(): print("严重错误：全局历史数据保存失败。实验无法继续。") error_message_ui = "错误：全局历史数据保存失败。实验已停止，请联系管理员。" progress_message_ui = f"用户ID: {user_id_display_text} | 实验因错误停止" error_ui_updates = create_ui_error_tuple(error_message_ui, progress_message_ui, stop_experiment=True) return output_s_trial_idx, output_s_run_no-1, output_s_user_logs, output_s_current_trial_data, \ output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *error_ui_updates if is_first or is_completed_for_restart: if is_completed_for_restart: output_s_run_no += 1 available_master_images = [img for img in master_image_list if img not in exhausted_target_images] print(f"开始轮次 {output_s_run_no}: 从 {len(master_image_list)}个总目标图片中筛选，可用图片 {len(available_master_images)}个 (已排除 {len(exhausted_target_images)}个已耗尽图片).") if not available_master_images: msg = "所有目标图片的所有唯一图片对均已展示完毕！感谢您的参与。" prog_text = f"用户ID: {user_id_display_text} | 实验完成！" if output_s_user_logs: print(f"最终轮次结束，尝试保存剩余的 {len(output_s_user_logs)} 条用户选择日志...") batch_id_for_filename = f"run{output_s_run_no-1}_final_logcount{len(output_s_user_logs)}" save_collected_logs_batch(list(output_s_user_logs), user_identifier_for_logging, batch_id_for_filename) output_s_user_logs = [] if global_history_has_unsaved_changes: print("实验最终结束，尝试保存全局图片对历史...") save_global_shown_pairs() ui_updates = list(create_ui_error_tuple(msg, prog_text, stop_experiment=True)) return 0, output_s_run_no, [], {}, output_s_user_session_id, [], 0, *tuple(ui_updates) if REPEAT_SINGLE_TARGET_FOR_TESTING and available_master_images: print(f"测试模式 (重复单一目标图) 已激活。") single_image_to_repeat = available_master_images[0] output_s_current_run_image_list = [single_image_to_repeat] * NUM_REPEATED_TRIALS_FOR_TESTING output_s_num_trials_this_run = NUM_REPEATED_TRIALS_FOR_TESTING print(f"测试模式：本轮将重复目标图片 '{single_image_to_repeat}' 共 {output_s_num_trials_this_run} 次。") else: num_really_avail = len(available_master_images) current_run_max_trials = NUM_TRIALS_PER_RUN run_size = min(num_really_avail, current_run_max_trials) if run_size == 0: error_ui = create_ui_error_tuple("错误: 可用图片采样数为0!", f"用户ID: {user_id_display_text} | 进度: 0/0", stop_experiment=False) return 0, output_s_run_no, output_s_user_logs, {}, output_s_user_session_id, [], 0, *error_ui output_s_current_run_image_list = random.sample(available_master_images, run_size) output_s_num_trials_this_run = run_size output_s_trial_idx = 0 output_s_current_trial_data = {} if is_first: timestamp_str = datetime.now().strftime('%Y%m%d%H%M%S%f'); random_val = random.randint(10000, 99999) if not output_s_user_session_id: output_s_user_session_id = f"user_{timestamp_str}_{random_val}"; user_identifier_for_logging = output_s_user_session_id else: user_identifier_for_logging = output_s_user_session_id print(f"用户会话ID: {output_s_user_session_id}") print(f"开始/继续轮次 {output_s_run_no} (用户ID: {output_s_user_session_id}). 本轮共 {output_s_num_trials_this_run} 个试验。") else: print(f"用户 {user_identifier_for_logging} 在第 {output_s_run_no} 轮，试验 {output_s_trial_idx} 点击开始，但轮次未完成。忽略。") no_change_ui = create_no_change_tuple() return output_s_trial_idx, output_s_run_no, output_s_user_logs, output_s_current_trial_data, output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *no_change_ui current_actual_trial_index_for_get_next = output_s_trial_idx if current_actual_trial_index_for_get_next >= output_s_num_trials_this_run and output_s_num_trials_this_run > 0: prog_text = f"用户ID: {output_s_user_session_id} | 进度：{output_s_num_trials_this_run}/{output_s_num_trials_this_run} | 第 {output_s_run_no} 轮 🎉" ui_updates = list(create_ui_error_tuple(f"🎉 第 {output_s_run_no} 轮完成！请点击“开始试验 / 下一轮”继续。", prog_text, stop_experiment=False)) ui_updates[7]=gr.update(interactive=True); ui_updates[8]=gr.update(interactive=False); ui_updates[9]=gr.update(interactive=False) ui_updates[0]=gr.update(value=None,visible=False); ui_updates[1]=gr.update(value=None,visible=False); ui_updates[2]=gr.update(value=None,visible=False) yield output_s_trial_idx, output_s_run_no, output_s_user_logs, {"data": None}, output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *ui_updates; return if not output_s_current_run_image_list or output_s_num_trials_this_run == 0: error_ui = create_ui_error_tuple("错误: 无法加载试验图片 (列表为空或试验数为0)", f"用户ID: {user_id_display_text} | 进度: N/A", stop_experiment=False) return output_s_trial_idx, output_s_run_no, output_s_user_logs, {"data": None}, output_s_user_session_id, [], 0, *error_ui trial_info = None next_s_trial_idx_for_state_loop = current_actual_trial_index_for_get_next while next_s_trial_idx_for_state_loop < output_s_num_trials_this_run: current_target_image_for_trial = output_s_current_run_image_list[next_s_trial_idx_for_state_loop] if current_target_image_for_trial in exhausted_target_images: print(f"信息：目标图 '{current_target_image_for_trial}' 已在全局耗尽列表中，跳过此试验。") next_s_trial_idx_for_state_loop += 1 output_s_trial_idx = next_s_trial_idx_for_state_loop continue _trial_info_candidate, _returned_next_idx = get_next_trial_info(next_s_trial_idx_for_state_loop, output_s_current_run_image_list, output_s_num_trials_this_run) if _trial_info_candidate is not None: trial_info = _trial_info_candidate output_s_trial_idx = _returned_next_idx break else: print(f"信息：目标图 '{current_target_image_for_trial}' 无法生成有效试验。尝试列表中的下一个。") next_s_trial_idx_for_state_loop +=1 output_s_trial_idx = next_s_trial_idx_for_state_loop if trial_info is None: print(f"轮次 {output_s_run_no} 中没有更多可用的有效试验了。结束本轮。") if output_s_user_logs: print(f"轮次 {output_s_run_no} 无更多有效试验，尝试保存剩余 {len(output_s_user_logs)} 条日志...") batch_id_for_filename = f"run{output_s_run_no}_no_more_trials_logcount{len(output_s_user_logs)}" if not save_collected_logs_batch(list(output_s_user_logs), user_identifier_for_logging, batch_id_for_filename): print("严重错误：保存剩余日志失败。实验可能需要停止。") output_s_user_logs = [] if global_history_has_unsaved_changes: print("轮次无更多有效试验，尝试保存全局图片对历史...") if not save_global_shown_pairs(): print("严重错误：全局历史数据保存失败。实验可能需要停止。") prog_text = f"用户ID: {output_s_user_session_id} | 进度：{output_s_num_trials_this_run}/{output_s_num_trials_this_run} | 第 {output_s_run_no} 轮 (无更多可用试验)" ui_updates = list(create_ui_error_tuple(f"第 {output_s_run_no} 轮因无更多可用试验而结束。请点击“开始试验 / 下一轮”。", prog_text, stop_experiment=False)) ui_updates[7]=gr.update(interactive=True); ui_updates[8]=gr.update(interactive=False); ui_updates[9]=gr.update(interactive=False) ui_updates[0]=gr.update(value=None,visible=False); ui_updates[1]=gr.update(value=None,visible=False); ui_updates[2]=gr.update(value=None,visible=False) yield output_s_num_trials_this_run, output_s_run_no, output_s_user_logs, {"data": None}, output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *ui_updates; return output_s_current_trial_data = {"data": trial_info} prog_text = f"用户ID: {output_s_user_session_id} | 进度：{trial_info['cur_no']}/{output_s_num_trials_this_run} | 第 {output_s_run_no} 轮" ui_show_target_updates = list(create_no_change_tuple()) ui_show_target_updates[0]=gr.update(value=trial_info["target_path"],visible=True); ui_show_target_updates[1]=gr.update(value=None,visible=False); ui_show_target_updates[2]=gr.update(value=None,visible=False) ui_show_target_updates[3]=""; ui_show_target_updates[4]=""; ui_show_target_updates[5]="请观察原图…"; ui_show_target_updates[6]=prog_text ui_show_target_updates[7]=gr.update(interactive=False); ui_show_target_updates[8]=gr.update(interactive=False); ui_show_target_updates[9]=gr.update(interactive=False) yield output_s_trial_idx, output_s_run_no, output_s_user_logs, output_s_current_trial_data, output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *ui_show_target_updates time.sleep(3) ui_show_candidates_updates = list(create_no_change_tuple()) ui_show_candidates_updates[0]=gr.update(value=None,visible=False); ui_show_candidates_updates[1]=gr.update(value=trial_info["left_path"],visible=True); ui_show_candidates_updates[2]=gr.update(value=trial_info["right_path"],visible=True) ui_show_candidates_updates[3]=gr.update(value=trial_info["left_display_label"], visible=True); ui_show_candidates_updates[4]=gr.update(value=trial_info["right_display_label"], visible=True) ui_show_candidates_updates[5]="请选择更像原图的一张"; ui_show_candidates_updates[6]=prog_text ui_show_candidates_updates[7]=gr.update(interactive=False); ui_show_candidates_updates[8]=gr.update(interactive=True); ui_show_candidates_updates[9]=gr.update(interactive=True) yield output_s_trial_idx, output_s_run_no, output_s_user_logs, output_s_current_trial_data, output_s_user_session_id, output_s_current_run_image_list, output_s_num_trials_this_run, *ui_show_candidates_updates # ==== Gradio UI 定义 和 程序入口 ==== def handle_download_history_file(): global GLOBAL_HISTORY_FILE if os.path.exists(GLOBAL_HISTORY_FILE): try: if os.path.getsize(GLOBAL_HISTORY_FILE) > 0: print(f"准备提供文件下载: {GLOBAL_HISTORY_FILE}") return GLOBAL_HISTORY_FILE, gr.update(value=f"点击上面的链接下载 '{os.path.basename(GLOBAL_HISTORY_FILE)}'") else: print(f"历史文件 '{GLOBAL_HISTORY_FILE}' 为空，不提供下载。") return None, gr.update(value=f"提示: 历史文件 '{os.path.basename(GLOBAL_HISTORY_FILE)}' 当前为空。") except Exception as e: print(f"检查历史文件大小时出错 '{GLOBAL_HISTORY_FILE}': {e}") return None, gr.update(value=f"错误: 检查历史文件时出错。") else: print(f"请求下载历史文件，但文件 '{GLOBAL_HISTORY_FILE}' 未找到。") return None, gr.update(value=f"错误: JSON历史文件 '{os.path.basename(GLOBAL_HISTORY_FILE)}' 未找到。请先运行实验以生成数据并触发保存。") welcome_page_markdown = """ ## 欢迎加入实验！ 您好！非常感谢您抽出宝贵时间参与我们的视觉偏好评估实验。您的选择将帮助我们改进重建算法，让机器生成的图像更贴近人类视觉体验! 1. **实验目的**：通过比较两幅 重建图像 与原始 目标图像 的相似度。 2. **操作流程**： * 点击下方的「我已阅读并同意开始实验」按钮。 * 然后点击主实验界面的「开始试验 / 下一轮」按钮。 * 系统先展示一张 **目标图像**，持续 3 秒。 * 随后自动切换到 **两张重建图像**。 * 根据刚才的观察记忆，选出您认为与目标图像最相似的一张。 * 选择后系统会自动进入下一轮比较。 3. **温馨提示**： * 请勿刷新或关闭页面，以免中断实验。 * 若图片加载稍有延迟，请耐心等待；持续异常可联系邮箱 yangminghan@bupt.edu.cn。 * 本实验将保护您的任何个人隐私信息，所有数据仅用于学术研究，请您认真选择和填写。 4. **奖励说明**： * 完成全部轮次后，请截图记录您所完成的实验总数（可累积,页面左下角将显示进度,请保证截取到为您分配的ID，轮次）。 * 将截图发送至邮箱 yangminghan@bupt.edu.cn，我们将在核验后发放奖励。 再次感谢您的参与与支持！您每一次认真选择都对我们的研究意义重大。祝您一切顺利，实验愉快！ """ def handle_agree_and_start(name, gender, age, education, request: gr.Request): error_messages_list = [] if not name or str(name).strip() == "": error_messages_list.append("姓名 不能为空。") if gender is None or str(gender).strip() == "": error_messages_list.append("性别 必须选择。") if age is None: error_messages_list.append("年龄 不能为空。") elif not (isinstance(age, (int, float)) and 1 <= age <= 120): try: num_age = float(age); except (ValueError, TypeError): error_messages_list.append("年龄必须是一个有效的数字。") else: if not (1 <= num_age <= 120): error_messages_list.append("年龄必须在 1 到 120 之间。") if education is None or str(education).strip() == "其他": error_messages_list.append("学历 必须选择。") if error_messages_list: full_error_message = "请修正以下错误：\n" + "\n".join([f"- {msg}" for msg in error_messages_list]) print(f"用户输入验证失败: {full_error_message}") return gr.update(), False, gr.update(visible=True), gr.update(visible=False), full_error_message s_name = str(name).strip().replace(" ","_").replace("/","_").replace("\\","_") s_gender = str(gender).strip().replace(" ","_").replace("/","_").replace("\\","_") s_age = str(int(float(age))) s_education = str(education).strip().replace(" ","_").replace("/","_").replace("\\","_") user_id_str = f"N-{s_name}_G-{s_gender}_A-{s_age}_E-{s_education}" print(f"用户信息收集完毕，生成用户ID: {user_id_str}") return user_id_str, True, gr.update(visible=False), gr.update(visible=True), "" with gr.Blocks(css=CSS, title="图像重建主观评估") as demo: s_show_experiment_ui = gr.State(False); s_trial_index = gr.State(0); s_run_no = gr.State(1) s_user_logs = gr.State([]); s_current_trial_data = gr.State({}); s_user_session_id = gr.State(None) s_current_run_image_list = gr.State([]); s_num_trials_this_run = gr.State(0) welcome_container = gr.Column(visible=True) experiment_container = gr.Column(visible=False) with welcome_container: gr.Markdown(welcome_page_markdown) with gr.Row(): user_name_input = gr.Textbox(label="请输入您的姓名或代号 (例如 张三 或 User001)", placeholder="例如：张三 -> ZS"); user_gender_input = gr.Radio(label="性别", choices=["男", "女"]) with gr.Row(): user_age_input = gr.Number(label="年龄 (请输入1-120的整数)", minimum=1, maximum=120, step=1); user_education_input = gr.Dropdown(label="学历", choices=["其他","初中及以下","高中（含中专）", "大专（含在读）", "本科（含在读）", "硕士（含在读）", "博士（含在读）"]) welcome_error_msg = gr.Markdown(value="") btn_agree_and_start = gr.Button("我已阅读上述说明并同意参与实验") with experiment_container: gr.Markdown("## 🧠 图像重建主观评估实验"); gr.Markdown(f"每轮实验大约有 {NUM_TRIALS_PER_RUN} 次比较。") with gr.Row(): with gr.Column(scale=1, min_width=300): left_img = gr.Image(label="左候选图", visible=False, height=400, interactive=False); left_lbl = gr.Textbox(label="左图信息", value="", visible=True, interactive=False, max_lines=1); btn_left = gr.Button("选择左图 (更相似)", interactive=False, elem_classes="compact_button") with gr.Column(scale=1, min_width=300): right_img = gr.Image(label="右候选图", visible=False, height=400, interactive=False); right_lbl = gr.Textbox(label="右图信息",value="", visible=True, interactive=False, max_lines=1); btn_right = gr.Button("选择右图 (更相似)", interactive=False, elem_classes="compact_button") with gr.Row(): target_img = gr.Image(label="目标图像 (观察3秒后消失)", visible=False, height=400, interactive=False) with gr.Row(): status_text = gr.Markdown(value="请点击“开始试验 / 下一轮”按钮。") with gr.Row(): progress_text = gr.Markdown() with gr.Row(): btn_start = gr.Button("开始试验 / 下一轮") btn_download_json = gr.Button("下载JSON历史记录") json_download_output = gr.File(label="下载的文件会在此处提供", interactive=False) file_out_placeholder = gr.File(label=" ", visible=False, interactive=False) outputs_ui_components_definition = [ target_img, left_img, right_img, left_lbl, right_lbl, status_text, progress_text, btn_start, btn_left, btn_right, file_out_placeholder ] click_inputs_base = [ s_trial_index, s_run_no, s_user_logs, s_current_trial_data, s_user_session_id, s_current_run_image_list, s_num_trials_this_run ] event_outputs = [ s_trial_index, s_run_no, s_user_logs, s_current_trial_data, s_user_session_id, s_current_run_image_list, s_num_trials_this_run, *outputs_ui_components_definition ] btn_agree_and_start.click(fn=handle_agree_and_start, inputs=[user_name_input, user_gender_input, user_age_input, user_education_input], outputs=[s_user_session_id, s_show_experiment_ui, welcome_container, experiment_container, welcome_error_msg]) btn_start.click(fn=partial(process_experiment_step, action_type="start_experiment"), inputs=click_inputs_base, outputs=event_outputs, queue=True) btn_left.click(fn=partial(process_experiment_step, action_type="record_choice", choice_value="left"), inputs=click_inputs_base, outputs=event_outputs, queue=True) btn_right.click(fn=partial(process_experiment_step, action_type="record_choice", choice_value="right"), inputs=click_inputs_base, outputs=event_outputs, queue=True) btn_download_json.click(fn=handle_download_history_file, inputs=None, outputs=[json_download_output, status_text]) if __name__ == "__main__": if not master_image_list: print("\n关键错误：程序无法启动，因无目标图片。"); exit() else: print(f"从 '{TARGET_DIR}' 加载 {len(master_image_list)} 张目标图片。") if not METHOD_ROOTS: print(f"警告: '{BASE_IMAGE_DIR}' 无候选方法子目录。") if not SUBJECTS: print("警告: SUBJECTS 列表为空。") print(f"用户选择日志保存到 Dataset: '{DATASET_REPO_ID}' 的 '{BATCH_LOG_FOLDER}/ 文件夹") if not os.getenv("HF_TOKEN"): print("警告: HF_TOKEN 未设置。日志无法保存到Hugging Face Dataset。\n 请在 Space Secrets 中设置 HF_TOKEN。") else: print("HF_TOKEN 已找到。") print(f"全局图片对历史将从 '{GLOBAL_HISTORY_FILE}' 加载/保存到此文件。") allowed_paths_list = [] image_base_dir_to_allow = BASE_IMAGE_DIR if os.path.exists(image_base_dir_to_allow) and os.path.isdir(image_base_dir_to_allow): allowed_paths_list.append(os.path.abspath(image_base_dir_to_allow)) else: print(f"关键警告：图片基础目录 '{image_base_dir_to_allow}' 不存在或非目录。") if os.path.exists(PERSISTENT_STORAGE_BASE) and os.path.isdir(PERSISTENT_STORAGE_BASE): allowed_paths_list.append(os.path.abspath(PERSISTENT_STORAGE_BASE)) else: print(f"警告：持久化存储基础目录 '{PERSISTENT_STORAGE_BASE}' 不存在。JSON历史文件下载可能受影响。") try: os.makedirs(PERSISTENT_STORAGE_BASE, exist_ok=True) print(f"信息：已尝试创建目录 '{PERSISTENT_STORAGE_BASE}'。") if os.path.exists(PERSISTENT_STORAGE_BASE) and os.path.isdir(PERSISTENT_STORAGE_BASE): allowed_paths_list.append(os.path.abspath(PERSISTENT_STORAGE_BASE)) except Exception as e_mkdir_main: print(f"错误：在 main 中创建目录 '{PERSISTENT_STORAGE_BASE}' 失败: {e_mkdir_main}") final_allowed_paths = list(set(allowed_paths_list)) if final_allowed_paths: print(f"Gradio demo.launch() 配置最终 allowed_paths: {final_allowed_paths}") else: print("警告：没有有效的 allowed_paths 被配置。Gradio文件访问可能受限。") print("启动 Gradio 应用...") if final_allowed_paths: demo.launch(allowed_paths=final_allowed_paths) else: demo.launch()