import json import os def get_relative_position(bbox, img_width, img_height): """ 根据边界框在图像中的相对位置生成描述。 bbox: [x, y, width, height] (Label Studio 导出的百分比坐标) img_width, img_height: 图像的原始宽度和高度 """ if img_width == 0 or img_height == 0: return "未知位置" center_x_percent = bbox[0] + bbox[2] / 2 center_y_percent = bbox[1] + bbox[3] / 2 parts = [] if center_y_percent < 33.3: parts.append("顶部") elif center_y_percent > 66.6: parts.append("底部") else: parts.append("中部") if center_x_percent < 33.3: parts.append("偏左") elif center_x_percent > 66.6: parts.append("偏右") else: parts.append("中心") return "".join(parts) + "区域" def generate_qwen_vl_data(label_studio_json_path, output_jsonl_path, image_base_dir): """ 将 Label Studio 导出的 JSON 转换为 Qwen-VL 微调所需的 JSONL 格式。 Args: label_studio_json_path (str): Label Studio 导出的 JSON 文件路径。 output_jsonl_path (str): 输出的 JSONL 文件路径。 image_base_dir (str): 图片文件所在的根目录，Label Studio JSON 中的路径是相对此目录的。 例如，如果JSON中是"/data/upload/1/image.jpg"，而实际图片在"your_project_data/images/image.jpg"， 那么image_base_dir就是"your_project_data/images"。 """ with open(label_studio_json_path, 'r', encoding='utf-8') as f: data = json.load(f) qwen_vl_samples = [] for task in data: image_relative_path = task['data']['image'] # --- 修改这里 --- # 使用 os.path.basename 来安全地提取文件名 image_filename = os.path.basename(image_relative_path) image_absolute_path = os.path.join(image_base_dir, image_filename).replace('\\', '/') # --- 修改结束 --- annotations = task.get('annotations', []) if not annotations: print(f"Warning: Task {task['id']} has no annotations. Skipping.") continue # 取第一个 annotation 结果，通常只有一个 annotation_results = annotations[0].get('result', []) if not annotation_results: print(f"Warning: Task {task['id']} annotation has no results. Skipping.") continue component_type = "未知构件" crack_type = "未知裂缝类型" text_description = "" bbox_descriptions = [] original_width = None original_height = None # 遍历所有标注结果 for res in annotation_results: if res['from_name'] == 'componentClassification' and res['type'] == 'choices': component_type = res['value']['choices'][0] if res['value']['choices'] else component_type elif res['from_name'] == 'cracksClassification' and res['type'] == 'choices': crack_type = res['value']['choices'][0] if res['value']['choices'] else crack_type elif res['from_name'] == 'textTool' and res['type'] == 'textarea': text_description = res['value']['text'][0] if res['value']['text'] else text_description elif res['type'] in ['rectanglelabels', 'polygonlabels']: # 获取图像原始尺寸，用于计算相对位置 if original_width is None: # 只需要获取一次 original_width = res.get('original_width', 0) original_height = res.get('original_height', 0) # 如果当前结果没有原始尺寸，尝试从task['annotations'][0]中查找 if original_width == 0 and annotations[0].get('result'): for r_sub in annotations[0]['result']: if r_sub.get('original_width') and r_sub.get('original_height'): original_width = r_sub['original_width'] original_height = r_sub['original_height'] break if res['type'] == 'rectanglelabels': bbox = res['value'] # Label Studio 导出的 x, y, width, height 是百分比 x_percent = bbox['x'] y_percent = bbox['y'] width_percent = bbox['width'] height_percent = bbox['height'] label = bbox['rectanglelabels'][0] if bbox['rectanglelabels'] else "区域" # 尝试生成更具描述性的位置 position_desc = get_relative_position([x_percent, y_percent, width_percent, height_percent], original_width, original_height) bbox_descriptions.append(f"一个 '{label}' 位于图像的 {position_desc}。") elif res['type'] == 'polygonlabels': # 多边形通常更复杂，这里简化为描述其标签和大致位置 # 可以通过计算多边形中心来获取大致位置 points = res['value']['points'] if points: # 计算多边形中心的大致位置 avg_x = sum([p[0] for p in points]) / len(points) avg_y = sum([p[1] for p in points]) / len(points) label = res['value']['polygonlabels'][0] if res['value']['polygonlabels'] else "区域" position_desc = get_relative_position([avg_x, avg_y, 0, 0], original_width, original_height) # width/height设为0，只用中心点 bbox_descriptions.append(f"一个 '{label}' 区域位于图像的 {position_desc}。") # 构建 Instruction 和 Response instruction = "请描述这张图片中的所有裂缝信息，包括构件、裂缝类型、位置和详细描述。" response_parts = [] if component_type != "未知构件": response_parts.append(f"图片显示的是一个 '{component_type}'。") if crack_type != "未知裂缝类型": response_parts.append(f"主要裂缝类型是 '{crack_type}'。") if text_description: response_parts.append(f"详细情况描述为：'{text_description}'。") if bbox_descriptions: response_parts.append("图中发现以下缺陷区域：") for desc in bbox_descriptions: response_parts.append(f"- {desc}") if not response_parts: # 如果没有任何有效信息，跳过此任务 print(f"Warning: Task {task['id']} has no meaningful annotations to generate a response. Skipping.") continue response = " ".join(response_parts) # 构建 Qwen-VL 格式的样本 qwen_vl_sample = { "image": image_absolute_path, "conversations": [ {"from": "user", "value": instruction}, {"from": "assistant", "value": response} ] } qwen_vl_samples.append(qwen_vl_sample) with open(output_jsonl_path, 'w', encoding='utf-8') as f: for sample in qwen_vl_samples: f.write(json.dumps(sample, ensure_ascii=False) + '\n') print(f"成功生成 {len(qwen_vl_samples)} 条 Qwen-VL 微调数据到 {output_jsonl_path}") # --- 使用示例 --- if __name__ == "__main__": # 你的 Label Studio 导出 JSON 文件路径 label_studio_json_file = 'annotations.json' # 输出的 Qwen-VL 微调 JSONL 文件路径 output_jsonl_file = 'annotations.jsonl' # 你的图片文件所在的根目录。 # Label Studio 导出的 JSON 中的图片路径是相对的，例如 "/data/upload/1/f238f029-0319.jpg_wh300.jpg" # 你需要根据实际情况调整 `image_base_dir`，使其指向你存储 Label Studio 图片的实际文件夹。 # 例如：如果你的图片是 `my_project/images/f238f029-0319.jpg_wh300.jpg` # 那么 `image_base_dir` 就应该是 `my_project/images` image_base_directory = 'images' # <-- 请务必修改为你的实际图片根目录！ # 运行转换 generate_qwen_vl_data(label_studio_json_file, output_jsonl_file, image_base_directory) # 打印生成的第一个样本，方便检查 with open(output_jsonl_file, 'r', encoding='utf-8') as f: first_sample = json.loads(f.readline()) print("\n--- 第一个生成的样本示例 ---") print(json.dumps(first_sample, indent=2, ensure_ascii=False))