在使用SAST算法实现印章弯曲文本检测时出现训练loss下降，但验证时精度为0的情况

[Dhao007]

请提供下述完整信息以便快速定位问题/Please provide the following information to quickly locate the problem

• 系统环境/System Environment：Ubuntu 18.04.6
• 版本号/Version：Paddle：paddlepaddle-gpu2.5.1 PaddleOCR：paddleocr2.7.0.2
• 运行指令/Command Code： python3 -m paddle.distributed.launch --gpus '0,1' tools/train.py \ -c myconfig/det/det_r50_vd_sast_icdar15.yml \ 在官方配置文件上进行修改https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/configs/det/det_r50_vd_sast_icdar15.yml

修改的地方为： epoch_num: 5000 -> epoch_num: 1000 save_model_dir: ./output/sast_r50_vd_ic15/ -> save_model_dir: ./output/det_icdar2023ReST/sast_r50/ save_epoch_step: 1000 -> save_epoch_step: 200 eval_batch_step: [4000, 5000] -> eval_batch_step: [0, 4000] pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained -> ./output/det_icdar2023ReST/sast_r50/best_accuracy.pdparams 以及自定义的数据集： 对印章文字区域进行多点标注，标注点数不固定，我标注的方式是根据文字数量多少进行标注，上下点数量是相而且对称，如下图所示：

部分训练日志： [2023/11/16 18:20:21] ppocr INFO: epoch: [9/1000], global_step: 3996, lr: 0.001000, loss: 0.953236, score_loss: 0.212883, border_loss: 0.325157, tvo_loss: 0.238184, tco_loss: 0.055055, avg_reader_cost: 0.00020 s, avg_batch_cost: 0.26244 s, avg_samples: 4.0, ips: 15.24167 samples/s, eta: 1 day, 12:07:28 [2023/11/16 18:20:22] ppocr INFO: epoch: [9/1000], global_step: 3998, lr: 0.001000, loss: 1.105541, score_loss: 0.212883, border_loss: 0.357708, tvo_loss: 0.238184, tco_loss: 0.055588, avg_reader_cost: 0.00019 s, avg_batch_cost: 0.26301 s, avg_samples: 4.0, ips: 15.20844 samples/s, eta: 1 day, 12:07:25 [2023/11/16 18:20:22] ppocr INFO: epoch: [9/1000], global_step: 4000, lr: 0.001000, loss: 1.110364, score_loss: 0.224519, border_loss: 0.380232, tvo_loss: 0.245484, tco_loss: 0.056110, avg_reader_cost: 0.00021 s, avg_batch_cost: 0.26051 s, avg_samples: 4.0, ips: 15.35465 samples/s, eta: 1 day, 12:07:22 [2023/11/16 18:25:44] ppocr INFO: cur metric, precision: 0.0, recall: 0.0, hmean: 0, fps: 7.835941393421171 [2023/11/16 18:25:53] ppocr INFO: save best model is to ./output/det_icdar2023ReST/sast_r50/best_accuracy [2023/11/16 18:25:53] ppocr INFO: best metric, hmean: 0, is_float16: False, precision: 0.0, recall: 0.0, fps: 7.835941393421171, best_epoch: 9 [2023/11/16 18:25:54] ppocr INFO: epoch: [9/1000], global_step: 4002, lr: 0.001000, loss: 1.110364, score_loss: 0.224519, border_loss: 0.380232, tvo_loss: 0.262929, tco_loss: 0.058905, avg_reader_cost: 0.00142 s, avg_batch_cost: 0.24582 s, avg_samples: 4.0, ips: 16.27197 samples/s, eta: 1 day, 12:07:15 [2023/11/16 18:25:54] ppocr INFO: epoch: [9/1000], global_step: 4004, lr: 0.001000, loss: 1.203508, score_loss: 0.224519, border_loss: 0.380232, tvo_loss: 0.245484, tco_loss: 0.058905, avg_reader_cost: 0.00140 s, avg_batch_cost: 0.25095 s, avg_samples: 4.0, ips: 15.93964 samples/s, eta: 1 day, 12:07:09 [2023/11/16 18:25:55] ppocr INFO: epoch: [9/1000], global_step: 4006, lr: 0.001000, loss: 1.121378, score_loss: 0.214334, border_loss: 0.392141, tvo_loss: 0.245484, tco_loss: 0.056110, avg_reader_cost: 0.00020 s, avg_batch_cost: 0.25236 s, avg_samples: 4.0, ips: 15.85016 samples/s, eta: 1 day, 12:07:04

其他： 我在DB++算法上跑相同的数据集，验证时精度有结果，同个数据集在SAST算法上就没有结果，很奇怪，有没有大佬遇到类似的问题。