PP-OCRv5_server_det 模型微调后效果更差

[sf9ehf9fe]

基于PPOCRLabel 制作数据集用于微调模型，微调后进行模型评估正确率将近0.95，但模型导出后使用没有一点效果，这是怎么回事？

det模型训练
python3 tools/train.py -c configs/det/PP-OCRv5/PP-OCRv5_server_det.yml
-o Global.pretrained_model=./PP-OCRv5_server_det_pretrained.pdparams
Train.dataset.data_dir=/export/DKyProject/PaddleOCR-main/dataset/det_imges
Train.dataset.label_file_list='[/export/DKyProject/PaddleOCR-main/dataset/det_train.txt]'
Eval.dataset.data_dir=/export/DKyProject/PaddleOCR-main/dataset/det_imges
Eval.dataset.label_file_list='[/export/DKyProject/PaddleOCR-main/dataset/det_val.txt]'
[2026/01/29 14:53:57] ppocr INFO: Architecture :
[2026/01/29 14:53:57] ppocr INFO: Backbone :
[2026/01/29 14:53:57] ppocr INFO: det : True
[2026/01/29 14:53:57] ppocr INFO: name : PPHGNetV2_B4
[2026/01/29 14:53:57] ppocr INFO: Head :
[2026/01/29 14:53:57] ppocr INFO: k : 50
[2026/01/29 14:53:57] ppocr INFO: mode : large
[2026/01/29 14:53:57] ppocr INFO: name : PFHeadLocal
[2026/01/29 14:53:57] ppocr INFO: Neck :
[2026/01/29 14:53:57] ppocr INFO: intracl : True
[2026/01/29 14:53:57] ppocr INFO: name : LKPAN
[2026/01/29 14:53:57] ppocr INFO: out_channels : 256
[2026/01/29 14:53:57] ppocr INFO: Transform : None
[2026/01/29 14:53:57] ppocr INFO: algorithm : DB
[2026/01/29 14:53:57] ppocr INFO: model_type : det
[2026/01/29 14:53:57] ppocr INFO: Eval :
[2026/01/29 14:53:57] ppocr INFO: dataset :
[2026/01/29 14:53:57] ppocr INFO: data_dir : /export/DKyProject/PaddleOCR-main/dataset/det_imges
[2026/01/29 14:53:57] ppocr INFO: label_file_list : ['/export/DKyProject/PaddleOCR-main/dataset/det_val.txt']
[2026/01/29 14:53:57] ppocr INFO: name : SimpleDataSet
[2026/01/29 14:53:57] ppocr INFO: transforms :
[2026/01/29 14:53:57] ppocr INFO: DecodeImage :
[2026/01/29 14:53:57] ppocr INFO: channel_first : False
[2026/01/29 14:53:57] ppocr INFO: img_mode : BGR
[2026/01/29 14:53:57] ppocr INFO: DetLabelEncode : None
[2026/01/29 14:53:57] ppocr INFO: DetResizeForTest : None
[2026/01/29 14:53:57] ppocr INFO: NormalizeImage :
[2026/01/29 14:53:57] ppocr INFO: mean : [0.485, 0.456, 0.406]
[2026/01/29 14:53:57] ppocr INFO: order : hwc
[2026/01/29 14:53:57] ppocr INFO: scale : 1./255.
[2026/01/29 14:53:57] ppocr INFO: std : [0.229, 0.224, 0.225]
[2026/01/29 14:53:57] ppocr INFO: ToCHWImage : None
[2026/01/29 14:53:57] ppocr INFO: KeepKeys :
[2026/01/29 14:53:57] ppocr INFO: keep_keys : ['image', 'shape', 'polys', 'ignore_tags']
[2026/01/29 14:53:57] ppocr INFO: loader :
[2026/01/29 14:53:57] ppocr INFO: batch_size_per_card : 1
[2026/01/29 14:53:57] ppocr INFO: drop_last : False
[2026/01/29 14:53:57] ppocr INFO: num_workers : 2
[2026/01/29 14:53:57] ppocr INFO: shuffle : False
[2026/01/29 14:53:57] ppocr INFO: Global :
[2026/01/29 14:53:57] ppocr INFO: cal_metric_during_train : False
[2026/01/29 14:53:57] ppocr INFO: checkpoints : None
[2026/01/29 14:53:57] ppocr INFO: debug : False
[2026/01/29 14:53:57] ppocr INFO: distributed : False
[2026/01/29 14:53:57] ppocr INFO: epoch_num : 500
[2026/01/29 14:53:57] ppocr INFO: eval_batch_step : [0, 1500]
[2026/01/29 14:53:57] ppocr INFO: infer_img : doc/imgs_en/img_10.jpg
[2026/01/29 14:53:57] ppocr INFO: log_smooth_window : 20
[2026/01/29 14:53:57] ppocr INFO: model_name : PP-OCRv5_server_det
[2026/01/29 14:53:57] ppocr INFO: pretrained_model : ./PP-OCRv5_server_det_pretrained.pdparams
[2026/01/29 14:53:57] ppocr INFO: print_batch_step : 10
[2026/01/29 14:53:57] ppocr INFO: save_epoch_step : 10
[2026/01/29 14:53:57] ppocr INFO: save_inference_dir : None
[2026/01/29 14:53:57] ppocr INFO: save_model_dir : ./output/PP-OCRv5_server_det
[2026/01/29 14:53:57] ppocr INFO: save_res_path : ./checkpoints/det_db/predicts_db.txt
[2026/01/29 14:53:57] ppocr INFO: use_gpu : True
[2026/01/29 14:53:57] ppocr INFO: use_visualdl : False
[2026/01/29 14:53:57] ppocr INFO: Loss :
[2026/01/29 14:53:57] ppocr INFO: alpha : 5
[2026/01/29 14:53:57] ppocr INFO: balance_loss : True
[2026/01/29 14:53:57] ppocr INFO: beta : 10
[2026/01/29 14:53:57] ppocr INFO: main_loss_type : DiceLoss
[2026/01/29 14:53:57] ppocr INFO: name : DBLoss
[2026/01/29 14:53:57] ppocr INFO: ohem_ratio : 3
[2026/01/29 14:53:57] ppocr INFO: Metric :
[2026/01/29 14:53:57] ppocr INFO: main_indicator : hmean
[2026/01/29 14:53:57] ppocr INFO: name : DetMetric
[2026/01/29 14:53:57] ppocr INFO: Optimizer :
[2026/01/29 14:53:57] ppocr INFO: beta1 : 0.9
[2026/01/29 14:53:57] ppocr INFO: beta2 : 0.999
[2026/01/29 14:53:57] ppocr INFO: lr :
[2026/01/29 14:53:57] ppocr INFO: learning_rate : 0.001
[2026/01/29 14:53:57] ppocr INFO: name : Cosine
[2026/01/29 14:53:57] ppocr INFO: warmup_epoch : 2
[2026/01/29 14:53:57] ppocr INFO: name : Adam
[2026/01/29 14:53:57] ppocr INFO: regularizer :
[2026/01/29 14:53:57] ppocr INFO: factor : 1e-6
[2026/01/29 14:53:57] ppocr INFO: name : L2
[2026/01/29 14:53:57] ppocr INFO: PostProcess :
[2026/01/29 14:53:57] ppocr INFO: box_thresh : 0.6
[2026/01/29 14:53:57] ppocr INFO: max_candidates : 1000
[2026/01/29 14:53:57] ppocr INFO: name : DBPostProcess
[2026/01/29 14:53:57] ppocr INFO: thresh : 0.3
[2026/01/29 14:53:57] ppocr INFO: unclip_ratio : 1.5
[2026/01/29 14:53:57] ppocr INFO: Train :
[2026/01/29 14:53:57] ppocr INFO: dataset :
[2026/01/29 14:53:57] ppocr INFO: data_dir : /export/DKyProject/PaddleOCR-main/dataset/det_imges
[2026/01/29 14:53:57] ppocr INFO: label_file_list : ['/export/DKyProject/PaddleOCR-main/dataset/det_train.txt']
[2026/01/29 14:53:57] ppocr INFO: name : SimpleDataSet
[2026/01/29 14:53:57] ppocr INFO: ratio_list : [1.0]
[2026/01/29 14:53:57] ppocr INFO: transforms :
[2026/01/29 14:53:57] ppocr INFO: DecodeImage :
[2026/01/29 14:53:57] ppocr INFO: channel_first : False
[2026/01/29 14:53:57] ppocr INFO: img_mode : BGR
[2026/01/29 14:53:57] ppocr INFO: DetLabelEncode : None
[2026/01/29 14:53:57] ppocr INFO: CopyPaste : None
[2026/01/29 14:53:57] ppocr INFO: IaaAugment :
[2026/01/29 14:53:57] ppocr INFO: augmenter_args :
[2026/01/29 14:53:57] ppocr INFO: args :
[2026/01/29 14:53:57] ppocr INFO: p : 0.5
[2026/01/29 14:53:57] ppocr INFO: type : Fliplr
[2026/01/29 14:53:57] ppocr INFO: args :
[2026/01/29 14:53:57] ppocr INFO: rotate : [-10, 10]
[2026/01/29 14:53:57] ppocr INFO: type : Affine
[2026/01/29 14:53:57] ppocr INFO: args :
[2026/01/29 14:53:57] ppocr INFO: size : [0.5, 3]
[2026/01/29 14:53:57] ppocr INFO: type : Resize
[2026/01/29 14:53:57] ppocr INFO: EastRandomCropData :
[2026/01/29 14:53:57] ppocr INFO: keep_ratio : True
[2026/01/29 14:53:57] ppocr INFO: max_tries : 50
[2026/01/29 14:53:57] ppocr INFO: size : [640, 640]
[2026/01/29 14:53:57] ppocr INFO: MakeBorderMap :
[2026/01/29 14:53:57] ppocr INFO: shrink_ratio : 0.4
[2026/01/29 14:53:57] ppocr INFO: thresh_max : 0.7
[2026/01/29 14:53:57] ppocr INFO: thresh_min : 0.3
[2026/01/29 14:53:57] ppocr INFO: total_epoch : 500
[2026/01/29 14:53:57] ppocr INFO: MakeShrinkMap :
[2026/01/29 14:53:57] ppocr INFO: min_text_size : 8
[2026/01/29 14:53:57] ppocr INFO: shrink_ratio : 0.4
[2026/01/29 14:53:57] ppocr INFO: total_epoch : 500
[2026/01/29 14:53:57] ppocr INFO: NormalizeImage :
[2026/01/29 14:53:57] ppocr INFO: mean : [0.485, 0.456, 0.406]
[2026/01/29 14:53:57] ppocr INFO: order : hwc
[2026/01/29 14:53:57] ppocr INFO: scale : 1./255.
[2026/01/29 14:53:57] ppocr INFO: std : [0.229, 0.224, 0.225]
[2026/01/29 14:53:57] ppocr INFO: ToCHWImage : None
[2026/01/29 14:53:57] ppocr INFO: KeepKeys :
[2026/01/29 14:53:57] ppocr INFO: keep_keys : ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask']
[2026/01/29 14:53:57] ppocr INFO: loader :
[2026/01/29 14:53:57] ppocr INFO: batch_size_per_card : 8
[2026/01/29 14:53:57] ppocr INFO: drop_last : False
[2026/01/29 14:53:57] ppocr INFO: num_workers : 8
[2026/01/29 14:53:57] ppocr INFO: shuffle : True
[2026/01/29 14:53:57] ppocr INFO: profiler_options : None
[2026/01/29 14:53:57] ppocr INFO: train with paddle 3.2.2 and device Place(gpu:0)
[2026/01/29 14:53:57] ppocr INFO: Initialize indexes of datasets:['/export/DKyProject/PaddleOCR-main/dataset/det_train.txt']
[2026/01/29 14:53:57] ppocr INFO: Initialize indexes of datasets:['/export/DKyProject/PaddleOCR-main/dataset/det_val.txt']
[2026/01/29 14:53:57] ppocr INFO: train dataloader has 114 iters
[2026/01/29 14:53:57] ppocr INFO: valid dataloader has 267 iters
[2026/01/29 14:53:57] ppocr INFO: load pretrain successful from ./PP-OCRv5_server_det_pretrained
[2026/01/29 14:53:57] ppocr INFO: During the training process, after the 0th iteration, an evaluation is run every 1500 iterations
[2026/01/29 14:54:47] ppocr INFO: Architecture :
[2026/01/29 14:54:47] ppocr INFO: Backbone :
[2026/01/29 14:54:47] ppocr INFO: det : True
[2026/01/29 14:54:47] ppocr INFO: name : PPHGNetV2_B4
[2026/01/29 14:54:47] ppocr INFO: Head :
[2026/01/29 14:54:47] ppocr INFO: k : 50
[2026/01/29 14:54:47] ppocr INFO: mode : large
[2026/01/29 14:54:47] ppocr INFO: name : PFHeadLocal
[2026/01/29 14:54:47] ppocr INFO: Neck :
[2026/01/29 14:54:47] ppocr INFO: intracl : True
[2026/01/29 14:54:47] ppocr INFO: name : LKPAN
[2026/01/29 14:54:47] ppocr INFO: out_channels : 256
[2026/01/29 14:54:47] ppocr INFO: Transform : None
[2026/01/29 14:54:47] ppocr INFO: algorithm : DB
[2026/01/29 14:54:47] ppocr INFO: model_type : det
[2026/01/29 14:54:47] ppocr INFO: Eval :
[2026/01/29 14:54:47] ppocr INFO: dataset :
[2026/01/29 14:54:47] ppocr INFO: data_dir : /export/DKyProject/PaddleOCR-main/dataset/det_imges
[2026/01/29 14:54:47] ppocr INFO: label_file_list : ['/export/DKyProject/PaddleOCR-main/dataset/det_val.txt']
[2026/01/29 14:54:47] ppocr INFO: name : SimpleDataSet
[2026/01/29 14:54:47] ppocr INFO: transforms :
[2026/01/29 14:54:47] ppocr INFO: DecodeImage :
[2026/01/29 14:54:47] ppocr INFO: channel_first : False
[2026/01/29 14:54:47] ppocr INFO: img_mode : BGR
[2026/01/29 14:54:47] ppocr INFO: DetLabelEncode : None
[2026/01/29 14:54:47] ppocr INFO: DetResizeForTest : None
[2026/01/29 14:54:47] ppocr INFO: NormalizeImage :
[2026/01/29 14:54:47] ppocr INFO: mean : [0.485, 0.456, 0.406]
[2026/01/29 14:54:47] ppocr INFO: order : hwc
[2026/01/29 14:54:47] ppocr INFO: scale : 1./255.
[2026/01/29 14:54:47] ppocr INFO: std : [0.229, 0.224, 0.225]
[2026/01/29 14:54:47] ppocr INFO: ToCHWImage : None
[2026/01/29 14:54:47] ppocr INFO: KeepKeys :
[2026/01/29 14:54:47] ppocr INFO: keep_keys : ['image', 'shape', 'polys', 'ignore_tags']
[2026/01/29 14:54:47] ppocr INFO: loader :
[2026/01/29 14:54:47] ppocr INFO: batch_size_per_card : 1
[2026/01/29 14:54:47] ppocr INFO: drop_last : False
[2026/01/29 14:54:47] ppocr INFO: num_workers : 2
[2026/01/29 14:54:47] ppocr INFO: shuffle : False
[2026/01/29 14:54:47] ppocr INFO: Global :
[2026/01/29 14:54:47] ppocr INFO: cal_metric_during_train : False
[2026/01/29 14:54:47] ppocr INFO: checkpoints : None
[2026/01/29 14:54:47] ppocr INFO: debug : False
[2026/01/29 14:54:47] ppocr INFO: distributed : False
[2026/01/29 14:54:47] ppocr INFO: epoch_num : 500
[2026/01/29 14:54:47] ppocr INFO: eval_batch_step : [0, 1500]
[2026/01/29 14:54:47] ppocr INFO: infer_img : doc/imgs_en/img_10.jpg
[2026/01/29 14:54:47] ppocr INFO: log_smooth_window : 20
[2026/01/29 14:54:47] ppocr INFO: model_name : PP-OCRv5_server_det
[2026/01/29 14:54:47] ppocr INFO: pretrained_model : ./PP-OCRv5_server_det_pretrained.pdparams
[2026/01/29 14:54:47] ppocr INFO: print_batch_step : 10
[2026/01/29 14:54:47] ppocr INFO: save_epoch_step : 10
[2026/01/29 14:54:47] ppocr INFO: save_inference_dir : None
[2026/01/29 14:54:47] ppocr INFO: save_model_dir : ./output/PP-OCRv5_server_det
[2026/01/29 14:54:47] ppocr INFO: save_res_path : ./checkpoints/det_db/predicts_db.txt
[2026/01/29 14:54:47] ppocr INFO: use_gpu : True
[2026/01/29 14:54:47] ppocr INFO: use_visualdl : False
[2026/01/29 14:54:47] ppocr INFO: Loss :
[2026/01/29 14:54:47] ppocr INFO: alpha : 5
[2026/01/29 14:54:47] ppocr INFO: balance_loss : True
[2026/01/29 14:54:47] ppocr INFO: beta : 10
[2026/01/29 14:54:47] ppocr INFO: main_loss_type : DiceLoss
[2026/01/29 14:54:47] ppocr INFO: name : DBLoss
[2026/01/29 14:54:47] ppocr INFO: ohem_ratio : 3
[2026/01/29 14:54:47] ppocr INFO: Metric :
[2026/01/29 14:54:47] ppocr INFO: main_indicator : hmean
[2026/01/29 14:54:47] ppocr INFO: name : DetMetric
[2026/01/29 14:54:47] ppocr INFO: Optimizer :
[2026/01/29 14:54:47] ppocr INFO: beta1 : 0.9
[2026/01/29 14:54:47] ppocr INFO: beta2 : 0.999
[2026/01/29 14:54:47] ppocr INFO: lr :
[2026/01/29 14:54:47] ppocr INFO: learning_rate : 0.001
[2026/01/29 14:54:47] ppocr INFO: name : Cosine
[2026/01/29 14:54:47] ppocr INFO: warmup_epoch : 2
[2026/01/29 14:54:47] ppocr INFO: name : Adam
[2026/01/29 14:54:47] ppocr INFO: regularizer :
[2026/01/29 14:54:47] ppocr INFO: factor : 1e-6
[2026/01/29 14:54:47] ppocr INFO: name : L2
[2026/01/29 14:54:47] ppocr INFO: PostProcess :
[2026/01/29 14:54:47] ppocr INFO: box_thresh : 0.6
[2026/01/29 14:54:47] ppocr INFO: max_candidates : 1000
[2026/01/29 14:54:47] ppocr INFO: name : DBPostProcess
[2026/01/29 14:54:47] ppocr INFO: thresh : 0.3
[2026/01/29 14:54:47] ppocr INFO: unclip_ratio : 1.5
[2026/01/29 14:54:47] ppocr INFO: Train :
[2026/01/29 14:54:47] ppocr INFO: dataset :
[2026/01/29 14:54:47] ppocr INFO: data_dir : /export/DKyProject/PaddleOCR-main/dataset/det_imges
[2026/01/29 14:54:47] ppocr INFO: label_file_list : ['/export/DKyProject/PaddleOCR-main/dataset/det_train.txt']
[2026/01/29 14:54:47] ppocr INFO: name : SimpleDataSet
[2026/01/29 14:54:47] ppocr INFO: ratio_list : [1.0]
[2026/01/29 14:54:47] ppocr INFO: transforms :
[2026/01/29 14:54:47] ppocr INFO: DecodeImage :
[2026/01/29 14:54:47] ppocr INFO: channel_first : False
[2026/01/29 14:54:47] ppocr INFO: img_mode : BGR
[2026/01/29 14:54:47] ppocr INFO: DetLabelEncode : None
[2026/01/29 14:54:47] ppocr INFO: CopyPaste : None
[2026/01/29 14:54:47] ppocr INFO: IaaAugment :
[2026/01/29 14:54:47] ppocr INFO: augmenter_args :
[2026/01/29 14:54:47] ppocr INFO: args :
[2026/01/29 14:54:47] ppocr INFO: p : 0.5
[2026/01/29 14:54:47] ppocr INFO: type : Fliplr
[2026/01/29 14:54:47] ppocr INFO: args :
[2026/01/29 14:54:47] ppocr INFO: rotate : [-10, 10]
[2026/01/29 14:54:47] ppocr INFO: type : Affine
[2026/01/29 14:54:47] ppocr INFO: args :
[2026/01/29 14:54:47] ppocr INFO: size : [0.5, 3]
[2026/01/29 14:54:47] ppocr INFO: type : Resize
[2026/01/29 14:54:47] ppocr INFO: EastRandomCropData :
[2026/01/29 14:54:47] ppocr INFO: keep_ratio : True
[2026/01/29 14:54:47] ppocr INFO: max_tries : 50
[2026/01/29 14:54:47] ppocr INFO: size : [640, 640]
[2026/01/29 14:54:47] ppocr INFO: MakeBorderMap :
[2026/01/29 14:54:47] ppocr INFO: shrink_ratio : 0.4
[2026/01/29 14:54:47] ppocr INFO: thresh_max : 0.7
[2026/01/29 14:54:47] ppocr INFO: thresh_min : 0.3
[2026/01/29 14:54:47] ppocr INFO: total_epoch : 500
[2026/01/29 14:54:47] ppocr INFO: MakeShrinkMap :
[2026/01/29 14:54:47] ppocr INFO: min_text_size : 8
[2026/01/29 14:54:47] ppocr INFO: shrink_ratio : 0.4
[2026/01/29 14:54:47] ppocr INFO: total_epoch : 500
[2026/01/29 14:54:47] ppocr INFO: NormalizeImage :
[2026/01/29 14:54:47] ppocr INFO: mean : [0.485, 0.456, 0.406]
[2026/01/29 14:54:47] ppocr INFO: order : hwc
[2026/01/29 14:54:47] ppocr INFO: scale : 1./255.
[2026/01/29 14:54:47] ppocr INFO: std : [0.229, 0.224, 0.225]
[2026/01/29 14:54:47] ppocr INFO: ToCHWImage : None
[2026/01/29 14:54:47] ppocr INFO: KeepKeys :
[2026/01/29 14:54:47] ppocr INFO: keep_keys : ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask']
[2026/01/29 14:54:47] ppocr INFO: loader :
[2026/01/29 14:54:47] ppocr INFO: batch_size_per_card : 8
[2026/01/29 14:54:47] ppocr INFO: drop_last : False
[2026/01/29 14:54:47] ppocr INFO: num_workers : 8
[2026/01/29 14:54:47] ppocr INFO: shuffle : True
[2026/01/29 14:54:47] ppocr INFO: profiler_options : None
[2026/01/29 14:54:47] ppocr INFO: train with paddle 3.2.2 and device Place(gpu:0)
[2026/01/29 14:54:47] ppocr INFO: Initialize indexes of datasets:['/export/DKyProject/PaddleOCR-main/dataset/det_train.txt']
[2026/01/29 14:54:47] ppocr INFO: Initialize indexes of datasets:['/export/DKyProject/PaddleOCR-main/dataset/det_val.txt']
[2026/01/29 14:54:47] ppocr INFO: train dataloader has 114 iters
[2026/01/29 14:54:47] ppocr INFO: valid dataloader has 267 iters
[2026/01/29 14:54:47] ppocr INFO: load pretrain successful from ./PP-OCRv5_server_det_pretrained
[2026/01/29 14:54:47] ppocr INFO: During the training process, after the 0th iteration, an evaluation is run every 1500 iterations
[2026/01/29 14:54:52] ppocr INFO: epoch: [1/500], global_step: 10, lr: 0.000020, loss: 6.092871, loss_shrink_maps: 3.494185, loss_threshold_maps: 1.164770, loss_binary_maps: 0.711984, loss_cbn: 0.711984, avg_reader_cost: 0.13922 s, avg_batch_cost: 0.43454 s, avg_samples: 8.0, ips: 18.41029 samples/s, eta: 6:52:44, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:54:54] ppocr INFO: epoch: [1/500], global_step: 20, lr: 0.000042, loss: 5.290680, loss_shrink_maps: 2.953744, loss_threshold_maps: 1.133860, loss_binary_maps: 0.592265, loss_cbn: 0.592265, avg_reader_cost: 0.00192 s, avg_batch_cost: 0.24098 s, avg_samples: 8.0, ips: 33.19813 samples/s, eta: 5:20:45, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:54:57] ppocr INFO: epoch: [1/500], global_step: 30, lr: 0.000086, loss: 4.516792, loss_shrink_maps: 2.475835, loss_threshold_maps: 1.064550, loss_binary_maps: 0.497792, loss_cbn: 0.497792, avg_reader_cost: 0.00203 s, avg_batch_cost: 0.24145 s, avg_samples: 8.0, ips: 33.13362 samples/s, eta: 4:50:13, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:54:59] ppocr INFO: epoch: [1/500], global_step: 40, lr: 0.000129, loss: 3.353877, loss_shrink_maps: 1.722735, loss_threshold_maps: 0.986250, loss_binary_maps: 0.347502, loss_cbn: 0.347502, avg_reader_cost: 0.00269 s, avg_batch_cost: 0.24115 s, avg_samples: 8.0, ips: 33.17408 samples/s, eta: 4:34:51, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:01] ppocr INFO: epoch: [1/500], global_step: 50, lr: 0.000173, loss: 2.773363, loss_shrink_maps: 1.349964, loss_threshold_maps: 0.914764, loss_binary_maps: 0.267309, loss_cbn: 0.267309, avg_reader_cost: 0.00258 s, avg_batch_cost: 0.24275 s, avg_samples: 8.0, ips: 32.95578 samples/s, eta: 4:25:55, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:04] ppocr INFO: epoch: [1/500], global_step: 60, lr: 0.000217, loss: 2.578005, loss_shrink_maps: 1.244271, loss_threshold_maps: 0.839844, loss_binary_maps: 0.248888, loss_cbn: 0.248888, avg_reader_cost: 0.00309 s, avg_batch_cost: 0.24502 s, avg_samples: 8.0, ips: 32.65059 samples/s, eta: 4:20:19, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:06] ppocr INFO: epoch: [1/500], global_step: 70, lr: 0.000261, loss: 2.312374, loss_shrink_maps: 1.090576, loss_threshold_maps: 0.808326, loss_binary_maps: 0.218202, loss_cbn: 0.218202, avg_reader_cost: 0.00355 s, avg_batch_cost: 0.24509 s, avg_samples: 8.0, ips: 32.64082 samples/s, eta: 4:16:19, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:09] ppocr INFO: epoch: [1/500], global_step: 80, lr: 0.000305, loss: 2.250500, loss_shrink_maps: 1.047191, loss_threshold_maps: 0.777106, loss_binary_maps: 0.208833, loss_cbn: 0.208833, avg_reader_cost: 0.00331 s, avg_batch_cost: 0.24313 s, avg_samples: 8.0, ips: 32.90428 samples/s, eta: 4:13:04, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:11] ppocr INFO: epoch: [1/500], global_step: 90, lr: 0.000349, loss: 2.065512, loss_shrink_maps: 0.952183, loss_threshold_maps: 0.748836, loss_binary_maps: 0.189971, loss_cbn: 0.189971, avg_reader_cost: 0.00316 s, avg_batch_cost: 0.24263 s, avg_samples: 8.0, ips: 32.97240 samples/s, eta: 4:10:28, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:14] ppocr INFO: epoch: [1/500], global_step: 100, lr: 0.000393, loss: 2.023730, loss_shrink_maps: 0.910194, loss_threshold_maps: 0.739576, loss_binary_maps: 0.182733, loss_cbn: 0.182733, avg_reader_cost: 0.00246 s, avg_batch_cost: 0.24099 s, avg_samples: 8.0, ips: 33.19699 samples/s, eta: 4:08:14, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:16] ppocr INFO: epoch: [1/500], global_step: 110, lr: 0.000436, loss: 1.957737, loss_shrink_maps: 0.880228, loss_threshold_maps: 0.729160, loss_binary_maps: 0.177410, loss_cbn: 0.177410, avg_reader_cost: 0.00205 s, avg_batch_cost: 0.24093 s, avg_samples: 8.0, ips: 33.20463 samples/s, eta: 4:06:24, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:17] ppocr INFO: epoch: [1/500], global_step: 114, lr: 0.000454, loss: 1.929638, loss_shrink_maps: 0.861826, loss_threshold_maps: 0.724089, loss_binary_maps: 0.174014, loss_cbn: 0.174014, avg_reader_cost: 0.00089 s, avg_batch_cost: 0.09337 s, avg_samples: 3.1, ips: 33.20032 samples/s, eta: 4:05:30, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:17] ppocr INFO: save model in ./output/PP-OCRv5_server_det/latest
[2026/01/29 14:55:20] ppocr INFO: epoch: [2/500], global_step: 120, lr: 0.000480, loss: 1.955120, loss_shrink_maps: 0.881762, loss_threshold_maps: 0.724089, loss_binary_maps: 0.176784, loss_cbn: 0.176784, avg_reader_cost: 0.14446 s, avg_batch_cost: 0.29025 s, avg_samples: 4.8, ips: 16.53730 samples/s, eta: 4:16:08, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:22] ppocr INFO: epoch: [2/500], global_step: 130, lr: 0.000524, loss: 1.909565, loss_shrink_maps: 0.857019, loss_threshold_maps: 0.705454, loss_binary_maps: 0.171491, loss_cbn: 0.171491, avg_reader_cost: 0.00196 s, avg_batch_cost: 0.24159 s, avg_samples: 8.0, ips: 33.11376 samples/s, eta: 4:14:00, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:25] ppocr INFO: epoch: [2/500], global_step: 140, lr: 0.000568, loss: 1.942170, loss_shrink_maps: 0.887979, loss_threshold_maps: 0.704286, loss_binary_maps: 0.177524, loss_cbn: 0.177524, avg_reader_cost: 0.00332 s, avg_batch_cost: 0.24451 s, avg_samples: 8.0, ips: 32.71825 samples/s, eta: 4:12:22, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:27] ppocr INFO: epoch: [2/500], global_step: 150, lr: 0.000612, loss: 2.001076, loss_shrink_maps: 0.928299, loss_threshold_maps: 0.710574, loss_binary_maps: 0.186207, loss_cbn: 0.186207, avg_reader_cost: 0.00260 s, avg_batch_cost: 0.24557 s, avg_samples: 8.0, ips: 32.57761 samples/s, eta: 4:11:01, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:31] ppocr INFO: epoch: [2/500], global_step: 160, lr: 0.000656, loss: 1.966648, loss_shrink_maps: 0.898502, loss_threshold_maps: 0.711659, loss_binary_maps: 0.179243, loss_cbn: 0.179243, avg_reader_cost: 0.10427 s, avg_batch_cost: 0.34426 s, avg_samples: 8.0, ips: 23.23802 samples/s, eta: 4:15:40, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:33] ppocr INFO: epoch: [2/500], global_step: 170, lr: 0.000700, loss: 1.870944, loss_shrink_maps: 0.838800, loss_threshold_maps: 0.686665, loss_binary_maps: 0.167932, loss_cbn: 0.167932, avg_reader_cost: 0.00206 s, avg_batch_cost: 0.24304 s, avg_samples: 8.0, ips: 32.91608 samples/s, eta: 4:14:07, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:36] ppocr INFO: epoch: [2/500], global_step: 180, lr: 0.000743, loss: 1.828708, loss_shrink_maps: 0.820399, loss_threshold_maps: 0.666074, loss_binary_maps: 0.164451, loss_cbn: 0.164451, avg_reader_cost: 0.00213 s, avg_batch_cost: 0.24136 s, avg_samples: 8.0, ips: 33.14613 samples/s, eta: 4:12:40, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:38] ppocr INFO: epoch: [2/500], global_step: 190, lr: 0.000787, loss: 1.858571, loss_shrink_maps: 0.835134, loss_threshold_maps: 0.677803, loss_binary_maps: 0.166939, loss_cbn: 0.166939, avg_reader_cost: 0.00309 s, avg_batch_cost: 0.24304 s, avg_samples: 8.0, ips: 32.91606 samples/s, eta: 4:11:26, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:40] ppocr INFO: epoch: [2/500], global_step: 200, lr: 0.000831, loss: 1.864131, loss_shrink_maps: 0.847274, loss_threshold_maps: 0.689748, loss_binary_maps: 0.169673, loss_cbn: 0.169673, avg_reader_cost: 0.00177 s, avg_batch_cost: 0.24270 s, avg_samples: 8.0, ips: 32.96288 samples/s, eta: 4:10:18, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:43] ppocr INFO: epoch: [2/500], global_step: 210, lr: 0.000875, loss: 1.851803, loss_shrink_maps: 0.830013, loss_threshold_maps: 0.687670, loss_binary_maps: 0.166013, loss_cbn: 0.166013, avg_reader_cost: 0.00185 s, avg_batch_cost: 0.24180 s, avg_samples: 8.0, ips: 33.08488 samples/s, eta: 4:09:15, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:45] ppocr INFO: epoch: [2/500], global_step: 220, lr: 0.000919, loss: 1.791664, loss_shrink_maps: 0.813191, loss_threshold_maps: 0.676215, loss_binary_maps: 0.161869, loss_cbn: 0.161869, avg_reader_cost: 0.00198 s, avg_batch_cost: 0.24120 s, avg_samples: 8.0, ips: 33.16742 samples/s, eta: 4:08:15, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:47] ppocr INFO: epoch: [2/500], global_step: 228, lr: 0.000954, loss: 1.761136, loss_shrink_maps: 0.765036, loss_threshold_maps: 0.662249, loss_binary_maps: 0.152699, loss_cbn: 0.152699, avg_reader_cost: 0.00154 s, avg_batch_cost: 0.18907 s, avg_samples: 6.3, ips: 33.32077 samples/s, eta: 4:07:21, max_mem_reserved: 12330 MB, max_mem_allocated: 10314 MB
[2026/01/29 14:55:48] ppocr INFO: save model in ./output/PP-OCRv5_server_det/latest
...
模型评估
python3 tools/eval.py -c configs/det/PP-OCRv5/PP-OCRv5_server_det.yml
-o Global.pretrained_model=/export/DKyProject/PaddleOCR-main/output/PP-OCRv5_server_det/iter_epoch_380.pdparams
Eval.dataset.data_dir=/export/DKyProject/PaddleOCR-main/dataset/det_imges
Eval.dataset.label_file_list='[/export/DKyProject/PaddleOCR-main/dataset/det_train.txt]'

W0130 14:42:52.648741 57079 gpu_resources.cc:114] Please NOTE: device: 0, GPU Compute Capability: 8.9, Driver API Version: 12.4, Runtime API Version: 11.8
[2026/01/30 14:42:53] ppocr INFO: load pretrain successful from /export/DKyProject/PaddleOCR-main/output/PP-OCRv5_server_det/iter_epoch_380
[2026/01/30 14:42:53] ppocr INFO: metric in ckpt ***************
[2026/01/30 14:42:53] ppocr INFO: is_float16:False
eval model:: 0%| | 0/911 [00:00<?, ?it/s]W0130 14:42:53.386711 57079 gpu_resources.cc:243] WARNING: device: 0. The installed Paddle is compiled with CUDNN 8.9, but CUDNN version in your machine is 8.9, which may cause serious incompatible bug. Please recompile or reinstall Paddle with compatible CUDNN version.
eval model:: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 911/911 [00:24<00:00, 36.44it/s]
[2026/01/30 14:43:18] ppocr INFO: metric eval ***************
[2026/01/30 14:43:18] ppocr INFO: precision:0.9554831704668838
[2026/01/30 14:43:18] ppocr INFO: recall:0.9659714599341384
[2026/01/30 14:43:18] ppocr INFO: hmean:0.9606986899563319
[2026/01/30 14:43:18] ppocr INFO: fps:42.2227861193049

模型导出
python3 tools/export_model.py -c configs/det/PP-OCRv5/PP-OCRv5_server_det.yml -o
Global.pretrained_model=/export/DKyProject/PaddleOCR-main/output/PP-OCRv5_server_det/iter_epoch_380.pdparams
Global.save_inference_dir="./PP-OCRv5_server_det_infer/"

which: no ccache in (/root/anaconda3/envs/Paddle-doc-parser/bin:/root/anaconda3/condabin:/usr/local/gcc-9.4.0/bin:/opt/libreoffice7.4/program:/usr/local/cuda-12.1/bin:/usr/lib64/qt-3.3/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/export/server/lcd_bdzsgx/jdk/jdk1.8.0_202/bin:/usr/local/yasm/bin:/opt/ffmpeg-7.0.2-amd64-static:/root/anaconda3/bin:/root/bin)
/root/anaconda3/envs/Paddle-doc-parser/lib/python3.10/site-packages/paddle/utils/cpp_extension/extension_utils.py:718: UserWarning: No ccache found. Please be aware that recompiling all source files may be required. You can download and install ccache from: https://github.com/ccache/ccache/blob/master/doc/INSTALL.md
warnings.warn(warning_message)
Skipping import of the encryption module.
W0130 14:44:02.925645 59524 gpu_resources.cc:114] Please NOTE: device: 0, GPU Compute Capability: 8.9, Driver API Version: 12.4, Runtime API Version: 11.8
[2026/01/30 14:44:03] ppocr INFO: load pretrain successful from /export/DKyProject/PaddleOCR-main/output/PP-OCRv5_server_det/iter_epoch_380
[2026/01/30 14:44:03] ppocr INFO: Export inference config file to ./PP-OCRv5_server_det_infer/inference.yml
Skipping import of the encryption module
W0130 14:44:04.828503 59524 eager_utils.cc:3441] Paddle static graph(PIR) not support input out tensor for now!!!!!
[2026/01/30 14:44:06] ppocr INFO: inference model is saved to ./PP-OCRv5_server_det_infer/inference

使用微调后模型
from paddleocr import TextDetection
model = TextDetection(model_name="PP-OCRv5_server_det",model_dir = "/export/DKyProject/PaddleOCR-main/PP-OCRv5_server_det_infer")
output = model.predict("get_scale_value_1_0232fc33-7138-4c07-a16c-84ae8a5c87a8.png", batch_size=1)
for res in output:
res.print()

输出：

No model hoster is available! Please check your network connection to one of the following model hosts: HuggingFace (https://huggingface.co), ModelScope (https://modelscope.cn), AIStudio (https://aistudio.baidu.com), or BOS (https://paddle-model-ecology.bj.bcebos.com). Otherwise, only local models can be used.
which: no ccache in (/root/anaconda3/envs/Paddle-doc-parser/bin:/root/anaconda3/condabin:/usr/local/gcc-9.4.0/bin:/opt/libreoffice7.4/program:/usr/local/cuda-12.1/bin:/usr/lib64/qt-3.3/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/export/server/lcd_bdzsgx/jdk/jdk1.8.0_202/bin:/usr/local/yasm/bin:/opt/ffmpeg-7.0.2-amd64-static:/root/anaconda3/bin:/root/bin)
/root/anaconda3/envs/Paddle-doc-parser/lib/python3.10/site-packages/paddle/utils/cpp_extension/extension_utils.py:718: UserWarning: No ccache found. Please be aware that recompiling all source files may be required. You can download and install ccache from: https://github.com/ccache/ccache/blob/master/doc/INSTALL.md
warnings.warn(warning_message)
{'res': {'input_path': 'get_scale_value_1_0232fc33-7138-4c07-a16c-84ae8a5c87a8.png', 'page_index': None, 'dt_polys': array([], dtype=int16), 'dt_scores': []}}

没有任何效果

[wulawulaaaa]

请问你这个解决了吗

[sf9ehf9fe]

没有 我初步排查是训练方法的问题&nbsp; 你可以看看paddleocr和paddlex中的det模型训练方法。

…

---原始邮件--- 发件人: ***@***.***&gt; 发送时间: 2026年2月1日(周日) 下午2:26 收件人: ***@***.***&gt;; 抄送: "Mr. ***@***.******@***.***&gt;; 主题: Re: [PaddlePaddle/PaddleOCR] PP-OCRv5_server_det 模型微调后效果更差 (Discussion #17584) 请问你这个解决了吗 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***&gt;