训练时，ppocr的打印未出现eval，只是一直save model并且最后打印best metric, hmean: 0, is_float16: False #14951

Jasonchanlan · 2025-03-28T09:46:11Z

Jasonchanlan
Mar 28, 2025

python tools/train.py -c configs/det/ch_PP-OCRv4/ch_PP-OCRv4_det_student.yml
信息: 用提供的模式无法找到文件。
E:\anconda\envs\paddleOCR\lib\site-packages\paddle\utils\cpp_extension\extension_utils.py:711: 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)
[2025/03/28 17:45:00] ppocr WARNING: Skipping import of the encryption module.
[2025/03/28 17:45:00] ppocr INFO: Architecture :
[2025/03/28 17:45:00] ppocr INFO: Backbone :
[2025/03/28 17:45:00] ppocr INFO: det : True
[2025/03/28 17:45:00] ppocr INFO: name : PPLCNetV3
[2025/03/28 17:45:00] ppocr INFO: scale : 0.75
[2025/03/28 17:45:00] ppocr INFO: Head :
[2025/03/28 17:45:00] ppocr INFO: fix_nan : True
[2025/03/28 17:45:00] ppocr INFO: k : 50
[2025/03/28 17:45:00] ppocr INFO: name : DBHead
[2025/03/28 17:45:00] ppocr INFO: Neck :
[2025/03/28 17:45:00] ppocr INFO: name : RSEFPN
[2025/03/28 17:45:00] ppocr INFO: out_channels : 96
[2025/03/28 17:45:00] ppocr INFO: shortcut : True
[2025/03/28 17:45:00] ppocr INFO: Transform : None
[2025/03/28 17:45:00] ppocr INFO: algorithm : DB
[2025/03/28 17:45:00] ppocr INFO: model_type : det
[2025/03/28 17:45:00] ppocr INFO: Eval :
[2025/03/28 17:45:00] ppocr INFO: dataset :
[2025/03/28 17:45:00] ppocr INFO: data_dir : ./train_data/
[2025/03/28 17:45:00] ppocr INFO: label_file_list : ['./train_data/det/val.txt']
[2025/03/28 17:45:00] ppocr INFO: name : SimpleDataSet
[2025/03/28 17:45:00] ppocr INFO: transforms :
[2025/03/28 17:45:00] ppocr INFO: DecodeImage :
[2025/03/28 17:45:00] ppocr INFO: channel_first : False
[2025/03/28 17:45:00] ppocr INFO: img_mode : BGR
[2025/03/28 17:45:00] ppocr INFO: DetLabelEncode : None
[2025/03/28 17:45:00] ppocr INFO: DetResizeForTest : None
[2025/03/28 17:45:00] ppocr INFO: NormalizeImage :
[2025/03/28 17:45:00] ppocr INFO: mean : [0.485, 0.456, 0.406]
[2025/03/28 17:45:00] ppocr INFO: order : hwc
[2025/03/28 17:45:00] ppocr INFO: scale : 1./255.
[2025/03/28 17:45:00] ppocr INFO: std : [0.229, 0.224, 0.225]
[2025/03/28 17:45:00] ppocr INFO: ToCHWImage : None
[2025/03/28 17:45:00] ppocr INFO: KeepKeys :
[2025/03/28 17:45:00] ppocr INFO: keep_keys : ['image', 'shape', 'polys', 'ignore_tags']
[2025/03/28 17:45:00] ppocr INFO: loader :
[2025/03/28 17:45:00] ppocr INFO: batch_size_per_card : 1
[2025/03/28 17:45:00] ppocr INFO: drop_last : False
[2025/03/28 17:45:00] ppocr INFO: num_workers : 2
[2025/03/28 17:45:00] ppocr INFO: shuffle : False
[2025/03/28 17:45:00] ppocr INFO: Global :
[2025/03/28 17:45:00] ppocr INFO: cal_metric_during_train : True
[2025/03/28 17:45:00] ppocr INFO: checkpoints : None
[2025/03/28 17:45:00] ppocr INFO: d2s_train_image_shape : [3, 640, 640]
[2025/03/28 17:45:00] ppocr INFO: debug : True
[2025/03/28 17:45:00] ppocr INFO: distributed : False
[2025/03/28 17:45:00] ppocr INFO: epoch_num : 10
[2025/03/28 17:45:00] ppocr INFO: eval_batch_step : [0, 5]
[2025/03/28 17:45:00] ppocr INFO: infer_img : doc/imgs_en/img_10.jpg
[2025/03/28 17:45:00] ppocr INFO: log_smooth_window : 20
[2025/03/28 17:45:00] ppocr INFO: pretrained_model : ./pretrain_models/ch_PP-OCRv4_det_train/best_accuracy
[2025/03/28 17:45:00] ppocr INFO: print_batch_step : 2
[2025/03/28 17:45:00] ppocr INFO: save_epoch_step : 2
[2025/03/28 17:45:00] ppocr INFO: save_inference_dir : None
[2025/03/28 17:45:00] ppocr INFO: save_model_dir : ./output/ch_PP-OCRv4
[2025/03/28 17:45:00] ppocr INFO: save_res_path : ./checkpoints/det_db/predicts_db.txt
[2025/03/28 17:45:00] ppocr INFO: use_gpu : False
[2025/03/28 17:45:00] ppocr INFO: use_visualdl : False
[2025/03/28 17:45:00] ppocr INFO: Loss :
[2025/03/28 17:45:00] ppocr INFO: alpha : 5
[2025/03/28 17:45:00] ppocr INFO: balance_loss : True
[2025/03/28 17:45:00] ppocr INFO: beta : 10
[2025/03/28 17:45:00] ppocr INFO: main_loss_type : DiceLoss
[2025/03/28 17:45:00] ppocr INFO: name : DBLoss
[2025/03/28 17:45:00] ppocr INFO: ohem_ratio : 3
[2025/03/28 17:45:00] ppocr INFO: Metric :
[2025/03/28 17:45:00] ppocr INFO: main_indicator : hmean
[2025/03/28 17:45:00] ppocr INFO: name : DetMetric
[2025/03/28 17:45:00] ppocr INFO: Optimizer :
[2025/03/28 17:45:00] ppocr INFO: beta1 : 0.9
[2025/03/28 17:45:00] ppocr INFO: beta2 : 0.999
[2025/03/28 17:45:00] ppocr INFO: lr :
[2025/03/28 17:45:00] ppocr INFO: learning_rate : 0.001
[2025/03/28 17:45:00] ppocr INFO: name : Cosine
[2025/03/28 17:45:00] ppocr INFO: warmup_epoch : 2
[2025/03/28 17:45:00] ppocr INFO: name : Adam
[2025/03/28 17:45:00] ppocr INFO: regularizer :
[2025/03/28 17:45:00] ppocr INFO: factor : 5e-05
[2025/03/28 17:45:00] ppocr INFO: name : L2
[2025/03/28 17:45:00] ppocr INFO: PostProcess :
[2025/03/28 17:45:00] ppocr INFO: box_thresh : 0.6
[2025/03/28 17:45:00] ppocr INFO: max_candidates : 1000
[2025/03/28 17:45:00] ppocr INFO: name : DBPostProcess
[2025/03/28 17:45:00] ppocr INFO: thresh : 0.3
[2025/03/28 17:45:00] ppocr INFO: unclip_ratio : 1.5
[2025/03/28 17:45:00] ppocr INFO: Train :
[2025/03/28 17:45:00] ppocr INFO: dataset :
[2025/03/28 17:45:00] ppocr INFO: data_dir : ./train_data/
[2025/03/28 17:45:00] ppocr INFO: label_file_list : ['./train_data/det/train.txt']
[2025/03/28 17:45:00] ppocr INFO: name : SimpleDataSet
[2025/03/28 17:45:00] ppocr INFO: ratio_list : [1.0]
[2025/03/28 17:45:00] ppocr INFO: transforms :
[2025/03/28 17:45:00] ppocr INFO: DecodeImage :
[2025/03/28 17:45:00] ppocr INFO: channel_first : False
[2025/03/28 17:45:00] ppocr INFO: img_mode : BGR
[2025/03/28 17:45:00] ppocr INFO: DetLabelEncode : None
[2025/03/28 17:45:00] ppocr INFO: CopyPaste : None
[2025/03/28 17:45:00] ppocr INFO: IaaAugment :
[2025/03/28 17:45:00] ppocr INFO: augmenter_args :
[2025/03/28 17:45:00] ppocr INFO: args :
[2025/03/28 17:45:00] ppocr INFO: p : 0.5
[2025/03/28 17:45:00] ppocr INFO: type : Fliplr
[2025/03/28 17:45:00] ppocr INFO: args :
[2025/03/28 17:45:00] ppocr INFO: rotate : [-10, 10]
[2025/03/28 17:45:00] ppocr INFO: type : Affine
[2025/03/28 17:45:00] ppocr INFO: args :
[2025/03/28 17:45:00] ppocr INFO: size : [0.5, 3]
[2025/03/28 17:45:00] ppocr INFO: type : Resize
[2025/03/28 17:45:00] ppocr INFO: EastRandomCropData :
[2025/03/28 17:45:00] ppocr INFO: keep_ratio : True
[2025/03/28 17:45:00] ppocr INFO: max_tries : 50
[2025/03/28 17:45:00] ppocr INFO: size : [640, 640]
[2025/03/28 17:45:00] ppocr INFO: MakeBorderMap :
[2025/03/28 17:45:00] ppocr INFO: shrink_ratio : 0.4
[2025/03/28 17:45:00] ppocr INFO: thresh_max : 0.7
[2025/03/28 17:45:00] ppocr INFO: thresh_min : 0.3
[2025/03/28 17:45:00] ppocr INFO: total_epoch : 10
[2025/03/28 17:45:00] ppocr INFO: MakeShrinkMap :
[2025/03/28 17:45:00] ppocr INFO: min_text_size : 8
[2025/03/28 17:45:00] ppocr INFO: shrink_ratio : 0.4
[2025/03/28 17:45:00] ppocr INFO: total_epoch : 10
[2025/03/28 17:45:00] ppocr INFO: NormalizeImage :
[2025/03/28 17:45:00] ppocr INFO: mean : [0.485, 0.456, 0.406]
[2025/03/28 17:45:00] ppocr INFO: order : hwc
[2025/03/28 17:45:00] ppocr INFO: scale : 1./255.
[2025/03/28 17:45:00] ppocr INFO: std : [0.229, 0.224, 0.225]
[2025/03/28 17:45:00] ppocr INFO: ToCHWImage : None
[2025/03/28 17:45:00] ppocr INFO: KeepKeys :
[2025/03/28 17:45:00] ppocr INFO: keep_keys : ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask']
[2025/03/28 17:45:00] ppocr INFO: loader :
[2025/03/28 17:45:00] ppocr INFO: batch_size_per_card : 8
[2025/03/28 17:45:00] ppocr INFO: drop_last : False
[2025/03/28 17:45:00] ppocr INFO: num_workers : 8
[2025/03/28 17:45:00] ppocr INFO: shuffle : True
[2025/03/28 17:45:00] ppocr INFO: profiler_options : None
[2025/03/28 17:45:00] ppocr INFO: train with paddle 3.0.0-rc1 and device Place(cpu)
[2025/03/28 17:45:00] ppocr INFO: Initialize indexes of datasets:['./train_data/det/train.txt']
[2025/03/28 17:45:00] ppocr INFO: Initialize indexes of datasets:['./train_data/det/val.txt']
[2025/03/28 17:45:00] ppocr INFO: train dataloader has 1 iters
[2025/03/28 17:45:00] ppocr INFO: valid dataloader has 2 iters
[2025/03/28 17:45:00] ppocr WARNING: The pretrained params backbone.conv1.hardswish.scale not in model
[2025/03/28 17:45:00] ppocr WARNING: The pretrained params backbone.conv1.hardswish.bias not in model
[2025/03/28 17:45:00] ppocr INFO: load pretrain successful from ./pretrain_models/ch_PP-OCRv4_det_train/best_accuracy
[2025/03/28 17:45:00] ppocr INFO: During the training process, after the 0th iteration, an evaluation is run every 5 iterations
[2025/03/28 17:45:01] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:02] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:02] ppocr INFO: save model in ./output/ch_PP-OCRv4\iter_epoch_2
[2025/03/28 17:45:03] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:04] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:04] ppocr INFO: save model in ./output/ch_PP-OCRv4\iter_epoch_4
[2025/03/28 17:45:05] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:06] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:06] ppocr INFO: save model in ./output/ch_PP-OCRv4\iter_epoch_6
[2025/03/28 17:45:06] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:07] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:07] ppocr INFO: save model in ./output/ch_PP-OCRv4\iter_epoch_8
[2025/03/28 17:45:08] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:09] ppocr INFO: save model in ./output/ch_PP-OCRv4\latest
[2025/03/28 17:45:09] ppocr INFO: save model in ./output/ch_PP-OCRv4\iter_epoch_10
[2025/03/28 17:45:09] ppocr INFO: best metric, hmean: 0, is_float16: False

GreatV · 2025-03-28T09:52:13Z

GreatV
Mar 28, 2025
Maintainer

从你提供的训练日志和配置来看，训练过程中的确并没有明显进行eval（评估）的相关打印输出，最终best metric为 hmean: 0，说明整个训练过程中模型在验证集上的检测指标为0，可能的原因如下：

⚠️验证集数据标注错误或者无数据有效载入：
训练日志中提到：
[2025/03/28 17:45:00] ppocr INFO: valid dataloader has 2 iters

意味着验证集只有两条数据，数量非常少，考虑以下问题：

确保路径 ./train_data/det/val.txt 中的图像路径是否正确？
检查图像是否实际存在？
检查标注格式是否和 PaddleOCR 要求一致（例如：格式为 image_path\t[polygon list]\n）；
图像中多边形标注可能被认为是被忽略区域（如全为 ignore_tags: true），则对评估无贡献；

☑️评估机制配置并未生效或被跳过：
配置中包含以下信息：
[2025/03/28 17:45:00] ppocr INFO: eval_batch_step : [0, 5]，代表每5步进行一次评估（从第0步开始），但后续日志并未见有“Hmean”或“precision/recall”等打印。这种现象可能因为以下原因：

数据集中加载的数据为0或加载失败，导致实际没有评估可进行；
设置了 debug: True，导致部分代码路径更简化，仅输出 minimal 日志对调试友好，建议将 debug 改为 False 试试看。

📉模型输出为空（或预测框全为空）导致评估指标为 0：
即便验证集数据读取成功，如果模型在验证集上的预测输出为空，则 hmean 会为0。这可能由于：

学习率过低或过高，模型未有效学习；
数据量太少（如你的 val.txt 仅有两条图片），模型无法泛化；
模型预训练参数加载不匹配，从日志可见：

[2025/03/28 17:45:00] ppocr WARNING: The pretrained params backbone.conv1.hardswish.scale not in model

这说明你加载的预训练模型和当前网络结构可能不完全一致，可能导致模型初始化不好，影响性能；

✅参考解决建议：

确保验证集数据格式正确，标注没有问题（推荐手动检查 ./train_data/det/val.txt 和对应图片）；
增加验证集数据规模；
关闭 debug 模式，确认是否打印出 val 的预测信息；
检查训练集是否足够、是否过拟合（如训练很快 loss 收敛，但 val 无预测）；
确保模型结构和预训练参数完全匹配；
如有必要，可将 cal_metric_during_train 设置为 False，改为单独评估模型（tools/eval.py）以更可控地调试验证集效果；

你可以参考下相关讨论：https://github.com/PaddlePaddle/PaddleOCR/discussions/13839，其中也有类似关于 hmean: 0 的问题。

如果进一步检查后仍有问题，建议尝试使用 paddleOCR 提供的预训练数据集进行测试，以排除数据兼容性问题。

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.

0 replies

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

训练时，ppocr的打印未出现eval，只是一直save model并且最后打印best metric, hmean: 0, is_float16: False #14951

Uh oh!

{{title}}

Uh oh!

Replies: 1 comment

Uh oh!

{{title}}

Uh oh!

Select a reply

Uh oh!

训练时，ppocr的打印未出现eval，只是一直save model并且最后打印best metric, hmean: 0, is_float16: False #14951

Uh oh!

Jasonchanlan Mar 28, 2025

Replies: 1 comment

Uh oh!

GreatV Mar 28, 2025 Maintainer

Jasonchanlan
Mar 28, 2025

GreatV
Mar 28, 2025
Maintainer