v4 en-rec模型微调loss为0 #12353

iPengXPro · 2024-02-26T14:48:32Z

iPengXPro
Feb 26, 2024

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：win10
版本号/Version：Paddle： PaddleOCR：问题相关组件/Related components：ppocr2.7
运行指令/Command Code：D:\PYTHON\Anaconda3\envs\pd\python.exe D:\PYTHON\codes\PaddleOCR-release-2.7\train-muhao.py -c .\configs\rec\PP-OCRv4\en_PP-OCRv4_rec.yml -o Global.pretrained_model=./pre_train_models/en_PP-OCRv4_rec_train/best_accuracy
完整报错/Complete Error Message：训练几个epoch acc降为0，其他参数变为nanxx

我们提供了AceIssueSolver来帮助你解答问题，你是否想要它来解答(请填写yes/no)?/We provide AceIssueSolver to solve issues, do you want it? (Please write yes/no):yes

请尽量不要包含图片在问题中/Please try to not include the image in the issue.

D:\PYTHON\Anaconda3\envs\pd\python.exe D:\PYTHON\codes\PaddleOCR-release-2.7\train.py -c .\configs\rec\PP-OCRv4\en_PP-OCRv4_rec.yml -o Global.pretrained_model=./pre_train_models/en_PP-OCRv4_rec_train/best_accuracy
[2024/02/26 22:10:06] ppocr INFO: Architecture :
[2024/02/26 22:10:06] ppocr INFO: Backbone :
[2024/02/26 22:10:06] ppocr INFO: name : PPLCNetV3
[2024/02/26 22:10:06] ppocr INFO: scale : 0.95
[2024/02/26 22:10:06] ppocr INFO: Head :
[2024/02/26 22:10:06] ppocr INFO: head_list :
[2024/02/26 22:10:06] ppocr INFO: CTCHead :
[2024/02/26 22:10:06] ppocr INFO: Head :
[2024/02/26 22:10:06] ppocr INFO: fc_decay : 1e-05
[2024/02/26 22:10:06] ppocr INFO: Neck :
[2024/02/26 22:10:06] ppocr INFO: depth : 2
[2024/02/26 22:10:06] ppocr INFO: dims : 120
[2024/02/26 22:10:06] ppocr INFO: hidden_dims : 120
[2024/02/26 22:10:06] ppocr INFO: kernel_size : [1, 3]
[2024/02/26 22:10:06] ppocr INFO: name : svtr
[2024/02/26 22:10:06] ppocr INFO: use_guide : True
[2024/02/26 22:10:06] ppocr INFO: NRTRHead :
[2024/02/26 22:10:06] ppocr INFO: max_text_length : 10
[2024/02/26 22:10:06] ppocr INFO: nrtr_dim : 384
[2024/02/26 22:10:06] ppocr INFO: name : MultiHead
[2024/02/26 22:10:06] ppocr INFO: Transform : None
[2024/02/26 22:10:06] ppocr INFO: algorithm : SVTR_LCNet
[2024/02/26 22:10:06] ppocr INFO: model_type : rec
[2024/02/26 22:10:06] ppocr INFO: Eval :
[2024/02/26 22:10:06] ppocr INFO: dataset :
[2024/02/26 22:10:06] ppocr INFO: data_dir : D:\PYTHON\pictures
[2024/02/26 22:10:06] ppocr INFO: label_file_list : ['img\rec_eval.txt']
[2024/02/26 22:10:06] ppocr INFO: name : SimpleDataSet
[2024/02/26 22:10:06] ppocr INFO: transforms :
[2024/02/26 22:10:06] ppocr INFO: DecodeImage :
[2024/02/26 22:10:06] ppocr INFO: channel_first : False
[2024/02/26 22:10:06] ppocr INFO: img_mode : BGR
[2024/02/26 22:10:06] ppocr INFO: MultiLabelEncode :
[2024/02/26 22:10:06] ppocr INFO: gtc_encode : NRTRLabelEncode
[2024/02/26 22:10:06] ppocr INFO: RecResizeImg :
[2024/02/26 22:10:06] ppocr INFO: image_shape : [3, 48, 96]
[2024/02/26 22:10:06] ppocr INFO: KeepKeys :
[2024/02/26 22:10:06] ppocr INFO: keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio']
[2024/02/26 22:10:06] ppocr INFO: loader :
[2024/02/26 22:10:06] ppocr INFO: batch_size_per_card : 128
[2024/02/26 22:10:06] ppocr INFO: drop_last : False
[2024/02/26 22:10:06] ppocr INFO: num_workers : 1
[2024/02/26 22:10:06] ppocr INFO: shuffle : False
[2024/02/26 22:10:06] ppocr INFO: Global :
[2024/02/26 22:10:06] ppocr INFO: cal_metric_during_train : True
[2024/02/26 22:10:06] ppocr INFO: character_dict_path : ppocr/utils/en_dict.txt
[2024/02/26 22:10:06] ppocr INFO: checkpoints : None
[2024/02/26 22:10:06] ppocr INFO: debug : False
[2024/02/26 22:10:06] ppocr INFO: distributed : False
[2024/02/26 22:10:06] ppocr INFO: epoch_num : 50
[2024/02/26 22:10:06] ppocr INFO: eval_batch_step : [0, 2000]
[2024/02/26 22:10:06] ppocr INFO: infer_img : doc/imgs_words/ch/word_1.jpg
[2024/02/26 22:10:06] ppocr INFO: infer_mode : False
[2024/02/26 22:10:06] ppocr INFO: log_smooth_window : 20
[2024/02/26 22:10:06] ppocr INFO: max_text_length : 10
[2024/02/26 22:10:06] ppocr INFO: pretrained_model : ./pre_train_models/en_PP-OCRv4_rec_train/best_accuracy
[2024/02/26 22:10:06] ppocr INFO: print_batch_step : 10
[2024/02/26 22:10:06] ppocr INFO: save_epoch_step : 10
[2024/02/26 22:10:06] ppocr INFO: save_inference_dir : None
[2024/02/26 22:10:06] ppocr INFO: save_model_dir : ./output/rec_ppocr_v4
[2024/02/26 22:10:06] ppocr INFO: save_res_path : ./output/rec/predicts_ppocrv3.txt
[2024/02/26 22:10:06] ppocr INFO: use_gpu : True
[2024/02/26 22:10:06] ppocr INFO: use_space_char : True
[2024/02/26 22:10:06] ppocr INFO: use_visualdl : False
[2024/02/26 22:10:06] ppocr INFO: Loss :
[2024/02/26 22:10:06] ppocr INFO: loss_config_list :
[2024/02/26 22:10:06] ppocr INFO: CTCLoss : None
[2024/02/26 22:10:06] ppocr INFO: NRTRLoss : None
[2024/02/26 22:10:06] ppocr INFO: name : MultiLoss
[2024/02/26 22:10:06] ppocr INFO: Metric :
[2024/02/26 22:10:06] ppocr INFO: ignore_space : False
[2024/02/26 22:10:06] ppocr INFO: main_indicator : acc
[2024/02/26 22:10:06] ppocr INFO: name : RecMetric
[2024/02/26 22:10:06] ppocr INFO: Optimizer :
[2024/02/26 22:10:06] ppocr INFO: beta1 : 0.9
[2024/02/26 22:10:06] ppocr INFO: beta2 : 0.999
[2024/02/26 22:10:06] ppocr INFO: lr :
[2024/02/26 22:10:06] ppocr INFO: learning_rate : 0.0005
[2024/02/26 22:10:06] ppocr INFO: name : Cosine
[2024/02/26 22:10:06] ppocr INFO: warmup_epoch : 5
[2024/02/26 22:10:06] ppocr INFO: name : Adam
[2024/02/26 22:10:06] ppocr INFO: regularizer :
[2024/02/26 22:10:06] ppocr INFO: factor : 3e-05
[2024/02/26 22:10:06] ppocr INFO: name : L2
[2024/02/26 22:10:06] ppocr INFO: PostProcess :
[2024/02/26 22:10:06] ppocr INFO: name : CTCLabelDecode
[2024/02/26 22:10:06] ppocr INFO: Train :
[2024/02/26 22:10:06] ppocr INFO: dataset :
[2024/02/26 22:10:06] ppocr INFO: data_dir : D:\PYTHON\pictures
[2024/02/26 22:10:06] ppocr INFO: ds_width : False
[2024/02/26 22:10:06] ppocr INFO: ext_op_transform_idx : 1
[2024/02/26 22:10:06] ppocr INFO: label_file_list : ['img\rec_train.txt']
[2024/02/26 22:10:06] ppocr INFO: name : MultiScaleDataSet
[2024/02/26 22:10:06] ppocr INFO: transforms :
[2024/02/26 22:10:06] ppocr INFO: DecodeImage :
[2024/02/26 22:10:06] ppocr INFO: channel_first : False
[2024/02/26 22:10:06] ppocr INFO: img_mode : BGR
[2024/02/26 22:10:06] ppocr INFO: RecConAug :
[2024/02/26 22:10:06] ppocr INFO: ext_data_num : 2
[2024/02/26 22:10:06] ppocr INFO: image_shape : [48, 96, 3]
[2024/02/26 22:10:06] ppocr INFO: max_text_length : 10
[2024/02/26 22:10:06] ppocr INFO: prob : 0.5
[2024/02/26 22:10:06] ppocr INFO: RecAug : None
[2024/02/26 22:10:06] ppocr INFO: MultiLabelEncode :
[2024/02/26 22:10:06] ppocr INFO: gtc_encode : NRTRLabelEncode
[2024/02/26 22:10:06] ppocr INFO: KeepKeys :
[2024/02/26 22:10:06] ppocr INFO: keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio']
[2024/02/26 22:10:06] ppocr INFO: loader :
[2024/02/26 22:10:06] ppocr INFO: batch_size_per_card : 96
[2024/02/26 22:10:06] ppocr INFO: drop_last : True
[2024/02/26 22:10:06] ppocr INFO: num_workers : 1
[2024/02/26 22:10:06] ppocr INFO: shuffle : True
[2024/02/26 22:10:06] ppocr INFO: sampler :
[2024/02/26 22:10:06] ppocr INFO: divided_factor : [8, 16]
[2024/02/26 22:10:06] ppocr INFO: first_bs : 96
[2024/02/26 22:10:06] ppocr INFO: fix_bs : False
[2024/02/26 22:10:06] ppocr INFO: is_training : True
[2024/02/26 22:10:06] ppocr INFO: name : MultiScaleSampler
[2024/02/26 22:10:06] ppocr INFO: scales : [[96, 32], [96, 48], [96, 64]]
[2024/02/26 22:10:06] ppocr INFO: profiler_options : None
[2024/02/26 22:10:06] ppocr INFO: train with paddle 2.3.2 and device Place(gpu:0)
[2024/02/26 22:10:06] ppocr INFO: Initialize indexs of datasets:['img\rec_train.txt']
list index out of range
[2024/02/26 22:10:06] ppocr INFO: Initialize indexs of datasets:['img\rec_eval.txt']
W0226 22:10:07.103623 23948 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 8.6, Driver API Version: 11.7, Runtime API Version: 11.2
W0226 22:10:07.163653 23948 gpu_resources.cc:91] device: 0, cuDNN Version: 8.5.
INFO 2024-02-26 22:10:19,364 optimizer.py:162] If regularizer of a Parameter has been set by 'paddle.ParamAttr' or 'static.WeightNormParamAttr' already. The weight_decay[3e-05] in Optimizer will not take effect, and it will only be applied to other Parameters!
[2024/02/26 22:10:19] ppocr INFO: train dataloader has 72 iters
[2024/02/26 22:10:19] ppocr INFO: valid dataloader has 10 iters
[2024/02/26 22:10:20] ppocr INFO: load pretrain successful from ./pre_train_models/en_PP-OCRv4_rec_train/best_accuracy
[2024/02/26 22:10:20] ppocr INFO: During the training process, after the 0th iteration, an evaluation is run every 2000 iterations
[2024/02/26 22:10:43] ppocr INFO: epoch: [1/50], global_step: 10, lr: 0.000006, acc: 0.041667, norm_edit_dis: 0.264400, CTCLoss: 22.112082, NRTRLoss: 0.000000, loss: 21.547119, avg_reader_cost: 1.32500 s, avg_batch_cost: 2.37332 s, avg_samples: 76.8, ips: 32.35970 samples/s, eta: 2:22:00
[2024/02/26 22:10:56] ppocr INFO: epoch: [1/50], global_step: 20, lr: 0.000013, acc: 0.044271, norm_edit_dis: 0.263883, CTCLoss: 21.502228, NRTRLoss: 0.000000, loss: 21.449989, avg_reader_cost: 0.87371 s, avg_batch_cost: 1.30937 s, avg_samples: 72.0, ips: 54.98843 samples/s, eta: 1:49:52
[2024/02/26 22:11:06] ppocr INFO: epoch: [1/50], global_step: 30, lr: 0.000027, acc: 0.052083, norm_edit_dis: 0.269422, CTCLoss: 20.095638, NRTRLoss: 0.000000, loss: 20.095638, avg_reader_cost: 0.54818 s, avg_batch_cost: 0.96542 s, avg_samples: 70.4, ips: 72.92178 samples/s, eta: 1:32:11
[2024/02/26 22:11:13] ppocr INFO: epoch: [1/50], global_step: 40, lr: 0.000041, acc: 0.052083, norm_edit_dis: 0.286861, CTCLoss: 15.908986, NRTRLoss: 0.000000, loss: 16.115395, avg_reader_cost: 0.24486 s, avg_batch_cost: 0.67578 s, avg_samples: 65.6, ips: 97.07327 samples/s, eta: 1:18:58
[2024/02/26 22:11:18] ppocr INFO: epoch: [1/50], global_step: 50, lr: 0.000055, acc: 0.083333, norm_edit_dis: 0.339697, CTCLoss: 12.769976, NRTRLoss: 0.000000, loss: 12.738501, avg_reader_cost: 0.08555 s, avg_batch_cost: 0.54633 s, avg_samples: 60.8, ips: 111.28837 samples/s, eta: 1:09:27
[2024/02/26 22:11:23] ppocr INFO: epoch: [1/50], global_step: 60, lr: 0.000069, acc: 0.151042, norm_edit_dis: 0.416667, CTCLoss: 10.439315, NRTRLoss: 0.000000, loss: 10.450922, avg_reader_cost: 0.02620 s, avg_batch_cost: 0.48454 s, avg_samples: 64.0, ips: 132.08500 samples/s, eta: 1:02:29
[2024/02/26 22:11:28] ppocr INFO: epoch: [1/50], global_step: 70, lr: 0.000083, acc: 0.182292, norm_edit_dis: 0.438015, CTCLoss: 8.423864, NRTRLoss: 0.000000, loss: 8.423864, avg_reader_cost: 0.00480 s, avg_batch_cost: 0.47775 s, avg_samples: 73.6, ips: 154.05539 samples/s, eta: 0:57:25
[2024/02/26 22:11:29] ppocr INFO: save model in ./output/rec_ppocr_v4\latest
[2024/02/26 22:11:34] ppocr INFO: epoch: [2/50], global_step: 80, lr: 0.000097, acc: 0.231771, norm_edit_dis: 0.452865, CTCLoss: 7.541358, NRTRLoss: 0.000000, loss: 7.541358, avg_reader_cost: 0.06546 s, avg_batch_cost: 0.57578 s, avg_samples: 64.0, ips: 111.15327 samples/s, eta: 0:54:18
[2024/02/26 22:11:39] ppocr INFO: epoch: [2/50], global_step: 90, lr: 0.000110, acc: 0.234375, norm_edit_dis: 0.441927, CTCLoss: 7.055272, NRTRLoss: 0.000000, loss: 7.055272, avg_reader_cost: 0.00020 s, avg_batch_cost: 0.48734 s, avg_samples: 67.2, ips: 137.89159 samples/s, eta: 0:51:18
[2024/02/26 22:11:43] ppocr INFO: epoch: [2/50], global_step: 100, lr: 0.000124, acc: 0.250000, norm_edit_dis: 0.442758, CTCLoss: 6.601116, NRTRLoss: 0.000000, loss: 6.580836, avg_reader_cost: 0.00016 s, avg_batch_cost: 0.47722 s, avg_samples: 68.8, ips: 144.16936 samples/s, eta: 0:48:49
[2024/02/26 22:11:48] ppocr INFO: epoch: [2/50], global_step: 110, lr: 0.000138, acc: 0.250000, norm_edit_dis: 0.447743, CTCLoss: 6.323363, NRTRLoss: 0.000000, loss: 6.296344, avg_reader_cost: 0.00010 s, avg_batch_cost: 0.48471 s, avg_samples: 73.6, ips: 151.84265 samples/s, eta: 0:46:49
[2024/02/26 22:11:53] ppocr INFO: epoch: [2/50], global_step: 120, lr: 0.000152, acc: 0.250000, norm_edit_dis: 0.437240, CTCLoss: 6.173958, NRTRLoss: 0.000000, loss: 6.173958, avg_reader_cost: 0.00045 s, avg_batch_cost: 0.47134 s, avg_samples: 65.6, ips: 139.17896 samples/s, eta: 0:45:04
[2024/02/26 22:11:58] ppocr INFO: epoch: [2/50], global_step: 130, lr: 0.000166, acc: 0.276042, norm_edit_dis: 0.462761, CTCLoss: 5.208588, NRTRLoss: 0.000000, loss: 5.204882, avg_reader_cost: 0.00050 s, avg_batch_cost: 0.49596 s, avg_samples: 75.2, ips: 151.62433 samples/s, eta: 0:43:41
[2024/02/26 22:12:03] ppocr INFO: epoch: [2/50], global_step: 140, lr: 0.000180, acc: 0.286458, norm_edit_dis: 0.454948, CTCLoss: 5.117970, NRTRLoss: -0.000000, loss: 5.114107, avg_reader_cost: 0.00015 s, avg_batch_cost: 0.45830 s, avg_samples: 73.6, ips: 160.59267 samples/s, eta: 0:42:20
[2024/02/26 22:12:04] ppocr INFO: save model in ./output/rec_ppocr_v4\latest
[2024/02/26 22:12:08] ppocr INFO: epoch: [3/50], global_step: 150, lr: 0.000194, acc: 0.229167, norm_edit_dis: 0.407379, CTCLoss: 5.243576, NRTRLoss: 0.000000, loss: 5.243576, avg_reader_cost: 0.06199 s, avg_batch_cost: 0.57789 s, avg_samples: 73.6, ips: 127.36036 samples/s, eta: 0:41:36
[2024/02/26 22:12:13] ppocr INFO: epoch: [3/50], global_step: 160, lr: 0.000208, acc: 0.229167, norm_edit_dis: 0.414497, CTCLoss: 4.979708, NRTRLoss: 0.000000, loss: 4.993014, avg_reader_cost: 0.00000 s, avg_batch_cost: 0.46680 s, avg_samples: 68.8, ips: 147.38596 samples/s, eta: 0:40:34
[2024/02/26 22:12:18] ppocr INFO: epoch: [3/50], global_step: 170, lr: 0.000222, acc: 0.171875, norm_edit_dis: 0.334722, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.00010 s, avg_batch_cost: 0.48857 s, avg_samples: 68.8, ips: 140.82051 samples/s, eta: 0:39:42
[2024/02/26 22:12:23] ppocr INFO: epoch: [3/50], global_step: 180, lr: 0.000235, acc: 0.000000, norm_edit_dis: 0.000000, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.00010 s, avg_batch_cost: 0.48262 s, avg_samples: 70.4, ips: 145.87084 samples/s, eta: 0:38:55
[2024/02/26 22:12:27] ppocr INFO: epoch: [3/50], global_step: 190, lr: 0.000249, acc: 0.000000, norm_edit_dis: 0.000000, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.00025 s, avg_batch_cost: 0.47050 s, avg_samples: 76.8, ips: 163.23136 samples/s, eta: 0:38:10
[2024/02/26 22:12:32] ppocr INFO: epoch: [3/50], global_step: 200, lr: 0.000263, acc: 0.000000, norm_edit_dis: 0.000000, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.00025 s, avg_batch_cost: 0.48621 s, avg_samples: 64.0, ips: 131.63061 samples/s, eta: 0:37:32
[2024/02/26 22:12:37] ppocr INFO: epoch: [3/50], global_step: 210, lr: 0.000277, acc: 0.000000, norm_edit_dis: 0.000000, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.00025 s, avg_batch_cost: 0.46303 s, avg_samples: 68.8, ips: 148.58720 samples/s, eta: 0:36:53
[2024/02/26 22:12:38] ppocr INFO: save model in ./output/rec_ppocr_v4\latest
[2024/02/26 22:12:42] ppocr INFO: epoch: [4/50], global_step: 220, lr: 0.000291, acc: 0.000000, norm_edit_dis: 0.000000, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.05028 s, avg_batch_cost: 0.52260 s, avg_samples: 59.2, ips: 113.27984 samples/s, eta: 0:36:26
[2024/02/26 22:12:47] ppocr INFO: epoch: [4/50], global_step: 230, lr: 0.000305, acc: 0.000000, norm_edit_dis: 0.000000, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.00030 s, avg_batch_cost: 0.48810 s, avg_samples: 68.8, ips: 140.95372 samples/s, eta: 0:35:56
[2024/02/26 22:12:52] ppocr INFO: epoch: [4/50], global_step: 240, lr: 0.000319, acc: 0.000000, norm_edit_dis: 0.000000, CTCLoss: nanxxx, NRTRLoss: nanxxx, loss: nanxxx, avg_reader_cost: 0.00036 s, avg_batch_cost: 0.47478 s, avg_samples: 67.2, ips: 141.53904 samples/s, eta: 0:35:26
Process finished with exit code -1073741510 (0xC000013A: interrupted by Ctrl+C)

tink2123 · 2024-02-27T03:18:10Z

tink2123
Feb 27, 2024
Collaborator

训练数据大概有多少条呢，是几卡训练的。可以尝试把loss再减小10倍。

0 replies

iPengXPro · 2024-02-27T03:23:25Z

iPengXPro
Feb 27, 2024
Author

训练数据大概有多少条呢，是几卡训练的。可以尝试把loss再减小10倍。

您好，我是单卡训练的，数据集6000张，然后是这样的图片，我在v3是可以正常训练的，v4不行

0 replies

tink2123 · 2024-02-28T06:13:53Z

tink2123
Feb 28, 2024
Collaborator

文本长度较短，V4中有一些增广的策略可能在过短文本上不适用。可以尝试关闭：
删除sampler：

PaddleOCR/configs/rec/PP-OCRv4/en_PP-OCRv4_rec.yml

Lines 102 to 115 in 0525f6b

    
           name: MultiScaleSampler 
        
           scales: 
        
           - - 320 
        
             - 32 
        
           - - 320 
        
             - 48 
        
           - - 320 
        
             - 64 
        
           first_bs: 96 
        
           fix_bs: false 
        
           divided_factor: 
        
           - 8 
        
           - 16 
        
           is_training: true

修改：

PaddleOCR/configs/rec/PP-OCRv4/en_PP-OCRv4_rec.yml

Lines 73 to 78 in 0525f6b

    
           name: MultiScaleDataSet 
        
           ds_width: false 
        
           data_dir: ./train_data/ 
        
           ext_op_transform_idx: 1 
        
           label_file_list: 
        
           - ./train_data/train_list.txt

为：

PaddleOCR/configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml

Lines 79 to 83 in 0525f6b

    
           name: SimpleDataSet 
        
           data_dir: ./train_data/ 
        
           ext_op_transform_idx: 1 
        
           label_file_list: 
        
           - ./train_data/train_list.txt

0 replies

iPengXPro · 2024-02-28T12:54:52Z

iPengXPro
Feb 28, 2024
Author

文本长度较短，V4中有一些增广的策略可能在过短文本上不适用。可以尝试关闭：删除sampler：

PaddleOCR/configs/rec/PP-OCRv4/en_PP-OCRv4_rec.yml

Lines 102 to 115 in 0525f6b

name: MultiScaleSampler

scales:

- - 320

- 32

- - 320

- 48

- - 320

- 64

first_bs: 96

fix_bs: false

divided_factor:

- 8

- 16

is_training: true

修改：

PaddleOCR/configs/rec/PP-OCRv4/en_PP-OCRv4_rec.yml

Lines 73 to 78 in 0525f6b

name: MultiScaleDataSet

ds_width: false

data_dir: ./train_data/

ext_op_transform_idx: 1

label_file_list:

- ./train_data/train_list.txt

为：

PaddleOCR/configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml

Lines 79 to 83 in 0525f6b

name: SimpleDataSet

data_dir: ./train_data/

ext_op_transform_idx: 1

label_file_list:

- ./train_data/train_list.txt

感谢回复，按照您说的，去掉sampler，并且将配置文件中的数据集改为simpledataset，还是训练几个step acc 变为0，loss变为nanxx，您看还有其他需要调整的地方吗？

Global:
debug: false
use_gpu: true
epoch_num: 50
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/rec_ppocr_v4
save_epoch_step: 10
eval_batch_step:

0
2000
cal_metric_during_train: true
pretrained_model: refactor
checkpoints: null
save_inference_dir: null
use_visualdl: false
infer_img: doc/imgs_words/ch/word_1.jpg
character_dict_path: ppocr/utils/en_dict.txt
max_text_length: &max_text_length 10
infer_mode: false
use_space_char: true
distributed: true
save_res_path: ./output/rec/predicts_ppocrv3.txt
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
warmup_epoch: 5
regularizer:
name: L2
factor: 3.0e-05
Architecture:
model_type: rec
algorithm: SVTR_LCNet
Transform: null
Backbone:
name: PPLCNetV3
scale: 0.95
Head:
name: MultiHead
head_list:
- CTCHead:
  Neck:
  name: svtr
  dims: 120
  depth: 2
  hidden_dims: 120
  kernel_size:
  - 1
  - 3
  use_guide: true
  Head:
  fc_decay: 1.0e-05
- NRTRHead:
  nrtr_dim: 384
  max_text_length: *max_text_length
  Loss:
  name: MultiLoss
  loss_config_list:
CTCLoss: null
NRTRLoss: null
PostProcess:
name: CTCLabelDecode
Metric:
name: RecMetric
main_indicator: acc
ignore_space: false
Train:
dataset:
name: SimpleDataSet
data_dir: D:\PYTHON\pictures
ext_op_transform_idx: 1
label_file_list:
- D:\PYTHON\pictures\img\rec_train.txt
  transforms:
- DecodeImage:
  img_mode: BGR
  channel_first: false
- RecConAug:
  prob: 0.5
  ext_data_num: 2
  image_shape:
  - 48
  - 96
  - 3
    max_text_length: *max_text_length
- RecAug: null
- MultiLabelEncode:
  gtc_encode: NRTRLabelEncode
- RecResizeImg:
  image_shape: [ 3, 48, 96 ]
- KeepKeys:
  keep_keys:
  - image
  - label_ctc
  - label_gtc
  - length
  - valid_ratio

loader:
shuffle: true
batch_size_per_card: 16
drop_last: true
num_workers: 1
Eval:
dataset:
name: SimpleDataSet
data_dir: D:\PYTHON\pictures
label_file_list:
- D:\PYTHON\pictures\img\rec_eval.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- MultiLabelEncode:
gtc_encode: NRTRLabelEncode
- RecResizeImg:
image_shape:
- 3
- 48
- 96
- KeepKeys:
keep_keys:
- image
- label_ctc
- label_gtc
- length
- valid_ratio
loader:
shuffle: false
drop_last: false
batch_size_per_card: 16
num_workers: 1
profiler_options: null

0 replies

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

v4 en-rec模型微调loss为0 #12353

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Replies: 4 comments

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Select a reply

Uh oh!

v4 en-rec模型微调loss为0 #12353

Uh oh!

Uh oh!

iPengXPro Feb 26, 2024

Replies: 4 comments

Uh oh!

tink2123 Feb 27, 2024 Collaborator

Uh oh!

iPengXPro Feb 27, 2024 Author

Uh oh!

tink2123 Feb 28, 2024 Collaborator

Uh oh!

iPengXPro Feb 28, 2024 Author

iPengXPro
Feb 26, 2024

tink2123
Feb 27, 2024
Collaborator

iPengXPro
Feb 27, 2024
Author

tink2123
Feb 28, 2024
Collaborator

iPengXPro
Feb 28, 2024
Author