ppocrv4蒸馏模型训练报错

QQQTAO commented 10 months ago

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：paddlepaddle-gpu 2.3.2.post112
版本号/Version：Paddle： PaddleOCR：问题相关组件/Related components：paddleocr2.7
运行指令/Command Code：python -m paddle.distributed.launch --log_dir=./debug --gpus '0,1,3,4' tools/train.py -c configs/rec/PP-OCRv4/ch_PP-OCRv4_rec_distill.yml
完整报错/Complete Error Message：INFO 2023-08-11 10:10:52,331 launch_utils.py:563] details about PADDLE_TRAINER_ENDPOINTS can be found in ./debug/centerloss/endpoints.log, and detail running logs maybe found in ./debug/centerloss/workerlog.0 INFO 2023-08-11 10:10:52,331 launch_utils.py:563] details about PADDLE_TRAINER_ENDPOINTS can be found in ./debug/centerloss/endpoints.log, and detail running logs maybe found in ./debug/centerloss/workerlog.0 launch proc_id:22211 idx:0 launch proc_id:22216 idx:1 launch proc_id:22221 idx:2 launch proc_id:22229 idx:3 [2023/08/11 10:10:54] ppocr INFO: Architecture : [2023/08/11 10:10:54] ppocr INFO: Models : [2023/08/11 10:10:54] ppocr INFO: Student : [2023/08/11 10:10:54] ppocr INFO: Backbone : [2023/08/11 10:10:54] ppocr INFO: name : PPLCNetV3 [2023/08/11 10:10:54] ppocr INFO: scale : 0.95 [2023/08/11 10:10:54] ppocr INFO: Head : [2023/08/11 10:10:54] ppocr INFO: head_list : [2023/08/11 10:10:54] ppocr INFO: CTCHead : [2023/08/11 10:10:54] ppocr INFO: Head : [2023/08/11 10:10:54] ppocr INFO: fc_decay : 1e-05 [2023/08/11 10:10:54] ppocr INFO: Neck : [2023/08/11 10:10:54] ppocr INFO: depth : 2 [2023/08/11 10:10:54] ppocr INFO: dims : 120 [2023/08/11 10:10:54] ppocr INFO: hidden_dims : 120 [2023/08/11 10:10:54] ppocr INFO: kernel_size : [1, 3] [2023/08/11 10:10:54] ppocr INFO: name : svtr [2023/08/11 10:10:54] ppocr INFO: use_guide : True [2023/08/11 10:10:54] ppocr INFO: NRTRHead : [2023/08/11 10:10:54] ppocr INFO: max_text_length : 25 [2023/08/11 10:10:54] ppocr INFO: nrtr_dim : 384 [2023/08/11 10:10:54] ppocr INFO: name : MultiHead [2023/08/11 10:10:54] ppocr INFO: Transform : None [2023/08/11 10:10:54] ppocr INFO: algorithm : SVTR [2023/08/11 10:10:54] ppocr INFO: freeze_params : False [2023/08/11 10:10:54] ppocr INFO: model_type : rec [2023/08/11 10:10:54] ppocr INFO: pretrained : None [2023/08/11 10:10:54] ppocr INFO: return_all_feats : True [2023/08/11 10:10:54] ppocr INFO: Teacher : [2023/08/11 10:10:54] ppocr INFO: Backbone : [2023/08/11 10:10:54] ppocr INFO: depth : [3, 6, 3] [2023/08/11 10:10:54] ppocr INFO: embed_dim : [64, 128, 256] [2023/08/11 10:10:54] ppocr INFO: img_size : [48, 320] [2023/08/11 10:10:54] ppocr INFO: last_stage : False [2023/08/11 10:10:54] ppocr INFO: local_mixer : [[5, 5], [5, 5], [5, 5]] [2023/08/11 10:10:54] ppocr INFO: mixer : ['Conv', 'Conv', 'Conv', 'Conv', 'Conv', 'Conv', 'Global', 'Global', 'Global', 'Global', 'Global', 'Global'] [2023/08/11 10:10:54] ppocr INFO: name : SVTRNet [2023/08/11 10:10:54] ppocr INFO: num_heads : [2, 4, 8] [2023/08/11 10:10:54] ppocr INFO: out_channels : 192 [2023/08/11 10:10:54] ppocr INFO: out_char_num : 40 [2023/08/11 10:10:54] ppocr INFO: patch_merging : Conv [2023/08/11 10:10:54] ppocr INFO: prenorm : True [2023/08/11 10:10:54] ppocr INFO: Head : [2023/08/11 10:10:54] ppocr INFO: head_list : [2023/08/11 10:10:54] ppocr INFO: CTCHead : [2023/08/11 10:10:54] ppocr INFO: Head : [2023/08/11 10:10:54] ppocr INFO: fc_decay : 1e-05 [2023/08/11 10:10:54] ppocr INFO: Neck : [2023/08/11 10:10:54] ppocr INFO: depth : 2 [2023/08/11 10:10:54] ppocr INFO: dims : 120 [2023/08/11 10:10:54] ppocr INFO: hidden_dims : 120 [2023/08/11 10:10:54] ppocr INFO: kernel_size : [1, 3] [2023/08/11 10:10:54] ppocr INFO: name : svtr [2023/08/11 10:10:54] ppocr INFO: use_guide : True [2023/08/11 10:10:54] ppocr INFO: NRTRHead : [2023/08/11 10:10:54] ppocr INFO: max_text_length : 25 [2023/08/11 10:10:54] ppocr INFO: nrtr_dim : 384 [2023/08/11 10:10:54] ppocr INFO: name : MultiHead [2023/08/11 10:10:54] ppocr INFO: Transform : None [2023/08/11 10:10:54] ppocr INFO: algorithm : SVTR [2023/08/11 10:10:54] ppocr INFO: freeze_params : True [2023/08/11 10:10:54] ppocr INFO: model_type : rec [2023/08/11 10:10:54] ppocr INFO: pretrained : None [2023/08/11 10:10:54] ppocr INFO: return_all_feats : True [2023/08/11 10:10:54] ppocr INFO: algorithm : Distillation [2023/08/11 10:10:54] ppocr INFO: model_type : rec [2023/08/11 10:10:54] ppocr INFO: name : DistillationModel [2023/08/11 10:10:54] ppocr INFO: Eval : [2023/08/11 10:10:54] ppocr INFO: dataset : [2023/08/11 10:10:54] ppocr INFO: data_dir : ./train_data [2023/08/11 10:10:54] ppocr INFO: label_file_list : ['/home/qintao/ocr_data/real_data/new_label_test.txt'] [2023/08/11 10:10:54] ppocr INFO: name : SimpleDataSet [2023/08/11 10:10:54] ppocr INFO: transforms : [2023/08/11 10:10:54] ppocr INFO: DecodeImage : [2023/08/11 10:10:54] ppocr INFO: channel_first : False [2023/08/11 10:10:54] ppocr INFO: img_mode : BGR [2023/08/11 10:10:54] ppocr INFO: MultiLabelEncode : [2023/08/11 10:10:54] ppocr INFO: gtc_encode : NRTRLabelEncode [2023/08/11 10:10:54] ppocr INFO: RecResizeImg : [2023/08/11 10:10:54] ppocr INFO: image_shape : [3, 48, 320] [2023/08/11 10:10:54] ppocr INFO: KeepKeys : [2023/08/11 10:10:54] ppocr INFO: keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio'] [2023/08/11 10:10:54] ppocr INFO: loader : [2023/08/11 10:10:54] ppocr INFO: batch_size_per_card : 128 [2023/08/11 10:10:54] ppocr INFO: drop_last : False [2023/08/11 10:10:54] ppocr INFO: num_workers : 4 [2023/08/11 10:10:54] ppocr INFO: shuffle : False [2023/08/11 10:10:54] ppocr INFO: Global : [2023/08/11 10:10:54] ppocr INFO: cal_metric_during_train : True [2023/08/11 10:10:54] ppocr INFO: character_dict_path : ppocr/utils/ppocr_keys_v1.txt [2023/08/11 10:10:54] ppocr INFO: checkpoints : ./pretrain_model/ch_PP-OCRv4_rec_train/student.pdparams [2023/08/11 10:10:54] ppocr INFO: debug : False [2023/08/11 10:10:54] ppocr INFO: distributed : True [2023/08/11 10:10:54] ppocr INFO: epoch_num : 200 [2023/08/11 10:10:54] ppocr INFO: eval_batch_step : [0, 2000] [2023/08/11 10:10:54] ppocr INFO: infer_img : doc/imgs_words/ch/word_1.jpg [2023/08/11 10:10:54] ppocr INFO: infer_mode : False [2023/08/11 10:10:54] ppocr INFO: log_smooth_window : 20 [2023/08/11 10:10:54] ppocr INFO: max_text_length : 25 [2023/08/11 10:10:54] ppocr INFO: pretrained_model : None [2023/08/11 10:10:54] ppocr INFO: print_batch_step : 10 [2023/08/11 10:10:54] ppocr INFO: save_epoch_step : 40 [2023/08/11 10:10:54] ppocr INFO: save_inference_dir : None [2023/08/11 10:10:54] ppocr INFO: save_model_dir : ./output/rec_dkd_400w_svtr_ctc_lcnet_blank_dkd0.1/ [2023/08/11 10:10:54] ppocr INFO: save_res_path : ./output/rec/predicts_ppocrv3.txt [2023/08/11 10:10:54] ppocr INFO: use_gpu : True [2023/08/11 10:10:54] ppocr INFO: use_space_char : True [2023/08/11 10:10:54] ppocr INFO: use_visualdl : False [2023/08/11 10:10:54] ppocr INFO: Loss : [2023/08/11 10:10:54] ppocr INFO: loss_config_list : [2023/08/11 10:10:54] ppocr INFO: DistillationDKDLoss : [2023/08/11 10:10:54] ppocr INFO: alpha : 1.0 [2023/08/11 10:10:54] ppocr INFO: beta : 2.0 [2023/08/11 10:10:54] ppocr INFO: dis_head : gtc [2023/08/11 10:10:54] ppocr INFO: key : head_out [2023/08/11 10:10:54] ppocr INFO: model_name_pairs : [['Student', 'Teacher']] [2023/08/11 10:10:54] ppocr INFO: multi_head : True [2023/08/11 10:10:54] ppocr INFO: name : dkd [2023/08/11 10:10:54] ppocr INFO: weight : 0.1 [2023/08/11 10:10:54] ppocr INFO: DistillationCTCLoss : [2023/08/11 10:10:54] ppocr INFO: key : head_out [2023/08/11 10:10:54] ppocr INFO: model_name_list : ['Student'] [2023/08/11 10:10:54] ppocr INFO: multi_head : True [2023/08/11 10:10:54] ppocr INFO: weight : 1.0 [2023/08/11 10:10:54] ppocr INFO: DistillationNRTRLoss : [2023/08/11 10:10:54] ppocr INFO: key : head_out [2023/08/11 10:10:54] ppocr INFO: model_name_list : ['Student'] [2023/08/11 10:10:54] ppocr INFO: multi_head : True [2023/08/11 10:10:54] ppocr INFO: smoothing : False [2023/08/11 10:10:54] ppocr INFO: weight : 1.0 [2023/08/11 10:10:54] ppocr INFO: DistillCTCLogits : [2023/08/11 10:10:54] ppocr INFO: key : head_out [2023/08/11 10:10:54] ppocr INFO: model_name_pairs : [['Student', 'Teacher']] [2023/08/11 10:10:54] ppocr INFO: reduction : mean [2023/08/11 10:10:54] ppocr INFO: weight : 1.0 [2023/08/11 10:10:54] ppocr INFO: name : CombinedLoss [2023/08/11 10:10:54] ppocr INFO: Metric : [2023/08/11 10:10:54] ppocr INFO: base_metric_name : RecMetric [2023/08/11 10:10:54] ppocr INFO: ignore_space : False [2023/08/11 10:10:54] ppocr INFO: key : Student [2023/08/11 10:10:54] ppocr INFO: main_indicator : acc [2023/08/11 10:10:54] ppocr INFO: name : DistillationMetric [2023/08/11 10:10:54] ppocr INFO: Optimizer : [2023/08/11 10:10:54] ppocr INFO: beta1 : 0.9 [2023/08/11 10:10:54] ppocr INFO: beta2 : 0.999 [2023/08/11 10:10:54] ppocr INFO: lr : [2023/08/11 10:10:54] ppocr INFO: learning_rate : 0.0005 [2023/08/11 10:10:54] ppocr INFO: name : Cosine [2023/08/11 10:10:54] ppocr INFO: warmup_epoch : 2 [2023/08/11 10:10:54] ppocr INFO: name : Adam [2023/08/11 10:10:54] ppocr INFO: regularizer : [2023/08/11 10:10:54] ppocr INFO: factor : 3e-05 [2023/08/11 10:10:54] ppocr INFO: name : L2 [2023/08/11 10:10:54] ppocr INFO: PostProcess : [2023/08/11 10:10:54] ppocr INFO: key : head_out [2023/08/11 10:10:54] ppocr INFO: model_name : ['Student'] [2023/08/11 10:10:54] ppocr INFO: multi_head : True [2023/08/11 10:10:54] ppocr INFO: name : DistillationCTCLabelDecode [2023/08/11 10:10:54] ppocr INFO: Train : [2023/08/11 10:10:54] ppocr INFO: dataset : [2023/08/11 10:10:54] ppocr INFO: data_dir : ./train_data/ [2023/08/11 10:10:54] ppocr INFO: label_file_list : ['/home/qintao/ocr_data/real_data/new_label_train.txt', '/home/qintao/ocr_data/RecData2/new_labels_gen.txt', '/home/qintao/ocr_data/RecData2/hor_white/image_labels.txt'] [2023/08/11 10:10:54] ppocr INFO: name : SimpleDataSet [2023/08/11 10:10:54] ppocr INFO: ratio_list : [1.0, 0.1, 1.0] [2023/08/11 10:10:54] ppocr INFO: transforms : [2023/08/11 10:10:54] ppocr INFO: DecodeImage : [2023/08/11 10:10:54] ppocr INFO: channel_first : False [2023/08/11 10:10:54] ppocr INFO: img_mode : BGR [2023/08/11 10:10:54] ppocr INFO: RecAug : None [2023/08/11 10:10:54] ppocr INFO: MultiLabelEncode : [2023/08/11 10:10:54] ppocr INFO: gtc_encode : NRTRLabelEncode [2023/08/11 10:10:54] ppocr INFO: KeepKeys : [2023/08/11 10:10:54] ppocr INFO: keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio'] [2023/08/11 10:10:54] ppocr INFO: loader : [2023/08/11 10:10:54] ppocr INFO: batch_size_per_card : 128 [2023/08/11 10:10:54] ppocr INFO: drop_last : True [2023/08/11 10:10:54] ppocr INFO: num_workers : 4 [2023/08/11 10:10:54] ppocr INFO: shuffle : True [2023/08/11 10:10:54] ppocr INFO: use_shared_memory : True [2023/08/11 10:10:54] ppocr INFO: profiler_options : None [2023/08/11 10:10:54] ppocr INFO: train with paddle 2.3.2 and device Place(gpu:0) I0811 10:10:54.980728 22211 nccl_context.cc:83] init nccl context nranks: 4 local rank: 0 gpu id: 0 ring id: 0 W0811 10:10:56.348552 22211 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 8.6, Driver API Version: 11.2, Runtime API Version: 11.2 W0811 10:10:56.353502 22211 gpu_resources.cc:91] device: 0, cuDNN Version: 8.1. [2023/08/11 10:10:57] ppocr INFO: Initialize indexs of datasets:['/home/qintao/ocr_data/real_data/new_label_train.txt', '/home/qintao/ocr_data/RecData2/new_labels_gen.txt', '/home/qintao/ocr_data/RecData2/hor_white/image_labels.txt'] list index out of range [2023/08/11 10:10:59] ppocr INFO: Initialize indexs of datasets:['/home/qintao/ocr_data/real_data/new_label_test.txt'] Traceback (most recent call last): File "tools/train.py", line 227, in main(config, device, logger, vdl_writer) File "tools/train.py", line 135, in main model = build_model(config['Architecture']) File "/home/qintao/PaddleOCRV4/ppocr/modeling/architectures/init.py", line 34, in build_model arch = getattr(mod, name)(config) File "/home/qintao/PaddleOCRV4/ppocr/modeling/architectures/distillation_model.py", line 47, in init model = BaseModel(model_config) File "/home/qintao/PaddleOCRV4/ppocr/modeling/architectures/base_model.py", line 76, in init self.head = build_head(config["Head"]) File "/home/qintao/PaddleOCRV4/ppocr/modeling/heads/init.py", line 71, in build_head module_class = eval(module_name)(**config) File "/home/qintao/PaddleOCRV4/ppocr/modeling/heads/rec_multi_head.py", line 74, in init out_channels=out_channels_list['NRTRLabelDecode']) KeyError: 'NRTRLabelDecode' 使用代码中的配置文件只修改了数据集配置和最大字符长度，出现这种错误，调试的时候也发现报错的这个配置选项

QQQTAO commented 10 months ago

我调试的时候又仔细对比了一下配置文件模型参数和训练代码中模型创建需求配置，模型蒸馏训练需求条件和配置文件模型配置参数不一致

aydenk1 commented 10 months ago

I am seeing the same issue with config /rec/PP-OCRv4/ch_PP-OCRv4_rec_distill.yml. Some help would be appreciated!

jadehh commented 10 months ago

请问解决了吗

QQQTAO commented 10 months ago

请问解决了吗

进行调试是配置文件参数和训练代码不匹配，修改源码才能用吧

QQQTAO commented 10 months ago

I am seeing the same issue with config /rec/PP-OCRv4/ch_PP-OCRv4_rec_distill.yml. Some help would be appreciated!

This is caused by the mismatch between the configuration file parameters and the training code. It may be solved by modifying the source code according to the configuration file parameters.

toanphanminh commented 10 months ago

I am seeing the same issue with config /rec/PP-OCRv4/ch_PP-OCRv4_rec_distill.yml. Some help would be appreciated!

This is caused by the mismatch between the configuration file parameters and the training code. It may be solved by modifying the source code according to the configuration file parameters.

How to modify it, can you share the update code?

xuxiansheng2018 commented 10 months ago

怎么修改呢，能分享下更新代码吗？

QQQTAO commented 10 months ago

我没有进行代码修改，这个是官方代码和配置文件问题，等着官方更新代码吧

秦涛 @.***

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2023年8月28日(星期一) 下午4:00 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [PaddlePaddle/PaddleOCR] ppocrv4蒸馏模型训练报错 (Issue #10610)

怎么修改呢，能分享下更新代码吗？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

UserWangZz commented 1 month ago

蒸馏模型目前还存在问题，已经反馈问题，等待推进

PaddlePaddle / PaddleOCR

ppocrv4蒸馏模型训练报错 #10610