Ham714
commented
1 year 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：Centos8 Python3.7
- 版本号/Version：Paddle： PaddleOCR： 问题相关组件/Related components：
- 运行指令/Command Code：[root@localhost PaddleOCR-release-2.6]# python3.7 tools/train.py -c configs/rec/rec_icdar15_train.yml
- 完整报错/Complete Error Message： [2023/01/14 16:16:11] ppocr INFO: epoch: [2/72], global_step: 4, lr: 0.000015, acc: 0.000000, norm_edit_dis: 0.008908, loss: 296.518524, avg_reader_cost: 2.61998 s, avg_batch_cost: 188.85271 s, avg_samples: 51.2, ips: 0.27111 samples/s, eta: 1 day, 12:34:49 [2023/01/14 16:16:15] ppocr INFO: save model in ./output/db_mv3/plate_rec_v1/latest [2023/01/14 16:45:50] ppocr INFO: epoch: [3/72], global_step: 6, lr: 0.000025, acc: 0.000000, norm_edit_dis: 0.012447, loss: 291.637024, avg_reader_cost: 2.12681 s, avg_batch_cost: 177.91249 s, avg_samples: 51.2, ips: 0.28778 samples/s, eta: 1 day, 11:24:18
大老们，训练完后 acc 始终一直是0，压根就没有识别精度，不知道是什么问题了，请求大老们帮忙分析下是什么因为导致的，尝试过不同的模型最终都未能解决， 使用的数据样本是网上下载的 CCPD2019 车牌数据图片。
rec_icdar15_train.yml 配置参数： Global: use_gpu: false epoch_num: 72 log_smooth_window: 20 print_batch_step: 10 save_model_dir: ./output/db_mv3/plate_rec_v1 save_epoch_step: 5
evaluation is run every 2000 iterations
eval_batch_step: [0, 50] cal_metric_during_train: True pretrained_model: ./pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_train/best_accuracy
pretrain_weights: ./pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_train/best_accuracy
checkpoints: save_inference_dir: ./ use_visualdl: False infer_img: /home/ym/Downloads/r9888.png
for data or label process
character_dict_path: dataset/dict.txt #dataset/dict.txt character_type: ch max_text_length: 9 infer_mode: False use_space_char: True save_res_path: ./output/db_mv3/plate_rec_v1/predicts_ic15.txt
Optimizer: name: Adam beta1: 0.9 beta2: 0.999 lr: name: Cosine #add new learning_rate: 0.0001 #0.0005 warmup_epoch: 5 #add new regularizer: name: 'L2' factor: 0.00004 #0
Architecture: model_type: rec algorithm: CRNN Transform: Backbone: name: ResNet #MobileNetV3 layers: 34 #add new #scale: 0.5 #model_name: large Neck: name: SequenceEncoder encoder_type: rnn hidden_size: 256 #96 Head: name: CTCHead fc_decay: 0.00004 #0
Loss: name: CTCLoss
PostProcess: name: CTCLabelDecode
Metric: name: RecMetric main_indicator: acc
Train: dataset: name: SimpleDataSet data_dir: /usr/projects/PaddleOCR-release-2.6/ #/usr/projects/PaddleOCR-release-2.6 label_file_list: ["dataset/train_rec.txt"] transforms: - DecodeImage: # load image img_mode: BGR channel_first: False - RecAug: # add new - CTCLabelEncode: # Class handling label - RecResizeImg: image_shape: [3, 32, 320] # -> [3, 32, 100] - KeepKeys: keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order loader: shuffle: True batch_size_per_card: 256 drop_last: True num_workers: 8 #use_shared_memory: False
Eval: dataset: name: SimpleDataSet data_dir: /usr/projects/PaddleOCR-release-2.6/ #/usr/projects/PaddleOCR-release-2.6 label_file_list: ["dataset/dev_rec.txt"] transforms: - DecodeImage: # load image img_mode: BGR channel_first: False - CTCLabelEncode: # Class handling label - RecResizeImg: image_shape: [3, 32, 320] # -> [3, 32, 100] - KeepKeys: keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order loader: shuffle: False drop_last: False batch_size_per_card: 256 num_workers: 4 #use_shared_memory: False
Same issue for nearly 1 month and still didnt get resolved
WenmuZhou
ftmasadi
github-actions[bot]
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
大老们，训练完后 acc 始终一直是0，压根就没有识别精度，不知道是什么问题了，请求大老们帮忙分析下是什么因为导致的，尝试过不同的模型最终都未能解决， 使用的数据样本是网上下载的 CCPD2019 车牌数据图片。
rec_icdar15_train.yml 配置参数：