Getting zero accuracy acc = 0

[ftmasadi]

I use the following yaml file, I even spent a few days training it, but every time I change it, it doesn't give me any accuracy.Thank you for helping me, where is the problem with my work? `Global: use_gpu: true epoch_num: 5000 log_smooth_window: 20 print_batch_step: 10 save_model_dir: /content/drive/MyDrive/output save_epoch_step: 1 eval_batch_step:

• 0
• 200 cal_metric_during_train: true pretrained_model: /content/PaddleOCR/pretrained_model/arabic_PP-OCRv3_rec_train/best_accuracy checkpoints: save_inference_dir: null use_visualdl: false infer_img: /content/PaddleOCR/doc/imgs_words/persian/fa_2.jpg character_dict_path: /content/good char.txt max_text_length: 120 infer_mode: false use_space_char: true save_res_path: /content/drive/MyDrive/output/rec/predicts_svtr_tiny_bn.txt Optimizer: name: AdamW beta1: 0.9 beta2: 0.99 epsilon: 8.0e-08 weight_decay: 0.05 no_weight_decay_name: norm pos_embed one_dim_param_no_weight_decay: true lr: name: Cosine learning_rate: 0.00025 warmup_epoch: 10 Architecture: model_type: rec algorithm: SVTR Transform: name: STN_ON tps_inputsize:
  • 32
  • 64 tps_outputsize:
  • 32
  • 100 num_control_points: 20 tps_margins:
  • 0.05
  • 0.05 stn_activation: none Backbone: name: SVTRNet img_size:
  • 32
  • 100 out_char_num: 120 out_channels: 192 patch_merging: Conv embed_dim:
  • 64
  • 128
  • 256 depth:
  • 3
  • 6
  • 3 num_heads:
  • 2
  • 4
  • 8 mixer:
  • Local
  • Local
  • Local
  • Local
  • Local
  • Local
  • Global
  • Global
  • Global
  • Global
  • Global
  • Global local_mixer:
  • • 7
    • 11
  • • 7
    • 11
  • • 7
    • 11 last_stage: true prenorm: false Neck: name: SequenceEncoder encoder_type: reshape Head: name: CTCHead Loss: name: CTCLoss PostProcess: name: CTCLabelDecode Metric: name: RecMetric main_indicator: acc Train: dataset: name: SimpleDataSet data_dir: /content/PaddleOCR/train_data/train/ label_file_list:
  • ./train_data/train.txt transforms:
  • DecodeImage: img_mode: BGR channel_first: false
  • CTCLabelEncode: null
  • RecResizeImg: character_dict_path: null image_shape:
    • 3
    • 64
    • 256 padding: false
  • KeepKeys: keep_keys:
    • image
    • label
    • length loader: shuffle: true batch_size_per_card: 64 drop_last: true num_workers: 4 Eval: dataset: name: SimpleDataSet data_dir: ./train_data/val label_file_list:
  • ./train_data/val.txt ratio_list:
  • 0.1 transforms:
  • DecodeImage: img_mode: BGR channel_first: false
  • CTCLabelEncode: null
  • RecResizeImg: character_dict_path: null image_shape:
    • 3
    • 64
    • 256 padding: false
  • KeepKeys: keep_keys:
    • image
    • label
    • length loader: shuffle: false drop_last: false batch_size_per_card: 64 num_workers: 4 profiler_options: null `

[ftmasadi]

This is my result

[ftmasadi]

And some labels of my dataset photos: 00010.tif به‌منظور افزایش ظرفیتشان برای استفاده از اندازه‌گیری می‌کنند. همچنین میزان بهره‌برداری آن‌ها از در فعالیت‌های 00011.tif روزانه در ارتباط با دیگر گروه‌ها مورد اندازه‌گیری قرار می‌گیرد. به طور کلی نیز شاخص

[ftmasadi]

@MohieEldinMuha I think maybe I'm making a mistake in the shape of my pictures. I saw that text in your dataset photos is long. What shape did you choose? Thank you for helping me with this problem

[Topdu]

out_char_num should be set to W//4, W//8 or W//12, and max_text_length is set up the same as out_char_num.

Long Text Recognition Optimization

Recognition scenes with most texts containing more than 15 characters will be considered as long text recognition, such as the Chinese sentence level text recognition.

• Resize shape during pre-process

According to experience, Resize shape can be set according to the aspect ratio distribution of the text image during pre-process, i.e., set to the aspect ratio that can cover 90% of the data. For example, when the aspect ratio of 90% of the data is less than 15:1 after collecting the aspect ratio of the dataset, the resize shape should be set as [H, H*15]. When H is set as 32, it is [32, 480].

• Use Padding strategy to prevent distortion caused by direct resizing of images

Where there is a text with only few characters in the data, it will cause data distortion and affect training and inference when it is resized directly to a relatively large aspect ratio. Therefore, reserve the aspect ratio first when resizing, then set H as 32. When W<480, pad 0 to W=480., when W>480, resize it to W=480 to retain the original shape of the text image.

Adjust the output sequence length

The sequence length of SVTR output is 1*W/4 by default, i.e., SVTR samples the height of the feature down to 1, and the width down to a quarter of W. However, when we recognize long text, the width is W=480 or higher, and the sequence length is 480//4=120. In practical applications, there are few text images with 120 characters. Obviously, it is unreasonable to use the default structure configuration. Therefore, we should make appropriate adjustments to the SVTR structure, and adjust the value of out_char_num in the SVTR structure configuration in the configuration file. According to experience, out_char_num is generally set as 1/4, 1/8 or 1/12 of W.

In this project, W=320, and most of the text includes not more than 40 characters, so we adjust the output sequence length in the configuration file as follows:

  Backbone:
    name: SVTRNet
    img_size:
    - 32
    - 320
    out_char_num: 40 # W//4 or W//8 or W//12
    ...

In short text recognition scenes, out_char_num is generally set as 1/4 of W; in long text recognition scenes, out_char_num is generally set as 1/8 or 1/12 of W, which can be set according to actual needs. When W=480, out_char_num can be set as 60 or 40. Where there is data with more than 40 characters, out_char_num can be set as 60, or 40.

[Aminfaraji]

@Topdu My samples dataset has dimensions (30,200,3). How to set hyperparameters in config file? (inputs image, output model or num_char_out)

[masoudMZB]

Try to reverse labels. If you read the CRNN main paper this model has an LSTM, which looks at the timestamps from left to right. as it uses CTC so it is predicting chars from left to right.

ppocr v3 arabic is not using CRNN but I think this may help .

wrong label: اکبر corrected label : ربکا

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions.

[UserWangZz]

It has not been updated for a long time. This issue is temporarily closed and can be reopened if necessary.