Changing Image Size in PP-OCRv4 Configuration

[ntdgo]

🔎 Search before asking

• I have searched the PaddleOCR Docs and found no similar bug report.
• I have searched the PaddleOCR Issues and found no similar bug report.
• I have searched the PaddleOCR Discussions and found no similar bug report.

🐛 Bug (问题描述)

Hello,

I am training the PP-OCRv4 recognition model using my custom dataset. The average size of the images in my dataset is 144x320 (height x width). To align with this, I modified the image_shape, d2s_train_image_shape, scales parameter in the configuration file https://github.com/PaddlePaddle/PaddleOCR/blob/0d41ffc91d218d0abd6ed7e27c23983740ca3cb1/configs/rec/PP-OCRv4/ch_PP-OCRv4_rec_hgnet.yml to reflect my target image size.

After training, the evaluation phase, as shown as (2). Notably, with the default configuration (image_shape: [3, 48, 320]), the training and evaluation processes ran without any issues.

Could you please help me resolve this issue?
Thank you!

1. Configuration

Global:
  debug: false
  use_gpu: true
  epoch_num: 20000
  log_smooth_window: 20
  print_batch_step: 10
  save_model_dir: ./output/rec_ppocr_v4_hgnet
  save_epoch_step: 10
  eval_batch_step: [0, 50]
  cal_metric_during_train: true
  pretrained_model:
  checkpoints:
  save_inference_dir:
  use_visualdl: false
  infer_img: doc/imgs_words/ch/word_1.jpg
  character_dict_path: ./configs/dict.txt
  max_text_length: &max_text_length 25
  infer_mode: false
  use_space_char: true
  distributed: true
  save_res_path: ./output/rec/predicts_ppocrv3.txt
  d2s_train_image_shape: [3, 144, 320]

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_HGNet
  Transform:
  Backbone:
    name: PPHGNet_small
  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: 0.00001
      - NRTRHead:
          nrtr_dim: 384
          max_text_length: *max_text_length

Loss:
  name: MultiLoss
  loss_config_list:
    - CTCLoss:
    - NRTRLoss:

PostProcess:  
  name: CTCLabelDecode

Metric:
  name: RecMetric
  main_indicator: acc

Train:
  dataset:
    name: MultiScaleDataSet
    ds_width: false
    data_dir: ./dataset/ocr_rec
    ext_op_transform_idx: 1
    label_file_list:
    - ./dataset/ocr_rec/train.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - RecConAug:
        prob: 0.5
        ext_data_num: 2
        image_shape: [144, 320, 3]
        max_text_length: *max_text_length
    - RecAug:
    - MultiLabelEncode:
        gtc_encode: NRTRLabelEncode
    - KeepKeys:
        keep_keys:
        - image
        - label_ctc
        - label_gtc
        - length
        - valid_ratio
  sampler:
    name: MultiScaleSampler
    scales: [[320, 112], [320, 144], [320, 176]]
    first_bs: &bs 50
    fix_bs: false
    divided_factor: [8, 16] # w, h
    is_training: True
  loader:
    shuffle: true
    batch_size_per_card: *bs
    drop_last: true
    num_workers: 52
Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./dataset/ocr_rec
    label_file_list:
    - ./dataset/ocr_rec/val.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - MultiLabelEncode:
        gtc_encode: NRTRLabelEncode
    - RecResizeImg:
        image_shape: [3, 144, 320]
    - KeepKeys:
        keep_keys:
        - image
        - label_ctc
        - label_gtc
        - length
        - valid_ratio
  loader:
    shuffle: false
    drop_last: false
    batch_size_per_card: 128
    num_workers: 4

2. Console

(.venv) C:\GitHub\USER-OCR\scr\train>python ./PaddleOCR/tools/train.py -c ./configs/ch_PP-OCRv4_rec_hgnet.yml
[2024/12/26 00:19:22] ppocr INFO: Architecture : 
[2024/12/26 00:19:22] ppocr INFO:     Backbone : 
[2024/12/26 00:19:22] ppocr INFO:         name : PPHGNet_small
[2024/12/26 00:19:22] ppocr INFO:     Head : 
[2024/12/26 00:19:22] ppocr INFO:         head_list : 
[2024/12/26 00:19:22] ppocr INFO:             CTCHead : 
[2024/12/26 00:19:22] ppocr INFO:                 Head :
[2024/12/26 00:19:22] ppocr INFO:                     fc_decay : 1e-05
[2024/12/26 00:19:22] ppocr INFO:                 Neck :
[2024/12/26 00:19:22] ppocr INFO:                     depth : 2
[2024/12/26 00:19:22] ppocr INFO:                     dims : 120
[2024/12/26 00:19:22] ppocr INFO:                     hidden_dims : 120
[2024/12/26 00:19:22] ppocr INFO:                     kernel_size : [1, 3]
[2024/12/26 00:19:22] ppocr INFO:                     name : svtr
[2024/12/26 00:19:22] ppocr INFO:                     use_guide : True
[2024/12/26 00:19:22] ppocr INFO:             NRTRHead :
[2024/12/26 00:19:22] ppocr INFO:                 max_text_length : 25
[2024/12/26 00:19:22] ppocr INFO:                 nrtr_dim : 384
[2024/12/26 00:19:22] ppocr INFO:         name : MultiHead
[2024/12/26 00:19:22] ppocr INFO:     Transform : None
[2024/12/26 00:19:22] ppocr INFO:     algorithm : SVTR_HGNet
[2024/12/26 00:19:22] ppocr INFO:     model_type : rec
[2024/12/26 00:19:22] ppocr INFO: Eval :
[2024/12/26 00:19:22] ppocr INFO:     dataset :
[2024/12/26 00:19:22] ppocr INFO:         data_dir : ./dataset/ocr_rec
[2024/12/26 00:19:22] ppocr INFO:         label_file_list : ['./dataset/ocr_rec/val.txt']
[2024/12/26 00:19:22] ppocr INFO:         name : SimpleDataSet
[2024/12/26 00:19:22] ppocr INFO:         transforms :
[2024/12/26 00:19:22] ppocr INFO:             DecodeImage :
[2024/12/26 00:19:22] ppocr INFO:                 channel_first : False
[2024/12/26 00:19:22] ppocr INFO:                 img_mode : BGR
[2024/12/26 00:19:22] ppocr INFO:             MultiLabelEncode :
[2024/12/26 00:19:22] ppocr INFO:                 gtc_encode : NRTRLabelEncode
[2024/12/26 00:19:22] ppocr INFO:             RecResizeImg :
[2024/12/26 00:19:22] ppocr INFO:                 image_shape : [3, 144, 320]
[2024/12/26 00:19:22] ppocr INFO:             KeepKeys :
[2024/12/26 00:19:22] ppocr INFO:                 keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio']
[2024/12/26 00:19:22] ppocr INFO:     loader :
[2024/12/26 00:19:22] ppocr INFO:         batch_size_per_card : 128
[2024/12/26 00:19:22] ppocr INFO:         drop_last : False
[2024/12/26 00:19:22] ppocr INFO:         num_workers : 4
[2024/12/26 00:19:22] ppocr INFO:         shuffle : False
[2024/12/26 00:19:22] ppocr INFO: Global :
[2024/12/26 00:19:22] ppocr INFO:     cal_metric_during_train : True
[2024/12/26 00:19:22] ppocr INFO:     character_dict_path : ./configs/dict.txt
[2024/12/26 00:19:22] ppocr INFO:     checkpoints : None
[2024/12/26 00:19:22] ppocr INFO:     d2s_train_image_shape : [3, 144, 320]
[2024/12/26 00:19:22] ppocr INFO:     debug : False
[2024/12/26 00:19:22] ppocr INFO:     distributed : False
[2024/12/26 00:19:22] ppocr INFO:     epoch_num : 20000
[2024/12/26 00:19:22] ppocr INFO:     eval_batch_step : [0, 50]
[2024/12/26 00:19:22] ppocr INFO:     infer_img : doc/imgs_words/ch/word_1.jpg
[2024/12/26 00:19:22] ppocr INFO:     infer_mode : False
[2024/12/26 00:19:22] ppocr INFO:     log_smooth_window : 20
[2024/12/26 00:19:22] ppocr INFO:     max_text_length : 25
[2024/12/26 00:19:22] ppocr INFO:     pretrained_model : None
[2024/12/26 00:19:22] ppocr INFO:     print_batch_step : 10
[2024/12/26 00:19:22] ppocr INFO:     save_epoch_step : 10
[2024/12/26 00:19:22] ppocr INFO:     save_inference_dir : None
[2024/12/26 00:19:22] ppocr INFO:     save_model_dir : ./output/rec_ppocr_v4_hgnet
[2024/12/26 00:19:22] ppocr INFO:     save_res_path : ./output/rec/predicts_ppocrv3.txt
[2024/12/26 00:19:22] ppocr INFO:     use_gpu : True
[2024/12/26 00:19:22] ppocr INFO:     use_space_char : True
[2024/12/26 00:19:22] ppocr INFO:     use_visualdl : False
[2024/12/26 00:19:22] ppocr INFO: Loss :
[2024/12/26 00:19:22] ppocr INFO:     loss_config_list :
[2024/12/26 00:19:22] ppocr INFO:         CTCLoss : None
[2024/12/26 00:19:22] ppocr INFO:         NRTRLoss : None
[2024/12/26 00:19:22] ppocr INFO:     name : MultiLoss
[2024/12/26 00:19:22] ppocr INFO: Metric :
[2024/12/26 00:19:22] ppocr INFO:     main_indicator : acc
[2024/12/26 00:19:22] ppocr INFO:     name : RecMetric
[2024/12/26 00:19:22] ppocr INFO: Optimizer :
[2024/12/26 00:19:22] ppocr INFO:     beta1 : 0.9
[2024/12/26 00:19:22] ppocr INFO:     beta2 : 0.999
[2024/12/26 00:19:22] ppocr INFO:     lr :
[2024/12/26 00:19:22] ppocr INFO:         learning_rate : 0.001
[2024/12/26 00:19:22] ppocr INFO:         name : Cosine
[2024/12/26 00:19:22] ppocr INFO:         warmup_epoch : 5
[2024/12/26 00:19:22] ppocr INFO:     name : Adam
[2024/12/26 00:19:22] ppocr INFO:     regularizer :
[2024/12/26 00:19:22] ppocr INFO:         factor : 3e-05
[2024/12/26 00:19:22] ppocr INFO:         name : L2
[2024/12/26 00:19:22] ppocr INFO: PostProcess :
[2024/12/26 00:19:22] ppocr INFO:     name : CTCLabelDecode
[2024/12/26 00:19:22] ppocr INFO: Train :
[2024/12/26 00:19:22] ppocr INFO:     dataset :
[2024/12/26 00:19:22] ppocr INFO:         data_dir : ./dataset/ocr_rec
[2024/12/26 00:19:22] ppocr INFO:         ds_width : False
[2024/12/26 00:19:22] ppocr INFO:         ext_op_transform_idx : 1
[2024/12/26 00:19:22] ppocr INFO:         label_file_list : ['./dataset/ocr_rec/train.txt']
[2024/12/26 00:19:22] ppocr INFO:         name : MultiScaleDataSet
[2024/12/26 00:19:22] ppocr INFO:         transforms :
[2024/12/26 00:19:22] ppocr INFO:             DecodeImage :
[2024/12/26 00:19:22] ppocr INFO:                 channel_first : False
[2024/12/26 00:19:22] ppocr INFO:                 img_mode : BGR
[2024/12/26 00:19:22] ppocr INFO:             RecConAug :
[2024/12/26 00:19:22] ppocr INFO:                 ext_data_num : 2
[2024/12/26 00:19:22] ppocr INFO:                 image_shape : [144, 320, 3]
[2024/12/26 00:19:22] ppocr INFO:                 max_text_length : 25
[2024/12/26 00:19:22] ppocr INFO:                 prob : 0.5
[2024/12/26 00:19:22] ppocr INFO:             RecAug : None
[2024/12/26 00:19:22] ppocr INFO:             MultiLabelEncode :
[2024/12/26 00:19:22] ppocr INFO:                 gtc_encode : NRTRLabelEncode
[2024/12/26 00:19:22] ppocr INFO:             KeepKeys :
[2024/12/26 00:19:22] ppocr INFO:                 keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio']
[2024/12/26 00:19:22] ppocr INFO:     loader :
[2024/12/26 00:19:22] ppocr INFO:         batch_size_per_card : 50
[2024/12/26 00:19:22] ppocr INFO:         drop_last : True
[2024/12/26 00:19:22] ppocr INFO:         num_workers : 52
[2024/12/26 00:19:22] ppocr INFO:         shuffle : True
[2024/12/26 00:19:22] ppocr INFO:     sampler :
[2024/12/26 00:19:22] ppocr INFO:         divided_factor : [8, 16]
[2024/12/26 00:19:22] ppocr INFO:         first_bs : 50
[2024/12/26 00:19:22] ppocr INFO:         fix_bs : False
[2024/12/26 00:19:22] ppocr INFO:         is_training : True
[2024/12/26 00:19:22] ppocr INFO:         name : MultiScaleSampler
[2024/12/26 00:19:22] ppocr INFO:         scales : [[320, 112], [320, 144], [320, 176]]
[2024/12/26 00:19:22] ppocr INFO: profiler_options : None
[2024/12/26 00:19:22] ppocr INFO: train with paddle 3.0.0-beta1 and device Place(gpu:0)
[2024/12/26 00:19:22] ppocr INFO: Initialize indexs of datasets:['./dataset/ocr_rec/train.txt']
[2024/12/26 00:19:22] ppocr INFO: Initialize indexs of datasets:['./dataset/ocr_rec/val.txt']
W1226 00:19:22.586943 20068 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 8.9, Driver API Version: 12.6, Runtime API Version: 11.8
W1226 00:19:22.587942 20068 gpu_resources.cc:164] device: 0, cuDNN Version: 8.9.
[2024/12/26 00:19:22] ppocr INFO: train dataloader has 483 iters
[2024/12/26 00:19:22] ppocr INFO: valid dataloader has 27 iters
[2024/12/26 00:19:22] ppocr INFO: train from scratch
[2024/12/26 00:19:22] ppocr INFO: During the training process, after the 0th iteration, an evaluation is run every 50 iterations
[2024/12/26 00:19:26] ppocr INFO: epoch: [1/20000], global_step: 10, lr: 0.000002, acc: 0.000000, norm_edit_dis: 0.088496, CTCLoss: 78.706806, NRTRLoss: 2.975460, loss: 81.640991, avg_reader_cost: 0.09308 s, avg_batch_cost: 0.40738 s, avg_samples: 40.2, ips: 98.67942 samples/s, eta: 45 days, 13:08:04, max_mem_reserved: 13308 MB, max_mem_allocated: 10874 MB
[2024/12/26 00:19:30] ppocr INFO: epoch: [1/20000], global_step: 20, lr: 0.000004, acc: 0.000000, norm_edit_dis: 0.089483, CTCLoss: 78.610718, NRTRLoss: 2.823729, loss: 81.374657, avg_reader_cost: 0.08872 s, avg_batch_cost: 0.36821 s, avg_samples: 43.1, ips: 117.05319 samples/s, eta: 43 days, 8:34:44, max_mem_reserved: 13308 MB, max_mem_allocated: 10874 MB
[2024/12/26 00:19:34] ppocr INFO: epoch: [1/20000], global_step: 30, lr: 0.000008, acc: 0.000000, norm_edit_dis: 0.092362, CTCLoss: 78.000946, NRTRLoss: 2.486497, loss: 80.398026, avg_reader_cost: 0.11149 s, avg_batch_cost: 0.38249 s, avg_samples: 38.1, ips: 99.61123 samples/s, eta: 43 days, 3:49:51, max_mem_reserved: 13308 MB, max_mem_allocated: 10874 MB
[2024/12/26 00:19:38] ppocr INFO: epoch: [1/20000], global_step: 40, lr: 0.000012, acc: 0.000000, norm_edit_dis: 0.100448, CTCLoss: 76.662594, NRTRLoss: 2.300675, loss: 78.992405, avg_reader_cost: 0.11637 s, avg_batch_cost: 0.38905 s, avg_samples: 41.9, ips: 107.69955 samples/s, eta: 43 days, 5:51:21, max_mem_reserved: 13308 MB, max_mem_allocated: 10874 MB
[2024/12/26 00:19:41] ppocr INFO: epoch: [1/20000], global_step: 50, lr: 0.000016, acc: 0.000000, norm_edit_dis: 0.110871, CTCLoss: 72.200996, NRTRLoss: 2.175009, loss: 74.399948, avg_reader_cost: 0.09706 s, avg_batch_cost: 0.36275 s, avg_samples: 38.3, ips: 105.58313 samples/s, eta: 42 days, 16:57:25, max_mem_reserved: 13308 MB, max_mem_allocated: 10874 MB
eval model::   0%|                                                                                                                                                                                                                                  | 0/27 [00:00<?, ?it/s]Traceback (most recent call last):
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\tools\train.py", line 269, in <module>
    main(config, device, logger, vdl_writer, seed)
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\tools\train.py", line 222, in main
    program.train(
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\tools\program.py", line 486, in train
    cur_metric = eval(
                 ^^^^^
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\tools\program.py", line 689, in eval
    preds = model(images, data=batch[1:])
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\GitHub\USER-OCR\scr\train\.venv\Lib\site-packages\paddle\nn\layer\layers.py", line 1426, in __call__
    return self.forward(*inputs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\ppocr\modeling\architectures\base_model.py", line 99, in forward
    x = self.head(x, targets=data)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\GitHub\USER-OCR\scr\train\.venv\Lib\site-packages\paddle\nn\layer\layers.py", line 1426, in __call__
    return self.forward(*inputs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\ppocr\modeling\heads\rec_multi_head.py", line 139, in forward
    ctc_encoder = self.ctc_encoder(x)
                  ^^^^^^^^^^^^^^^^^^^
  File "c:\GitHub\USER-OCR\scr\train\.venv\Lib\site-packages\paddle\nn\layer\layers.py", line 1426, in __call__
    return self.forward(*inputs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\ppocr\modeling\necks\rnn.py", line 283, in forward
    x = self.encoder_reshape(x)
        ^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\GitHub\USER-OCR\scr\train\.venv\Lib\site-packages\paddle\nn\layer\layers.py", line 1426, in __call__
    return self.forward(*inputs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\GitHub\USER-OCR\scr\train\PaddleOCR\ppocr\modeling\necks\rnn.py", line 39, in forward
    assert H == 1
           ^^^^^^
AssertionError
eval model::   0%|                                                                                                                                                                                                                                  | 0/27 [00:01<?, ?it/s]

(.venv) C:\GitHub\USER-OCR\scr\train>

🏃‍♂️ Environment (运行环境)

1. Windows 11
2. Python 3.12.7
3. I9-13900k
4. GTX 4090

🌰 Minimal Reproducible Example (最小可复现问题的Demo)

python ./PaddleOCR/tools/train.py -c ./configs/ch_PP-OCRv4_rec_hgnet.yml

[GreatV]

The issue you are encountering ("AssertionError: assert H == 1") when changing the image height in the PaddleOCR PP-OCRv4 configuration is related to architectural constraints within the model's implementation. Specifically, the H == 1 check implies that the model expects the height of the feature maps (after processing by the backbone) to be 1. This is a common design choice in OCR models since many models are optimized for sequences, where the image height is fixed to a small value (e.g., 32 or 48), and the width can vary.

Here are some steps you can take to resolve the issue and align the configurations with your custom dataset:

---

1. Root Cause of H == 1 Assertion

The H == 1 assertion occurs because the RNN-based encoder in the CTCHead expects the height of the feature map to be exactly 1 after processing. This is achieved by setting a fixed input height and using down-sampling operations (e.g., pooling, strides) in the Backbone. If the input height (e.g., 144) is too large, the feature map height remains greater than 1, causing the assertion to fail.

---

2. Adjusting Image Height and Backbone Down-Sampling

If your images have a height of 144, ensure the Backbone down-samples the height sufficiently to make the feature map height 1. For example:

• Use a smaller input height: If possible, set the input height to a common value like 48 or 32. This reduces the need for excessive down-sampling.
• Modify the Backbone down-sampling:
  • Add more down-sampling layers (e.g., pooling or strides) in the Backbone configuration to reduce the feature map height to 1.
  • Check the configuration of PPHGNet_small (used in your configuration). You may need to adjust the stride or pooling behavior.

---

3. Configuration Changes

Update the configuration to ensure compatibility with your image dimensions.

Update image_shape:

In your case, the input image shape is [3, 144, 320]. To avoid the H == 1 assertion error, reduce the height or ensure the Backbone down-samples it appropriately.

Example:

Global:
  d2s_train_image_shape: [3, 48, 320]  # Update height to 48
Eval:
  dataset:
    transforms:
      - RecResizeImg:
          image_shape: [3, 48, 320]  # Match training shape

Modify Backbone Down-Sampling:

If you need to keep the height at 144, modify the Backbone to include additional down-sampling layers. For PPHGNet_small, check for stride or pooling configurations and adjust them.

Example:

Backbone:
  name: PPHGNet_small
  strides: [2, 2, 2]  # Example of increasing down-sampling

---

4. Verify the Dataset Transformations

Ensure that all dataset transformations are consistent with the input shape. Specifically:

• Training Data:
  • RecConAug and RecResizeImg must output images of the same shape as specified in image_shape.
• Evaluation Data:
  • Ensure RecResizeImg uses the same shape during evaluation.

Example:

Train:
  dataset:
    transforms:
      - RecConAug:
          image_shape: [144, 320, 3]  # Should match `d2s_train_image_shape`
Eval:
  dataset:
    transforms:
      - RecResizeImg:
          image_shape: [3, 144, 320]  # Match training image shape

---

5. Alternative: Resize Images

If increasing the input height is not critical for your dataset, resize your images to a smaller height (e.g., 48) to avoid architectural changes.

Example:

d2s_train_image_shape: [3, 48, 320]

---

6. Performance Considerations

• If your dataset contains very tall images, resizing them to a smaller height may introduce distortions. In such cases, carefully preprocess the dataset to preserve readability.
• Test the model's performance on resized images to ensure it meets your requirements.

---

7. Debugging Suggestions

• Print the feature map dimensions at different stages in the model to ensure the height becomes 1 before reaching the CTCHead.
• Use a debugging tool like paddle.summary or insert print statements in the Backbone and Head components to inspect the shapes.

---

8. References to Similar Issues

For more insights, you can refer to related issues in the PaddleOCR repository:

• Issue #9225: Discusses increasing image widths for long texts.
• Issue #13938: Discusses challenges with recognition accuracy for long texts.

---

Summary

To fix the H == 1 error:

1. Reduce the input height (e.g., to 48) or modify the Backbone for additional down-sampling.
2. Ensure all dataset transformations are consistent with the input shape.
3. Debug the feature map dimensions to verify the down-sampling behavior.

Let me know if you need further clarification or help with specific configurations!

---

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

[HungryFlo]

Thanks for your contributions and reply!

If I want to finetune the word recognition model for colomn-wise texts, like Chinese traditional books, what should I do, as my data are rather tall but not so wide.

Thanks in advance.