code bug while training parseq in paddleOCR

[SleepEarlyLiveLong]

Hi, thank you for your outstanding work! I encountered an error while trying to reproduce your work based on the PaddleOCR framework. It seems to be a bug in the code. Please take a look at the specific information below:

Here is the config file: ` Global: use_gpu: True epoch_num: 100 log_smooth_window: 20 print_batch_step: 5 save_model_dir: ./output/rec/parseq_cty_v1 save_epoch_step: 3 eval_batch_step: [0, 500] cal_metric_during_train: True pretrained_model: checkpoints: save_inference_dir: use_visualdl: False infer_img: doc/imgs_words_en/word_10.png character_dict_path: ppocr/utils/dict/parseq_dict_mixlang.txt character_type: ch max_text_length: 35 # 35 num_heads: 8 infer_mode: False use_space_char: False save_res_path: ./output/rec/predicts_parseq.txt

Optimizer: name: Adam beta1: 0.9 beta2: 0.999 lr: name: OneCycle max_lr: 0.0007

Architecture: model_type: rec algorithm: ParseQ in_channels: 3 Transform: Backbone: name: ViTParseQ img_size: [32, 128] patch_size: [4, 8] embed_dim: 384 depth: 12 num_heads: 6 mlp_ratio: 4 in_channels: 3 Head: name: ParseQHead # Architecture max_text_length: 35 embed_dim: 384 dec_num_heads: 12 dec_mlp_ratio: 4 dec_depth: 1 # Training perm_num: 6 perm_forward: true perm_mirrored: true dropout: 0.1 # Decoding mode (test) decode_ar: true refine_iters: 1

Loss: name: ParseQLoss

PostProcess: name: ParseQLabelDecode

Metric: name: RecMetric main_indicator: acc is_filter: True

Train: dataset: name: LMDBDataSet data_dir: /mnt/workspace/workgroup/sukunming/code/parseq/data/train/synth transforms: - DecodeImage: # load image img_mode: BGR channel_first: False - ParseQRecAug: aug_type: 0 # or 1 - ParseQLabelEncode: - SVTRRecResizeImg: image_shape: [3, 32, 128] padding: False - KeepKeys: keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order loader: shuffle: True batch_size_per_card: 192 drop_last: True num_workers: 4

Eval: dataset: name: LMDBDataSet data_dir: /mnt/workspace/workgroup/sukunming/code/parseq/data/val_label_data/synth transforms: - DecodeImage: # load image img_mode: BGR channel_first: False - ParseQLabelEncode: # Class handling label - SVTRRecResizeImg: image_shape: [3, 32, 128] padding: False - KeepKeys: keep_keys: ['image', 'label', 'length'] loader: shuffle: False drop_last: False batch_size_per_card: 384 num_workers: 4

`

Here is the what 'data_dir' looks like, each folder includes two file: 'data.mdb' and 'lock.mdb', which are generated by 'python tools/create_lmdb_dataset.py /path/to/img/root /path/to/gt /path/to/save/lmdb':

Based on infos above, I run order "python3 tools/train.py -c configs/rec/rec_vit_parseq_cty_v1.yml" and encountered a bug at 'ppocr/modeling/heads/rec_parseq_head.py' Line 498:

where targets[0]:

targets[1]:

And:

Is there something wrong with the code? How to solve the problem? Thank you a lot!