使用本地数据集上训练的RE进行推理时出错,KeyError: 'relations'

[justgoonz]

🔎 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)

我输入的命令是

python3 ./tools/infer_kie_token_ser_re.py \
  -c  ./configs/kie/vi_layoutxlm/RE_front.yml\
  -o Architecture.Backbone.checkpoints=./output/10_14/RE_front/best_accuracy\
  Global.infer_img=./train_data/OCR_front/test/image\
  -c_ser ./configs/kie/vi_layoutxlm/RE_front.yml \
  -o_ser Architecture.Backbone.checkpoints=./output/10_14/RE_front/best_accuracy

然后遇到报错

Traceback (most recent call last):
  File "./tools/infer_kie_token_ser_re.py", line 216, in <module>
    result = ser_re_engine(data)
  File "./tools/infer_kie_token_ser_re.py", line 147, in __call__
    ser_results, ser_inputs = self.ser_engine(data)
  File "/home/user/1_xsz_ocr/PaddleOCR/tools/infer_kie_token_ser.py", line 115, in __call__
    batch = transform(data, self.ops)
  File "/home/user/1_xsz_ocr/PaddleOCR/ppocr/data/imaug/__init__.py", line 72, in transform
    data = op(data)
  File "/home/user/1_xsz_ocr/PaddleOCR/ppocr/data/imaug/vqa/token/vqa_token_relation.py", line 31, in __call__
    relations = data["relations"]
KeyError: 'relations'

我查看源码时发现,在infer时,data相关的代码段为

for idx, info in enumerate(infer_imgs):
            if config["Global"].get("infer_mode", None) is False:
                data_line = info.decode("utf-8")
                substr = data_line.strip("\n").split("\t")
                img_path = os.path.join(data_dir, substr[0])
                data = {"img_path": img_path, "label": substr[1]}
                print(substr)
            else:
                img_path = info
                data = {"img_path": img_path}

即data只有img_path和label字段.请问哪里出了问题

完整的输出如下

(ocr_pp) user@gpu2:~/1_xsz_ocr/PaddleOCR$ python3 ./tools/infer_kie_token_ser_re.py   -c  ./configs/kie/vi_layoutxlm/RE_front.yml  -o Architecture.Backbone.checkpoints=./output/10_14/RE_front/best_accuracy  Global.infer_img=./train_data/OCR_front/test/image  -c_ser ./configs/kie/vi_layoutxlm/RE_front.yml   -o_ser Architecture.Backbone.checkpoints=./output/10_14/RE_front/best_accuracy
[2024/10/14 19:20:16] ppocr INFO: ********** re config **********
[2024/10/14 19:20:16] ppocr INFO: Architecture : 
[2024/10/14 19:20:16] ppocr INFO:     Backbone : 
[2024/10/14 19:20:16] ppocr INFO:         checkpoints : ./output/10_14/RE_front/best_accuracy
[2024/10/14 19:20:16] ppocr INFO:         mode : vi
[2024/10/14 19:20:16] ppocr INFO:         name : LayoutXLMForRe
[2024/10/14 19:20:16] ppocr INFO:         pretrained : True
[2024/10/14 19:20:16] ppocr INFO:     Transform : None
[2024/10/14 19:20:16] ppocr INFO:     algorithm : LayoutXLM
[2024/10/14 19:20:16] ppocr INFO:     model_type : kie
[2024/10/14 19:20:16] ppocr INFO: Eval : 
[2024/10/14 19:20:16] ppocr INFO:     dataset : 
[2024/10/14 19:20:16] ppocr INFO:         data_dir : /home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/test/image
[2024/10/14 19:20:16] ppocr INFO:         label_file_list : ['/home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/test/test.json']
[2024/10/14 19:20:16] ppocr INFO:         name : SimpleDataSet
[2024/10/14 19:20:16] ppocr INFO:         transforms : 
[2024/10/14 19:20:16] ppocr INFO:             DecodeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 channel_first : False
[2024/10/14 19:20:16] ppocr INFO:                 img_mode : RGB
[2024/10/14 19:20:16] ppocr INFO:             VQATokenLabelEncode : 
[2024/10/14 19:20:16] ppocr INFO:                 algorithm : LayoutXLM
[2024/10/14 19:20:16] ppocr INFO:                 class_path : train_data/OCR_front/class_list_xfun.txt
[2024/10/14 19:20:16] ppocr INFO:                 contains_re : True
[2024/10/14 19:20:16] ppocr INFO:                 order_method : tb-yx
[2024/10/14 19:20:16] ppocr INFO:                 use_textline_bbox_info : True
[2024/10/14 19:20:16] ppocr INFO:             VQATokenPad : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:                 return_attention_mask : True
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenRelation : None
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenChunk : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:             TensorizeEntitiesRelations : None
[2024/10/14 19:20:16] ppocr INFO:             Resize : 
[2024/10/14 19:20:16] ppocr INFO:                 size : [512, 512]
[2024/10/14 19:20:16] ppocr INFO:             NormalizeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 mean : [128, 128, 128]
[2024/10/14 19:20:16] ppocr INFO:                 order : hwc
[2024/10/14 19:20:16] ppocr INFO:                 scale : 1
[2024/10/14 19:20:16] ppocr INFO:                 std : [60, 60, 60]
[2024/10/14 19:20:16] ppocr INFO:             ToCHWImage : None
[2024/10/14 19:20:16] ppocr INFO:             KeepKeys : 
[2024/10/14 19:20:16] ppocr INFO:                 keep_keys : ['input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'entities', 'relations']
[2024/10/14 19:20:16] ppocr INFO:     loader : 
[2024/10/14 19:20:16] ppocr INFO:         batch_size_per_card : 8
[2024/10/14 19:20:16] ppocr INFO:         drop_last : False
[2024/10/14 19:20:16] ppocr INFO:         num_workers : 8
[2024/10/14 19:20:16] ppocr INFO:         shuffle : False
[2024/10/14 19:20:16] ppocr INFO: Global : 
[2024/10/14 19:20:16] ppocr INFO:     cal_metric_during_train : False
[2024/10/14 19:20:16] ppocr INFO:     epoch_num : 130
[2024/10/14 19:20:16] ppocr INFO:     eval_batch_step : [0, 19]
[2024/10/14 19:20:16] ppocr INFO:     infer_img : ./train_data/OCR_front/test/image
[2024/10/14 19:20:16] ppocr INFO:     kie_det_model_dir : None
[2024/10/14 19:20:16] ppocr INFO:     kie_rec_model_dir : None
[2024/10/14 19:20:16] ppocr INFO:     log_smooth_window : 10
[2024/10/14 19:20:16] ppocr INFO:     print_batch_step : 10
[2024/10/14 19:20:16] ppocr INFO:     save_epoch_step : 2000
[2024/10/14 19:20:16] ppocr INFO:     save_inference_dir : None
[2024/10/14 19:20:16] ppocr INFO:     save_model_dir : ./output/10_14/RE_front
[2024/10/14 19:20:16] ppocr INFO:     save_res_path : ./output/10_14/re/OCR/with_gt
[2024/10/14 19:20:16] ppocr INFO:     seed : 2022
[2024/10/14 19:20:16] ppocr INFO:     use_gpu : True
[2024/10/14 19:20:16] ppocr INFO:     use_visualdl : False
[2024/10/14 19:20:16] ppocr INFO: Loss : 
[2024/10/14 19:20:16] ppocr INFO:     key : loss
[2024/10/14 19:20:16] ppocr INFO:     name : LossFromOutput
[2024/10/14 19:20:16] ppocr INFO:     reduction : mean
[2024/10/14 19:20:16] ppocr INFO: Metric : 
[2024/10/14 19:20:16] ppocr INFO:     main_indicator : hmean
[2024/10/14 19:20:16] ppocr INFO:     name : VQAReTokenMetric
[2024/10/14 19:20:16] ppocr INFO: Optimizer : 
[2024/10/14 19:20:16] ppocr INFO:     beta1 : 0.9
[2024/10/14 19:20:16] ppocr INFO:     beta2 : 0.999
[2024/10/14 19:20:16] ppocr INFO:     clip_norm : 10
[2024/10/14 19:20:16] ppocr INFO:     lr : 
[2024/10/14 19:20:16] ppocr INFO:         learning_rate : 5e-05
[2024/10/14 19:20:16] ppocr INFO:         warmup_epoch : 10
[2024/10/14 19:20:16] ppocr INFO:     name : AdamW
[2024/10/14 19:20:16] ppocr INFO:     regularizer : 
[2024/10/14 19:20:16] ppocr INFO:         factor : 0.0
[2024/10/14 19:20:16] ppocr INFO:         name : L2
[2024/10/14 19:20:16] ppocr INFO: PostProcess : 
[2024/10/14 19:20:16] ppocr INFO:     name : VQAReTokenLayoutLMPostProcess
[2024/10/14 19:20:16] ppocr INFO: Train : 
[2024/10/14 19:20:16] ppocr INFO:     dataset : 
[2024/10/14 19:20:16] ppocr INFO:         data_dir : /home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/train/image
[2024/10/14 19:20:16] ppocr INFO:         label_file_list : ['/home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/train/train.json']
[2024/10/14 19:20:16] ppocr INFO:         name : SimpleDataSet
[2024/10/14 19:20:16] ppocr INFO:         ratio_list : [1.0]
[2024/10/14 19:20:16] ppocr INFO:         transforms : 
[2024/10/14 19:20:16] ppocr INFO:             DecodeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 channel_first : False
[2024/10/14 19:20:16] ppocr INFO:                 img_mode : RGB
[2024/10/14 19:20:16] ppocr INFO:             VQATokenLabelEncode : 
[2024/10/14 19:20:16] ppocr INFO:                 algorithm : LayoutXLM
[2024/10/14 19:20:16] ppocr INFO:                 class_path : train_data/OCR_front/class_list_xfun.txt
[2024/10/14 19:20:16] ppocr INFO:                 contains_re : True
[2024/10/14 19:20:16] ppocr INFO:                 order_method : tb-yx
[2024/10/14 19:20:16] ppocr INFO:                 use_textline_bbox_info : True
[2024/10/14 19:20:16] ppocr INFO:             VQATokenPad : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:                 return_attention_mask : True
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenRelation : None
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenChunk : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:             TensorizeEntitiesRelations : None
[2024/10/14 19:20:16] ppocr INFO:             Resize : 
[2024/10/14 19:20:16] ppocr INFO:                 size : [512, 512]
[2024/10/14 19:20:16] ppocr INFO:             NormalizeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 mean : [128, 128, 128]
[2024/10/14 19:20:16] ppocr INFO:                 order : hwc
[2024/10/14 19:20:16] ppocr INFO:                 scale : 1
[2024/10/14 19:20:16] ppocr INFO:                 std : [60, 60, 60]
[2024/10/14 19:20:16] ppocr INFO:             ToCHWImage : None
[2024/10/14 19:20:16] ppocr INFO:             KeepKeys : 
[2024/10/14 19:20:16] ppocr INFO:                 keep_keys : ['input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'entities', 'relations']
[2024/10/14 19:20:16] ppocr INFO:     loader : 
[2024/10/14 19:20:16] ppocr INFO:         batch_size_per_card : 2
[2024/10/14 19:20:16] ppocr INFO:         drop_last : False
[2024/10/14 19:20:16] ppocr INFO:         num_workers : 4
[2024/10/14 19:20:16] ppocr INFO:         shuffle : True
[2024/10/14 19:20:16] ppocr INFO: 

[2024/10/14 19:20:16] ppocr INFO: ********** ser config **********
[2024/10/14 19:20:16] ppocr INFO: Architecture : 
[2024/10/14 19:20:16] ppocr INFO:     Backbone : 
[2024/10/14 19:20:16] ppocr INFO:         checkpoints : ./output/10_14/RE_front/best_accuracy
[2024/10/14 19:20:16] ppocr INFO:         mode : vi
[2024/10/14 19:20:16] ppocr INFO:         name : LayoutXLMForRe
[2024/10/14 19:20:16] ppocr INFO:         pretrained : True
[2024/10/14 19:20:16] ppocr INFO:     Transform : None
[2024/10/14 19:20:16] ppocr INFO:     algorithm : LayoutXLM
[2024/10/14 19:20:16] ppocr INFO:     model_type : kie
[2024/10/14 19:20:16] ppocr INFO: Eval : 
[2024/10/14 19:20:16] ppocr INFO:     dataset : 
[2024/10/14 19:20:16] ppocr INFO:         data_dir : /home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/test/image
[2024/10/14 19:20:16] ppocr INFO:         label_file_list : ['/home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/test/test.json']
[2024/10/14 19:20:16] ppocr INFO:         name : SimpleDataSet
[2024/10/14 19:20:16] ppocr INFO:         transforms : 
[2024/10/14 19:20:16] ppocr INFO:             DecodeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 channel_first : False
[2024/10/14 19:20:16] ppocr INFO:                 img_mode : RGB
[2024/10/14 19:20:16] ppocr INFO:             VQATokenLabelEncode : 
[2024/10/14 19:20:16] ppocr INFO:                 algorithm : LayoutXLM
[2024/10/14 19:20:16] ppocr INFO:                 class_path : train_data/OCR_front/class_list_xfun.txt
[2024/10/14 19:20:16] ppocr INFO:                 contains_re : True
[2024/10/14 19:20:16] ppocr INFO:                 order_method : tb-yx
[2024/10/14 19:20:16] ppocr INFO:                 use_textline_bbox_info : True
[2024/10/14 19:20:16] ppocr INFO:             VQATokenPad : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:                 return_attention_mask : True
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenRelation : None
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenChunk : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:             TensorizeEntitiesRelations : None
[2024/10/14 19:20:16] ppocr INFO:             Resize : 
[2024/10/14 19:20:16] ppocr INFO:                 size : [512, 512]
[2024/10/14 19:20:16] ppocr INFO:             NormalizeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 mean : [128, 128, 128]
[2024/10/14 19:20:16] ppocr INFO:                 order : hwc
[2024/10/14 19:20:16] ppocr INFO:                 scale : 1
[2024/10/14 19:20:16] ppocr INFO:                 std : [60, 60, 60]
[2024/10/14 19:20:16] ppocr INFO:             ToCHWImage : None
[2024/10/14 19:20:16] ppocr INFO:             KeepKeys : 
[2024/10/14 19:20:16] ppocr INFO:                 keep_keys : ['input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'entities', 'relations']
[2024/10/14 19:20:16] ppocr INFO:     loader : 
[2024/10/14 19:20:16] ppocr INFO:         batch_size_per_card : 8
[2024/10/14 19:20:16] ppocr INFO:         drop_last : False
[2024/10/14 19:20:16] ppocr INFO:         num_workers : 8
[2024/10/14 19:20:16] ppocr INFO:         shuffle : False
[2024/10/14 19:20:16] ppocr INFO: Global : 
[2024/10/14 19:20:16] ppocr INFO:     cal_metric_during_train : False
[2024/10/14 19:20:16] ppocr INFO:     epoch_num : 130
[2024/10/14 19:20:16] ppocr INFO:     eval_batch_step : [0, 19]
[2024/10/14 19:20:16] ppocr INFO:     infer_img : /home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/test/image
[2024/10/14 19:20:16] ppocr INFO:     kie_det_model_dir : None
[2024/10/14 19:20:16] ppocr INFO:     kie_rec_model_dir : None
[2024/10/14 19:20:16] ppocr INFO:     log_smooth_window : 10
[2024/10/14 19:20:16] ppocr INFO:     print_batch_step : 10
[2024/10/14 19:20:16] ppocr INFO:     save_epoch_step : 2000
[2024/10/14 19:20:16] ppocr INFO:     save_inference_dir : None
[2024/10/14 19:20:16] ppocr INFO:     save_model_dir : ./output/10_14/RE_front
[2024/10/14 19:20:16] ppocr INFO:     save_res_path : ./output/10_14/re/OCR/with_gt
[2024/10/14 19:20:16] ppocr INFO:     seed : 2022
[2024/10/14 19:20:16] ppocr INFO:     use_gpu : True
[2024/10/14 19:20:16] ppocr INFO:     use_visualdl : False
[2024/10/14 19:20:16] ppocr INFO: Loss : 
[2024/10/14 19:20:16] ppocr INFO:     key : loss
[2024/10/14 19:20:16] ppocr INFO:     name : LossFromOutput
[2024/10/14 19:20:16] ppocr INFO:     reduction : mean
[2024/10/14 19:20:16] ppocr INFO: Metric : 
[2024/10/14 19:20:16] ppocr INFO:     main_indicator : hmean
[2024/10/14 19:20:16] ppocr INFO:     name : VQAReTokenMetric
[2024/10/14 19:20:16] ppocr INFO: Optimizer : 
[2024/10/14 19:20:16] ppocr INFO:     beta1 : 0.9
[2024/10/14 19:20:16] ppocr INFO:     beta2 : 0.999
[2024/10/14 19:20:16] ppocr INFO:     clip_norm : 10
[2024/10/14 19:20:16] ppocr INFO:     lr : 
[2024/10/14 19:20:16] ppocr INFO:         learning_rate : 5e-05
[2024/10/14 19:20:16] ppocr INFO:         warmup_epoch : 10
[2024/10/14 19:20:16] ppocr INFO:     name : AdamW
[2024/10/14 19:20:16] ppocr INFO:     regularizer : 
[2024/10/14 19:20:16] ppocr INFO:         factor : 0.0
[2024/10/14 19:20:16] ppocr INFO:         name : L2
[2024/10/14 19:20:16] ppocr INFO: PostProcess : 
[2024/10/14 19:20:16] ppocr INFO:     name : VQAReTokenLayoutLMPostProcess
[2024/10/14 19:20:16] ppocr INFO: Train : 
[2024/10/14 19:20:16] ppocr INFO:     dataset : 
[2024/10/14 19:20:16] ppocr INFO:         data_dir : /home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/train/image
[2024/10/14 19:20:16] ppocr INFO:         label_file_list : ['/home/user/1_xsz_ocr/PaddleOCR/train_data/OCR_front/train/train.json']
[2024/10/14 19:20:16] ppocr INFO:         name : SimpleDataSet
[2024/10/14 19:20:16] ppocr INFO:         ratio_list : [1.0]
[2024/10/14 19:20:16] ppocr INFO:         transforms : 
[2024/10/14 19:20:16] ppocr INFO:             DecodeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 channel_first : False
[2024/10/14 19:20:16] ppocr INFO:                 img_mode : RGB
[2024/10/14 19:20:16] ppocr INFO:             VQATokenLabelEncode : 
[2024/10/14 19:20:16] ppocr INFO:                 algorithm : LayoutXLM
[2024/10/14 19:20:16] ppocr INFO:                 class_path : train_data/OCR_front/class_list_xfun.txt
[2024/10/14 19:20:16] ppocr INFO:                 contains_re : True
[2024/10/14 19:20:16] ppocr INFO:                 order_method : tb-yx
[2024/10/14 19:20:16] ppocr INFO:                 use_textline_bbox_info : True
[2024/10/14 19:20:16] ppocr INFO:             VQATokenPad : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:                 return_attention_mask : True
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenRelation : None
[2024/10/14 19:20:16] ppocr INFO:             VQAReTokenChunk : 
[2024/10/14 19:20:16] ppocr INFO:                 max_seq_len : 512
[2024/10/14 19:20:16] ppocr INFO:             TensorizeEntitiesRelations : None
[2024/10/14 19:20:16] ppocr INFO:             Resize : 
[2024/10/14 19:20:16] ppocr INFO:                 size : [512, 512]
[2024/10/14 19:20:16] ppocr INFO:             NormalizeImage : 
[2024/10/14 19:20:16] ppocr INFO:                 mean : [128, 128, 128]
[2024/10/14 19:20:16] ppocr INFO:                 order : hwc
[2024/10/14 19:20:16] ppocr INFO:                 scale : 1
[2024/10/14 19:20:16] ppocr INFO:                 std : [60, 60, 60]
[2024/10/14 19:20:16] ppocr INFO:             ToCHWImage : None
[2024/10/14 19:20:16] ppocr INFO:             KeepKeys : 
[2024/10/14 19:20:16] ppocr INFO:                 keep_keys : ['input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'entities', 'relations']
[2024/10/14 19:20:16] ppocr INFO:     loader : 
[2024/10/14 19:20:16] ppocr INFO:         batch_size_per_card : 2
[2024/10/14 19:20:16] ppocr INFO:         drop_last : False
[2024/10/14 19:20:16] ppocr INFO:         num_workers : 4
[2024/10/14 19:20:16] ppocr INFO:         shuffle : True
[2024/10/14 19:20:16] ppocr INFO: train with paddle 3.0.0-beta1 and device Place(gpu:0)
INFO:datasets:PyTorch version 2.4.1 available.
W1014 19:20:18.021422 715557 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 12.2, Runtime API Version: 11.8
W1014 19:20:18.022987 715557 gpu_resources.cc:164] device: 0, cuDNN Version: 8.7.
[2024/10/14 19:20:21] ppocr INFO: resume from ./output/10_14/RE_front/best_accuracy
[2024/10/14 19:20:21] ppocr WARNING: The first GPU is used for inference by default, GPU ID: 0
[2024/10/14 19:20:22] ppocr WARNING: The first GPU is used for inference by default, GPU ID: 0
[2024-10-14 19:20:23,026] [    INFO] - Already cached /home/user/.paddlenlp/models/layoutxlm-base-uncased/sentencepiece.bpe.model
[2024-10-14 19:20:23,614] [    INFO] - tokenizer config file saved in /home/user/.paddlenlp/models/layoutxlm-base-uncased/tokenizer_config.json
[2024-10-14 19:20:23,617] [    INFO] - Special tokens file saved in /home/user/.paddlenlp/models/layoutxlm-base-uncased/special_tokens_map.json
[2024/10/14 19:20:26] ppocr INFO: resume from ./output/10_14/RE_front/best_accuracy
Traceback (most recent call last):
  File "./tools/infer_kie_token_ser_re.py", line 216, in <module>
    result = ser_re_engine(data)
  File "./tools/infer_kie_token_ser_re.py", line 147, in __call__
    ser_results, ser_inputs = self.ser_engine(data)
  File "/home/user/1_xsz_ocr/PaddleOCR/tools/infer_kie_token_ser.py", line 115, in __call__
    batch = transform(data, self.ops)
  File "/home/user/1_xsz_ocr/PaddleOCR/ppocr/data/imaug/__init__.py", line 72, in transform
    data = op(data)
  File "/home/user/1_xsz_ocr/PaddleOCR/ppocr/data/imaug/vqa/token/vqa_token_relation.py", line 31, in __call__
    relations = data["relations"]
KeyError: 'relations'

🏃‍♂️ Environment (运行环境)

paddlenlp                 2.5.2                    pypi_0    pypi
paddleocr                 2.8.1                    pypi_0    pypi
paddlepaddle-gpu          3.0.0b1                  pypi_0    pypi

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

完整的命令如下,在训练阶段没有异常,错误仅仅发生在推理阶段

conda activate ocr_pp

cd /home/user/1_xsz_ocr/PaddleOCR

python3 tools/train.py -c configs/kie/vi_layoutxlm/RE_front.yml

python3 ./tools/infer_kie_token_ser_re.py \
  -c  ./configs/kie/vi_layoutxlm/RE_front.yml\
  -o Architecture.Backbone.checkpoints=./output/10_14/RE_front/best_accuracy\
  Global.infer_img=./train_data/OCR_front/test/image\
  -c_ser ./configs/kie/vi_layoutxlm/RE_front.yml \
  -o_ser Architecture.Backbone.checkpoints=./output/10_14/RE_front/best_accuracy

[GreatV]

根据提供的错误信息和代码片段，推理阶段的 KeyError: 'relations' 问题可能是由于以下原因导致的：

---

问题分析

1. data 字典中缺少 relations 字段：

   • 在推理阶段，data 的构造是基于以下代码：

     if config["Global"].get("infer_mode", None) is False:
         data = {"img_path": img_path, "label": substr[1]}
     else:
         data = {"img_path": img_path}

     从这里可以看出，在推理模式下 (infer_mode=True)，data 仅包含 img_path 字段，而没有 relations 字段。

   • 然而，推理过程中调用了 VQAReTokenRelation，该模块在尝试访问 data["relations"] 时抛出了 KeyError。

2. 配置文件中 transforms 包含与关系抽取（RE）相关的处理：

   • 根据日志中的配置：

     transforms:
         - VQATokenLabelEncode:
             contains_re: True
         - VQAReTokenRelation: None
         - VQAReTokenChunk:
             max_seq_len: 512
         - KeepKeys:
             keep_keys: ['input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'entities', 'relations']

     这里的 VQATokenLabelEncode 和 VQAReTokenRelation 是与关系抽取相关的变换操作，其中 VQAReTokenRelation 明显依赖 relations 字段。

   • 在训练阶段，relations 是从标注文件（如 JSON 文件）中加载的。然而，在推理阶段，输入数据可能没有 relations 信息，这会导致问题。

3. 推理阶段的数据准备不完整：

   • 您的推理数据仅包含图片路径和标签（可选），但缺少与关系抽取相关的关键信息（如 relations 和 entities）。
   • 这与训练阶段的数据结构不一致，导致 VQAReTokenRelation 无法正确处理输入。

---

解决方案

方法 1：禁用与 relations 相关的操作

在推理阶段，如果 relations 字段不需要（仅需要文本实体的分类结果），可以通过修改配置文件以禁用与关系抽取相关的处理：

1. 修改 configs/kie/vi_layoutxlm/RE_front.yml 文件中的 transforms 部分：

   transforms:
       - VQATokenLabelEncode:
           contains_re: False  # 设置为 False，禁用关系抽取相关的处理
       - VQAReTokenChunk:
           max_seq_len: 512
       - KeepKeys:
           keep_keys: ['input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'entities']  # 移除 'relations'

2. 确保推理阶段不调用 VQAReTokenRelation 方法。

---

方法 2：补充 relations 字段

如果您确实需要推理阶段的关系抽取结果，则需要在推理数据中补充 relations 信息。

1. 修改推理数据的准备逻辑，确保 data 包含 relations 字段。可以从标注文件中加载或通过其他方式生成。
2. 确保推理数据与训练数据的结构一致。

例如，如果标注是 JSON 格式，请确保每个实体都包含 relations 信息：

{
    "img_path": "path/to/image.jpg",
    "entities": [{"text": "Entity1", "bbox": [x1, y1, x2, y2]}],
    "relations": [{"head": 0, "tail": 1, "type": "related"}]
}

---

方法 3：检查配置文件的兼容性

检查配置文件是否适用于推理阶段。训练配置和推理配置可能需要分开管理：

1. 创建一个新的推理配置文件（如 RE_front_infer.yml），去掉训练专用的部分，例如 contains_re: True 和 VQAReTokenRelation。
2. 使用新的配置文件运行推理：

   python3 ./tools/infer_kie_token_ser_re.py \
       -c ./configs/kie/vi_layoutxlm/RE_front_infer.yml \
       -o Architecture.Backbone.checkpoints=./output/10_14/RE_front/best_accuracy \
       Global.infer_img=./train_data/OCR_front/test/image

---

总结

推理阶段 KeyError: 'relations' 的根本原因在于推理数据缺少 relations 字段，而配置文件要求处理此字段。可以通过以下方式解决：

1. 禁用关系抽取相关的处理逻辑（适用于不需要关系抽取结果的场景）。
2. 补充推理数据中的 relations 字段（适用于需要关系抽取结果的场景）。
3. 使用专门的推理配置文件，避免训练和推理配置之间的不兼容问题。

根据您的实际需求选择适当的方法，修改配置或数据以解决问题。

---

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