SVTRv2(Rec Sever) eval modes gives error

[amandhapola]

Hi ,
I have been trying to finetune the SVTRv2 model url on ICDAR 15 dataset . The training runs fine but while running the eval mode as specified in here I get the following issue .

System : `
Linux CentOS
python 3.10.14
Paddle Docker image : paddlepaddle/paddle:3.0.0rc1-gpu-cuda12.3-cudnn9.0-trt8.6

`

Training Command : !python3 tools/train.py \ -c configs/rec/SVTRv2/rec_repsvtr_gtc.yml \ -o Global.pretrained_model=./pretrain_models/svtrv2_1/openatom_rec_svtrv2_ch_train/best_accuracy

Eval command : !python3 tools/eval.py \ -c configs/rec/SVTRv2/rec_repsvtr_gtc.yml \ -o Global.pretrained_model=./output/rec_repsvtr_gtc/latest

Eval command output :
grep: warning: GREP_OPTIONS is deprecated; please use an alias or script [2025/03/20 06:09:13] ppocr WARNING: Skipping import of the encryption module. [2025/03/20 06:09:13] ppocr INFO: Architecture : [2025/03/20 06:09:13] ppocr INFO: Backbone : [2025/03/20 06:09:13] ppocr INFO: name : RepSVTR [2025/03/20 06:09:13] ppocr INFO: Head : [2025/03/20 06:09:13] ppocr INFO: head_list : [2025/03/20 06:09:13] ppocr INFO: CTCHead : [2025/03/20 06:09:13] ppocr INFO: Head : [2025/03/20 06:09:13] ppocr INFO: fc_decay : 1e-05 [2025/03/20 06:09:13] ppocr INFO: Neck : [2025/03/20 06:09:13] ppocr INFO: depth : 2 [2025/03/20 06:09:13] ppocr INFO: dims : 256 [2025/03/20 06:09:13] ppocr INFO: hidden_dims : 256 [2025/03/20 06:09:13] ppocr INFO: kernel_size : [1, 3] [2025/03/20 06:09:13] ppocr INFO: name : svtr [2025/03/20 06:09:13] ppocr INFO: use_guide : True [2025/03/20 06:09:13] ppocr INFO: NRTRHead : [2025/03/20 06:09:13] ppocr INFO: max_text_length : 25 [2025/03/20 06:09:13] ppocr INFO: nrtr_dim : 384 [2025/03/20 06:09:13] ppocr INFO: num_decoder_layers : 2 [2025/03/20 06:09:13] ppocr INFO: name : MultiHead [2025/03/20 06:09:13] ppocr INFO: Transform : None [2025/03/20 06:09:13] ppocr INFO: algorithm : SVTR_HGNet [2025/03/20 06:09:13] ppocr INFO: model_type : rec [2025/03/20 06:09:13] ppocr INFO: Eval : [2025/03/20 06:09:13] ppocr INFO: dataset : [2025/03/20 06:09:13] ppocr INFO: data_dir : ./dataset/rec_dataset/ [2025/03/20 06:09:13] ppocr INFO: label_file_list : ./dataset/rec_dataset/rec_gt_test.txt [2025/03/20 06:09:13] ppocr INFO: name : MultiScaleDataSet [2025/03/20 06:09:13] ppocr INFO: transforms : [2025/03/20 06:09:13] ppocr INFO: DecodeImage : [2025/03/20 06:09:13] ppocr INFO: channel_first : False [2025/03/20 06:09:13] ppocr INFO: img_mode : BGR [2025/03/20 06:09:13] ppocr INFO: MultiLabelEncode : [2025/03/20 06:09:13] ppocr INFO: gtc_encode : NRTRLabelEncode [2025/03/20 06:09:13] ppocr INFO: RecResizeImg : [2025/03/20 06:09:13] ppocr INFO: image_shape : [3, 48, 320] [2025/03/20 06:09:13] ppocr INFO: KeepKeys : [2025/03/20 06:09:13] ppocr INFO: keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio'] [2025/03/20 06:09:13] ppocr INFO: loader : [2025/03/20 06:09:13] ppocr INFO: batch_size_per_card : 8 [2025/03/20 06:09:13] ppocr INFO: drop_last : False [2025/03/20 06:09:13] ppocr INFO: num_workers : 0 [2025/03/20 06:09:13] ppocr INFO: shuffle : False [2025/03/20 06:09:13] ppocr INFO: Global : [2025/03/20 06:09:13] ppocr INFO: cal_metric_during_train : False [2025/03/20 06:09:13] ppocr INFO: character_dict_path : ppocr/utils/ppocr_keys_v1.txt [2025/03/20 06:09:13] ppocr INFO: checkpoints : None [2025/03/20 06:09:13] ppocr INFO: debug : False [2025/03/20 06:09:13] ppocr INFO: distributed : False [2025/03/20 06:09:13] ppocr INFO: epoch_num : 10 [2025/03/20 06:09:13] ppocr INFO: eval_batch_step : [0, 1000] [2025/03/20 06:09:13] ppocr INFO: infer_img : doc/imgs_words/ch/word_1.jpg [2025/03/20 06:09:13] ppocr INFO: infer_mode : False [2025/03/20 06:09:13] ppocr INFO: log_smooth_window : 20 [2025/03/20 06:09:13] ppocr INFO: max_text_length : 25 [2025/03/20 06:09:13] ppocr INFO: pretrained_model : ./output/rec_repsvtr_gtc/latest [2025/03/20 06:09:13] ppocr INFO: print_batch_step : 10 [2025/03/20 06:09:13] ppocr INFO: save_epoch_step : 10 [2025/03/20 06:09:13] ppocr INFO: save_inference_dir : None [2025/03/20 06:09:13] ppocr INFO: save_model_dir : ./output/rec_repsvtr_gtc [2025/03/20 06:09:13] ppocr INFO: save_res_path : ./output/rec/predicts_repsvtr.txt [2025/03/20 06:09:13] ppocr INFO: use_gpu : True [2025/03/20 06:09:13] ppocr INFO: use_space_char : True [2025/03/20 06:09:13] ppocr INFO: use_visualdl : False [2025/03/20 06:09:13] ppocr INFO: Loss : [2025/03/20 06:09:13] ppocr INFO: loss_config_list : [2025/03/20 06:09:13] ppocr INFO: CTCLoss : None [2025/03/20 06:09:13] ppocr INFO: NRTRLoss : None [2025/03/20 06:09:13] ppocr INFO: name : MultiLoss [2025/03/20 06:09:13] ppocr INFO: Metric : [2025/03/20 06:09:13] ppocr INFO: main_indicator : acc [2025/03/20 06:09:13] ppocr INFO: name : RecMetric [2025/03/20 06:09:13] ppocr INFO: Optimizer : [2025/03/20 06:09:13] ppocr INFO: beta1 : 0.9 [2025/03/20 06:09:13] ppocr INFO: beta2 : 0.999 [2025/03/20 06:09:13] ppocr INFO: epsilon : 1e-08 [2025/03/20 06:09:13] ppocr INFO: lr : [2025/03/20 06:09:13] ppocr INFO: learning_rate : 0.001 [2025/03/20 06:09:13] ppocr INFO: name : Cosine [2025/03/20 06:09:13] ppocr INFO: warmup_epoch : 5 [2025/03/20 06:09:13] ppocr INFO: name : AdamW [2025/03/20 06:09:13] ppocr INFO: no_weight_decay_name : norm [2025/03/20 06:09:13] ppocr INFO: one_dim_param_no_weight_decay : True [2025/03/20 06:09:13] ppocr INFO: weight_decay : 0.025 [2025/03/20 06:09:13] ppocr INFO: PostProcess : [2025/03/20 06:09:13] ppocr INFO: name : CTCLabelDecode [2025/03/20 06:09:13] ppocr INFO: Train : [2025/03/20 06:09:13] ppocr INFO: dataset : [2025/03/20 06:09:13] ppocr INFO: data_dir : ./dataset/rec_dataset/ [2025/03/20 06:09:13] ppocr INFO: ds_width : False [2025/03/20 06:09:13] ppocr INFO: ext_op_transform_idx : 1 [2025/03/20 06:09:13] ppocr INFO: label_file_list : ./dataset/rec_dataset/rec_gt_train.txt [2025/03/20 06:09:13] ppocr INFO: name : MultiScaleDataSet [2025/03/20 06:09:13] ppocr INFO: transforms : [2025/03/20 06:09:13] ppocr INFO: DecodeImage : [2025/03/20 06:09:13] ppocr INFO: channel_first : False [2025/03/20 06:09:13] ppocr INFO: img_mode : BGR [2025/03/20 06:09:13] ppocr INFO: RecAug : None [2025/03/20 06:09:13] ppocr INFO: MultiLabelEncode : [2025/03/20 06:09:13] ppocr INFO: gtc_encode : NRTRLabelEncode [2025/03/20 06:09:13] ppocr INFO: KeepKeys : [2025/03/20 06:09:13] ppocr INFO: keep_keys : ['image', 'label_ctc', 'label_gtc', 'length', 'valid_ratio'] [2025/03/20 06:09:13] ppocr INFO: loader : [2025/03/20 06:09:13] ppocr INFO: batch_size_per_card : 8 [2025/03/20 06:09:13] ppocr INFO: drop_last : True [2025/03/20 06:09:13] ppocr INFO: num_workers : 0 [2025/03/20 06:09:13] ppocr INFO: shuffle : True [2025/03/20 06:09:13] ppocr INFO: sampler : [2025/03/20 06:09:13] ppocr INFO: divided_factor : [8, 16] [2025/03/20 06:09:13] ppocr INFO: first_bs : 192 [2025/03/20 06:09:13] ppocr INFO: fix_bs : False [2025/03/20 06:09:13] ppocr INFO: is_training : True [2025/03/20 06:09:13] ppocr INFO: name : MultiScaleSampler [2025/03/20 06:09:13] ppocr INFO: scales : [[320, 32], [320, 48], [320, 64]] [2025/03/20 06:09:13] ppocr INFO: profiler_options : None [2025/03/20 06:09:13] ppocr INFO: train with paddle 3.0.0-rc1 and device Place(gpu:0) [2025/03/20 06:09:13] ppocr INFO: Initialize indexs of datasets:./dataset/rec_dataset/rec_gt_test.txt W0320 06:09:13.830448 7523 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 8.9, Driver API Version: 12.8, Runtime API Version: 12.3 W0320 06:09:13.831583 7523 gpu_resources.cc:164] device: 0, cuDNN Version: 9.1. [2025/03/20 06:09:14] ppocr INFO: load pretrain successful from ./output/rec_repsvtr_gtc/latest [2025/03/20 06:09:14] ppocr INFO: metric in ckpt *************** [2025/03/20 06:09:14] ppocr INFO: is_float16:False eval model:: 0%| | 0/260 [00:00<?, ?it/s]Exception in thread Thread-2 (_thread_loop): Traceback (most recent call last): File "/usr/lib/python3.10/threading.py", line 1016, in _bootstrap_inner self.run() File "/usr/lib/python3.10/threading.py", line 953, in run self._target(*self._args, **self._kwargs) File "/usr/local/lib/python3.10/dist-packages/paddle/io/dataloader/dataloader_iter.py", line 245, in _thread_loop batch = self._dataset_fetcher.fetch( File "/usr/local/lib/python3.10/dist-packages/paddle/io/dataloader/fetcher.py", line 77, in fetch data.append(self.dataset[idx]) File "/paddle/PaddleOCR/ppocr/data/simple_dataset.py", line 210, in __getitem__ img_height = properties[1] TypeError: 'int' object is not subscriptable

[GreatV]

The error you are encountering during the evaluation (TypeError: 'int' object is not subscriptable) suggests an issue with how the dataset is being processed. Specifically, it points to this line in simple_dataset.py:

img_height = properties[1]

Possible Causes and Solutions:

1. Incorrect Dataset Format:

   • The dataset must be in the correct format. Check if the rec_gt_test.txt file adheres to the expected structure, which is usually:

     path/to/image label
     

   • Ensure that each line includes both an image path and a corresponding label.

2. Issues with Dataset Parsing:

   • It seems the dataset loader (MultiScaleDataSet) expects a list or tuple, but instead, it encounters an integer (properties[1] failing).
   • Add print debugging inside simple_dataset.py:

     print("Properties: ", properties)

     This should help identify what properties contains and whether it's being parsed correctly.

3. Mismatch in MultiLabelEncode Configuration:

   • Your configuration uses NRTRLabelEncode for gtc_encode (ground truth encoding).
   • Try explicitly setting Global.character_dict_path to the correct dictionary file.
   • Also, check if NRTRLabelEncode is correctly configured to handle your dataset.

4. Incompatible Version Issues:

   • You are using PaddlePaddle 3.0.0-rc1, which is a release candidate.
     • Consider downgrading to a stable version like 2.6 to check for compatibility issues.
   • Reinstall dependencies in a clean environment to prevent package conflicts:

     pip install -r requirements.txt
     

5. Pretrained Model Usage:

   • The path provided to Global.pretrained_model (./output/rec_repsvtr_gtc/latest) might be incomplete or corrupted.
   • Test loading the model separately:

     import paddle
     paddle.load("./output/rec_repsvtr_gtc/latest")

     If this fails, the model checkpoint might not be saving correctly.

Next Steps:

• Verify rec_gt_test.txt contents and print debug statements.
• Confirm the dataset format matches PaddleOCR requirements.
• Try a different PaddlePaddle version (e.g., 2.6).
• Manually check model loading.

If the issue persists, please provide an example line from your dataset (rec_gt_test.txt).

---

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

⚠️ This is an AI-generated response. Please verify any critical information.