wrong inference on SVTR trained network

[lschmidt7]

I already tried all possibilities and i cannot do these things work.
someone can help me?

I trained a model using rec_svtrnet.yml on paddle 2.6.2 version (on 3.0.0 version, only the train works, inference requires model registered and it is not the case with svtr).
Well, i trained my model using this script

[train.py]

import sys
import runpy
from pathlib import Path

ROOT = Path(__file__).resolve().parents[1]
PADDLEOCR = ROOT / "PaddleOCR"

# adiciona PaddleOCR ao path
sys.path.insert(0, str(PADDLEOCR))

sys.argv = [
    "tools/program.py",
    "-c", "/config/rec_svtrnet.yml",
]

runpy.run_path(str(PADDLEOCR / "tools/train.py"), run_name="__main__")

[rec_svtrnet.yml]

Global:
  use_gpu: True
  epoch_num: 200
  log_smooth_window: 20
  print_batch_step: 10
  save_model_dir: /output/rec/train1
  save_epoch_step: 50
  # evaluation is run every 2000 iterations after the 0th iteration
  eval_batch_step: [0, 2000]
  cal_metric_during_train: True
  pretrained_model:
  checkpoints:
  save_inference_dir:
  use_visualdl: False
  infer_img: doc/imgs_words_en/word_10.png
  # for data or label process
  character_dict_path: /config/placas_dict.txt
  character_type: en
  max_text_length: 7
  infer_mode: False
  use_space_char: False
  save_res_path: /output/rec/predicts_svtr_tiny_train1.txt
  d2s_train_image_shape: [3, 64, 256]

Optimizer:
  name: AdamW
  beta1: 0.9
  beta2: 0.99
  epsilon: 1.e-8
  weight_decay: 0.05
  no_weight_decay_name: norm pos_embed
  one_dim_param_no_weight_decay: True
  lr:
    name: Cosine
    learning_rate: 0.0005
    warmup_epoch: 2

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: [64, 256]
    out_char_num: 36 # W//4 or W//8 or W/12
    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: /data/paddle_train/images
    label_file_list:
      - /data/paddle_train/labels.txt
    delimiter: "\t"
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      # - SVTRRecAug:
      #     aug_type: 0 # or 1
      - CTCLabelEncode: # Class handling label
      - SVTRRecResizeImg:
          image_shape: [3, 64, 256]
          padding: False
      - KeepKeys:
          keep_keys: ["image", "label", "length"] # dataloader will return list in this order
  loader:
    shuffle: True
    batch_size_per_card: 32
    drop_last: True
    num_workers: 8

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: /data/paddle_train/images
    label_file_list:
      - /data/paddle_train/labels.txt
    delimiter: "\t"
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - CTCLabelEncode: # Class handling label
      - SVTRRecResizeImg:
          image_shape: [3, 64, 256]
          padding: False
      - KeepKeys:
          keep_keys: ["image", "label", "length"] # dataloader will return list in this order
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: 32
    num_workers: 2

After the training (that reached acc: 1.0) i exported my model with this script (export.sh)

export FLAGS_enable_pir_api=0
python3 /PaddleOCR/tools/export_model.py \
    -c /config/rec_svtrnet.yml \
    -o Global.pretrained_model=/output/rec/train1/best_accuracy  Global.save_inference_dir=/output/inference

It generated the files listed below:

• best_accturacy.pdparams
• best_accuracy.states
• best_accuracy.pdopt

[inference.yml]

PreProcess:
  transform_ops:
  - DecodeImage:
      channel_first: false
      img_mode: BGR
  - CTCLabelEncode: null
  - SVTRRecResizeImg:
      image_shape:
      - 3
      - 64
      - 256
      padding: false
  - KeepKeys:
      keep_keys:
      - image
      - label
      - length
PostProcess:
  name: CTCLabelDecode
  character_dict:
  - A
  - B
  - C
  - D
  - E
  - F
  - G
  - H
  - I
  - J
  - K
  - L
  - M
  - N
  - O
  - P
  - Q
  - R
  - S
  - T
  - U
  - V
  - W
  - X
  - Y
  - Z
  - '0'
  - '1'
  - '2'
  - '3'
  - '4'
  - '5'
  - '6'
  - '7'
  - '8'
  - '9'

then i ran the inference with the following bash script

 python /PaddleOCR/tools/infer/predict_rec.py \
    --image_dir="/data/teste" \
    --rec_model_dir="/output/inference/" \
    --rec_char_dict_path="/config/placas_dict.txt" \
    --rec_image_shape="3, 64, 256"

and the result is some text completly different from the truth text.
and i cannot understand.
It seems to me that the pipeline and/or pre/post processing that is running on inference is different from that is running on the trainning.