No Best Accuracy Model and Inference Failure with Exported PP-OCRv5_mobile_rec

[BaeSyoon]

Hello,

I trained the PP-OCRv5_mobile_rec model and exported it.
Here are the results after train and export:

1. There is no best accuracy model after training — only iter_epoch and latest checkpoints are saved.

2. I cannot run inference with my exported model. When I try, I get the following error:

I used 'test.py' code.

import os
import json
from paddleocr import PaddleOCR
from paddleocr import TextRecognition

INPUT_DIR = "data/ml_rc_crops"
OUTPUT_DIR = "output"
DICT_PATH = "digit_dict.txt"
os.makedirs(OUTPUT_DIR, exist_ok = True)

model = PaddleOCR(
    text_recognition_model_name="PP-OCRv5_mobile_rec",
    text_recognition_model_dir="./model/inference/",
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False,
) 

# model = TextRecognition(model_name="PP-OCRv5_mobile_rec", model_dir="./model")

results = []

file_list = sorted(
    [f for f in os.listdir(INPUT_DIR) if f.lower().endswith(".jpg")],
    key=lambda x: int(os.path.splitext(x)[0].split("_")[-1])  # img_00001.jpg → 1
)

for filename in file_list:
    img_path = os.path.join(INPUT_DIR, filename)
    output = model.predict(input=img_path)

    for res in output:
        res.print()
        data = res.json 
        results.append(data["res"])

with open(os.path.join(OUTPUT_DIR, "results.json"), "w", encoding="utf-8") as f:
    json.dump(results, f, ensure_ascii=False, indent=2)

[BaeSyoon]

For reference, Here is inference.yml code.

PreProcess:
  transform_ops:
  - DecodeImage:
      channel_first: false
      img_mode: BGR
  - CTCLabelEncode: null
  - RecResizeImg:
      image_shape:
      - 3
      - 32
      - 100
  - KeepKeys:
      keep_keys:
      - image
      - label
      - length
PostProcess:
  name: CTCLabelDecode
  character_dict:
  - '0'
  - '1'
  - '2'
  - '3'
  - '4'
  - '5'
  - '6'
  - '7'
  - '8'
  - '9'
  - ':'
  - ;
  - <
  - '='
  - '>'
  - '?'
  - '@'
  - 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
  - '['
  - \
  - ']'
  - ^
  - _
  - '`'
  - 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
  - '{'
  - '|'
  - '}'
  - '~'
  - '!'
  - '"'
  - '#'
  - $
  - '%'
  - '&'
  - ''''
  - (
  - )
  - '*'
  - +
  - ','
  - '-'
  - .
  - /
  - ' '

[lyan-chhay]

for best_accuracy model missing, it is due to you might set the eval_batch_step too big . let say let set [0,1000], it will train 1000 iteration before doing eval, if you have small dataset and train for small epoch, maybe it does not reach that iteration.

whereas for the error during inference part, I AM FACING THE SAME PROBLEM NOW ;( STILL LOOKING FOR SOLUTION

[BaeSyoon]

Thank you for your response. I didn't solve this problem yet, but i used another config file.