ValueError: not find model.pdmodel or inference.pdmodel in output/v3_en_mobile/best_model

[kishore-work-hard]

I found answers in Chinese but I don't know Chinese

But i trained a model successfully. and in output/v3_en_mobile/best_model path folder i have model.pdopt and model.pdparams.

after training my recognition model
i used the command to test my script -

python tools/infer/predict_rec.py --image_dir="img/a (6).jpg" --rec_model_dir="output/v3_en_mobile/best_model" --rec_image_shape="3, 32, 100" --rec_char_dict_path="ppocr/utils/en_dict.txt"

and got this error -

Traceback (most recent call last):
  File "/home/elvicto/.ghostshell/Test/OCR/PaddleOCR/tools/infer/predict_rec.py", line 756, in <module>
    main(utility.parse_args())
  File "/home/elvicto/.ghostshell/Test/OCR/PaddleOCR/tools/infer/predict_rec.py", line 718, in main
    text_recognizer = TextRecognizer(args)
  File "/home/elvicto/.ghostshell/Test/OCR/PaddleOCR/tools/infer/predict_rec.py", line 149, in __init__
    ) = utility.create_predictor(args, "rec", logger)
  File "/home/elvicto/.ghostshell/Test/OCR/PaddleOCR/tools/infer/utility.py", line 214, in create_predictor
    raise ValueError(
ValueError: not find model.pdmodel or inference.pdmodel in output/v3_en_mobile/best_model

dict used - ppocr/utils/en_dict.txt

after getting this error I tried the below code.
and using the code - I figured it that this script downloads file automatically from this URL
https://paddleocr.bj.bcebos.com/PP-OCRv4/chinese/ch_PP-OCRv4_rec_infer.tar

import os
import cv2
import numpy as np
import paddle
from paddleocr import PaddleOCR

# Define paths
model_dir = "D:/DATA/OCR_MODEL/best_model"
image_path = "../images/a (2).jpg"

# Load custom OCR model
ocr = PaddleOCR(
    rec_model_dir=model_dir,
    rec_char_dict_path=os.path.join(model_dir, "dict.txt"),  # Update with actual dictionary path if needed
    use_space_char=False
)

# Load image
img = cv2.imread(image_path)
if img is None:
    raise ValueError(f"Image at path {image_path} could not be loaded.")

# Perform OCR
result = ocr.ocr(img, det=False)

print(result)

But I want english model. so i replaced it with en and english
should i be downloading the file from this -
https://paddleocr.bj.bcebos.com/PP-OCRv4/english/en_PP-OCRv4_rec_infer.tar

what i mean is...
after custom training my rec model. should I be using this files from en_PP-OCRv4_rec_infer.tar after downloading it.?

Or

should this file be used for pretrained model and should not be used by custom trained model?
please explain what's happening. and what is this inference.pdmodel and inference.pdiparams ?

[SWHL]

You can refer this doc

inference.pdmodel and inference.pdiparams are converted from the model.pdopt and model.params.

[kishore-work-hard]

You can refer this doc

inference.pdmodel and inference.pdiparams are converted from the model.pdopt and model.params.

SO YOU ARE TELLING ME NOT TO DOWNLOAD THE MISSING FILES FILE FROM THE URL. ?
https://paddleocr.bj.bcebos.com/PP-OCRv4/english/en_PP-OCRv4_rec_infer.tar

I didn't get the file inference.pdmodel and inference.pdiparams after training.

can you tell me how to get it?

[pcycccccc]

你是拿训练完的模型文件进行测试，所以出现了这样的问题。
正确做法应该是：将你的训练完的模型转为推理模型，示例指令为：python tools/export_model.py -c "./configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml" -o Global.pretrained_model="./output/v3_en_rec/best_accuracy.pdparams" Global.save_inference_dir="./inference _model/rec/"
你需要根据你的实际情况，更换一下路径。然后再利用你之前的测试脚本指令：python tools/infer/predict_rec.py来指向你的推理模型

[pcycccccc]

@kishore-work-hard
You encountered issues with finding model.pdmodel or inference.pdmodel because you are testing with the model file obtained after training.

If you want to test your trained model, you need to follow these steps:

• 1、 Convert the trained model into an inference model. The example command is:
  ‘python tools/export_model.py -c "./configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml" -o Global.pretrained_model="./output/v3_en_rec/best_accuracy.pdparams" Global.save_inference_dir="./inference_model/rec/" ’
  You need to adjust the paths according to your situation. It will automatically generate model files such as model.pdmodel or inference.pdmodel in the specified directory.

• 2、Use your previous test script command:
  ‘python tools/infer/predict_rec.py’
  Make sure to set rec_model_dir to point to the inference model generated in Step One.

By following these two steps, you will resolve the issue you encountered.

[kishore-work-hard]

• python tools/export_model.py -c "./configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml" -o Global.pretrained_model="./output/v3_en_rec/best_accuracy.pdparams" Global.save_inference_dir="./inference_model/rec/"

thanks for the English answer. I will try it out.

[pcycccccc]

@kishore-work-hard
If you want to test the text recognition effect of your trained model after completing the training, you only need to use the model file obtained from your training. This does not involve the inference model files provided by the official source, such as ch_PP-OCRv4_rec_infer.tar or other inference model files.

I guess the script automatically downloads URL files because it cannot find your model.pdmodel or inference.pdmodel files during inference testing, so it downloads the inference model from the default URL.

[kishore-work-hard]

@kishore-work-hard If you want to test the text recognition effect of your trained model after completing the training, you only need to use the model file obtained from your training. This does not involve the inference model files provided by the official source, such as ch_PP-OCRv4_rec_infer.tar or other inference model files.

I guess the script automatically downloads URL files because it cannot find your model.pdmodel or inference.pdmodel files during inference testing, so it downloads the inference model from the default URL.

I will followed ur instructions and create inference.pdmodel. using tools/export_model.py.
It worked.
results are still bad.
how many images are needed to train an rec model for good accuracy. currently i used 10k cropped number plate images without using data argumentation. any insight would be helpfull

[pcycccccc]

@kishore-work-hard Did you use crop_image for both training and testing? Additionally, does your --rec_char_dict_path point to your dictionary file?

[kishore-work-hard]

@kishore-work-hard Did you use crop_image for both training and testing? Additionally, does your --rec_char_dict_path point to your dictionary file?

dictionary path are all correct. i am getting one or two letters recognised out of 10 letters.
both my training and testing data contains custom dataset from real life cropped number plates from the images of cars and bikes taken from ip cams.

i didnt train paddle's text det model. i disabled it and enabled only rec model when doing inference
i only trained paddle's text rec model
(set rec=True, Det=False)

sample camera image(without cropping) -

sample dataset images (after cropping the number plate using yolo)-

These Cropped Number Plates are the datasets used as training and testing dataset.