1. 写在前面

模型（推理模型）分为预测模型和识别模型，模型拿到数据需要先使用预测模型将文本框选出来，然后再使用识别模型来识别文本

训练模型总体分为4个步骤：第一步：标注数据，并生成相关配置文件第二步：训练预测模型第三步：训练识别模型第四步：将两个模型导出为推理模型

首先标注数据使用PPOCRLabel，将准备好的图片打开，可以使用自动标注将图片中需要识别的文本标注出来，然后人工修改模型预测错的数据，也可以直接自己框选并写入数据

PPOCRLabel使用教程

可以从训练曲线中参考到训练模型过程中模型在识别时学习的过程

1. 准确率 (Accuracy)

在训练一百轮中，整体表现稳定提升，说明模型学习效果不断增强。

2. 总损失（Loss）

Loss：模型在一次前向计算后，所有任务（子损失）误差的加权和。它反映了模型整体预测结果与真实标签之间的差距。

$Total Loss = CTCLoss + NRTRLoss$

第 2 轮约 93，迅速降到第 10 轮 ~17，然后平滑下降到第 50 轮 ~1.7，再到第 100 轮 ~0.85。训练初期收敛最快，中后期趋于平缓。

3. CTCLoss

CTCLoss：它惩罚模型把字符预测到错误的时刻或多余的空白，数值越小说明模型在序列对齐和字符识别上越准确。

从 ~90 降到 ~0.23，第 10–20 轮后就降至个位数，第 40 轮后接近 1，至第 100 轮降至 ~0.23。

是整体 loss 降低的主力。

4. 归一化编辑距离 (Norm Edit Distance)

编辑距离：衡量预测文本和真实文本在字符层面需要多少次“增删改”操作才能互相转换。

归一化：用编辑距离除以真实文本长度，得到一个 0–1 之间的比值。如果值越接近0，表示预测文本与真实文本差异越小，预测的越准确。

第 30+ 轮后即超过 0.90，到第 60 轮接近 0.98，表明预测文本与真实文本高度一致。

5. 模型收敛

收敛是训练进入“平稳”阶段，模型已基本学到数据规律，还未出现明显分歧。

在训练开始时，损失（Loss）通常会快速下降，准确率迅速上升；当训练到一定轮数后，指标变化变得非常缓慢，曲线趋于平稳，这种状态就叫“收敛”

即模型已经基本学到了数据中的规律，再继续训练改进也会很有限。

训练模型到”收敛“的时候表明训练达到了最佳或接近最佳状态，可以考虑停止训练、保存模型，或者改用更细的学习率再做微调。

如果epoch_num（迭代次数）过高，会导致模型过拟合

6. 过拟合

过拟合 是训练和泛化能力开始脱钩，虽然训练集表现更好，但验证／测试集性能变差，需加以干预。

过拟合示例：

训练 loss 从 0.5 → 0.2 → 0.1 → 0.05；验证 loss 从 0.6 → 0.3 → 0.25 → 0.35。
验证 loss 在第三次迭代后回升，说明模型开始记忆训练集噪声，泛化能力下降。

过拟合会导致什么？

正常学习：你看了很多苹果、香蕉、橙子的照片，还是真实的水果。考试时，给你没见过角度的苹果、香蕉、橙子，你都能认出来——这就是「泛化好」，模型没过拟合。

过拟合：你只记住了教科书里那几张苹果、香蕉、橙子的照片的每一个细节（比如苹果旁边那片叶子的形状、香蕉背景的光线角度），连截图里边的阴影都背下来了。但考试时给你一张真实市场上卖的苹果照片，因为背景不同、光线不同，你反而认不出这是苹果了。——模型把训练集（教科书图片）记得滚瓜烂熟，却学不到“苹果是圆圆的、有柄、红色或青色”这个本质，就叫过拟合。

过拟合就像死记硬背教科书上的样本，而不是理解背后的规律，结果一遇到新场景就傻眼。

判定依据：

Loss 曲线趋于平缓：每个 epoch 的 loss 下降很少。
Accuracy / norm_edit_dis 曲线趋于平稳：指标几乎不再提升。

2. 训练步骤

2.1 标注数据

ppocrlabel安装位置

D:\\anaconda3\\envs\\paddle38\\lib\\site-packages

安装路径

C:\\Users\\wa1yb\\.conda\\envs\\paddle39\\Lib\\site-packages\\PPOCRLabel>

启动命令

python PPOCRLabel.py --lang ch

将数据集划分(训练集/验证集/测试集)

**python gen_ocr_train_val_test.py --trainValTestRatio 6:2:2 –-datasetRootPath ./train_data --detRootPath ./train_data/det --recRootPath ./train_data/rec**

参数说明：

trainValTestRatio 是训练集、验证集、测试集的图像数量划分比例，根据实际情况设定，默认是6:2:2
datasetRootPath 是PPOCRLabel标注的完整数据集存放路径。默认路径是 …/train_data/

就是步骤6的结果文件夹train_data文件夹。
detRootPath 是输出训练文字检测的数据集存放路径。默认路径是 …/train_data/det
recRootPath 是输出训练文字识别的数据集存放路径。默认路径是 …/train_data/rec

2.2 训练检测模型

2.2.1 修改配置文件

配置文件

Global:
  debug: false
  use_gpu: true
  epoch_num: 300 # 迭代次数
  log_smooth_window: 20
  print_batch_step: 10
  save_model_dir: ./output/en_PP-OCRv3_det/
  pretrained_model: ./Temp/myModel/en_PP-OCRv3_det/latest
  save_epoch_step: 5
  eval_batch_step:
  - 0
  - 100  # 验证频率 只在最后评估 让训练连续进行更久
  cal_metric_during_train: false
  checkpoints: null
  save_inference_dir: null
  infer_img: ./Temp/testImage/
  save_res_path: ./checkpoints/det_db/predicts_db.txt
  distributed: true
  d2s_train_image_shape: [3, -1, -1]
  amp_dtype: bfloat16
#  amp_dtype: float16

Architecture:
  name: DistillationModel
  algorithm: Distillation
  model_type: det
  Models:
    Student:
      pretrained:
      model_type: det
      algorithm: DB
      Transform: null
      Backbone:
        name: MobileNetV3
        scale: 0.5
        model_name: large
        disable_se: true
      Neck:
        name: RSEFPN
        out_channels: 96
        shortcut: True
      Head:
        name: DBHead
        k: 50
    Student2:
      pretrained:
      model_type: det
      algorithm: DB
      Transform: null
      Backbone:
        name: MobileNetV3
        scale: 0.5
        model_name: large
        disable_se: true
      Neck:
        name: RSEFPN
        out_channels: 96
        shortcut: True
      Head:
        name: DBHead
        k: 50
    Teacher:
      freeze_params: true
      return_all_feats: false
      model_type: det
      algorithm: DB
      Backbone:
        name: ResNet_vd
        in_channels: 3
        layers: 50
      Neck:
        name: LKPAN
        out_channels: 256
      Head:
        name: DBHead
        kernel_list: [7,2,2]
        k: 50

Loss:
  name: CombinedLoss
  loss_config_list:
  - DistillationDilaDBLoss:
      weight: 1.0
      model_name_pairs:
      - ["Student", "Teacher"]
      - ["Student2", "Teacher"]
      key: maps
      balance_loss: true
      main_loss_type: DiceLoss
      alpha: 5
      beta: 10
      ohem_ratio: 3
  - DistillationDMLLoss:
      model_name_pairs:
      - ["Student", "Student2"]
      maps_name: "thrink_maps"
      weight: 1.0
      key: maps
  - DistillationDBLoss:
      weight: 1.0
      model_name_list: ["Student", "Student2"]
      balance_loss: true
      main_loss_type: DiceLoss
      alpha: 5
      beta: 10
      ohem_ratio: 3

Optimizer:
  name: Adam
  beta1: 0.9
  beta2: 0.999
  lr:
    name: Cosine
    learning_rate: 0.001
    warmup_epoch: 2
  regularizer:
    name: L2
    factor: 5.0e-05

PostProcess:
  name: DistillationDBPostProcess
  model_name: ["Student"]
  key: head_out
  thresh: 0.3
  box_thresh: 0.6
  max_candidates: 1000
  unclip_ratio: 1.5

Metric:
  name: DistillationMetric
  base_metric_name: DetMetric
  main_indicator: hmean
  key: "Student"

Train:
  dataset:
    name: SimpleDataSet
    data_dir: ./Temp/train_data/images
    label_file_list:
      - ./Temp/train_data/det/train.txt
    ratio_list: [1.0]
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - DetLabelEncode: null
    - CopyPaste:
    - IaaAugment:
        augmenter_args:
        - type: Fliplr
          args:
            p: 0.5
        - type: Affine
          args:
            rotate:
            - -10
            - 10
        - type: Resize
          args:
            size:
            - 0.5
            - 3
    - EastRandomCropData:
        size:
        - 960
        - 960
        max_tries: 50
        keep_ratio: true
    - MakeBorderMap:
        shrink_ratio: 0.4
        thresh_min: 0.3
        thresh_max: 0.7
    - MakeShrinkMap:
        shrink_ratio: 0.4
        min_text_size: 8
    - NormalizeImage:
        scale: 1./255.
        mean:
        - 0.485
        - 0.456
        - 0.406
        std:
        - 0.229
        - 0.224
        - 0.225
        order: hwc
    - ToCHWImage: null
    - KeepKeys:
        keep_keys:
        - image
        - threshold_map
        - threshold_mask
        - shrink_map
        - shrink_mask
  loader:
    shuffle: true
    drop_last: false
    batch_size_per_card: 8
    num_workers: 4

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./Temp/train_data/images
    label_file_list:
      - ./Temp/train_data/det/val.txt
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - DetLabelEncode: # Class handling label
      - DetResizeForTest:
      - NormalizeImage:
          scale: 1./255.
          mean: [0.485, 0.456, 0.406]
          std: [0.229, 0.224, 0.225]
          order: 'hwc'
      - ToCHWImage:
      - KeepKeys:
          keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: 1 # must be 1
    num_workers: 2

2.2.2 启动训练

在PaddleOCR-release-2.8.1根目录下，执行命令开始训练：

python .\\tools\\train.py -c .\\Temp\\configs\\en_PP-OCRv3_det_cml.yml

2.2.3 测试模型

测试训练好的检测模型检测图片是否准确

python tools/infer_det.py -c .\\Temp\\configs\\en_PP-OCRv3_det_cml.yml -o Global.pretrained_model=/output/en_PP-OCR_v3_det/latest.pdparams Global.infer_img="D:\\Files\\PythonProject\\Important final code\\PaddleOCR-3.0.2\\testImages\\001.png"

测试导出后的检测模型 检测图片是否准确

python .\\tools\\infer\\predict_det.py --det_model_dir="./inference_model/v3det_model/Student" --image_dir="./Temp/testImages/003.png" --use_gpu=true

2.3 训练识别模型

配置文件

Global:
  debug: false
  use_gpu: true
  epoch_num: 300
  save_model_dir: ./output/en_PP-OCRv4_rec
  pretrained_model: ./Temp/en_PP-OCRv4_rec_train/best_accuracy
#  checkpoints: ./output/en_PP-OCRv4_rec/latest
  checkpoints: null
  log_smooth_window: 20
  print_batch_step: 10
  save_epoch_step: 10
  eval_batch_step:
  - 0
  - 9999999
  cal_metric_during_train: true
  save_inference_dir: null
  use_visualdl: false
  infer_img: ./Temp/testImage/
  character_dict_path: ppocr/utils/en_dict.txt
  max_text_length: 25
  infer_mode: false
  use_space_char: true
  distributed: true
  save_res_path: ./output/rec/predicts_ppocrv3.txt
Optimizer:
  name: Adam
  beta1: 0.9
  beta2: 0.999
  lr:
    name: Cosine
    learning_rate: 0.0005
    warmup_epoch: 5
  regularizer:
    name: L2
    factor: 3.0e-05
Architecture:
  model_type: rec
  algorithm: SVTR_LCNet
  Transform: null
  Backbone:
    name: PPLCNetV3
    scale: 0.95
  Head:
    name: MultiHead
    head_list:
    - CTCHead:
        Neck:
          name: svtr
          dims: 120
          depth: 2
          hidden_dims: 120
          kernel_size:
          - 1
          - 3
          use_guide: true
        Head:
          fc_decay: 1.0e-05
    - NRTRHead:
        nrtr_dim: 384
        max_text_length: 25
Loss:
  name: MultiLoss
  loss_config_list:
  - CTCLoss: null
  - NRTRLoss: null
PostProcess:
  name: CTCLabelDecode
Metric:
  name: RecMetric
  main_indicator: acc
  ignore_space: false
Train:
  dataset:
    name: MultiScaleDataSet
    ds_width: false
    data_dir: ./Temp/train_data/images
    ext_op_transform_idx: 1
    label_file_list:
    - ./Temp/train_data/rec/train.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - RecConAug:
        prob: 0.5
        ext_data_num: 2
        image_shape:
        - 48
        - 320
        - 3
        max_text_length: 25
    - RecAug: null
    - MultiLabelEncode:
        gtc_encode: NRTRLabelEncode
    - KeepKeys:
        keep_keys:
        - image
        - label_ctc
        - label_gtc
        - length
        - valid_ratio
  sampler:
    name: MultiScaleSampler
    scales:
    - - 320
      - 32
    - - 320
      - 48
    - - 320
      - 64
    first_bs: 96
    fix_bs: false
    divided_factor:
    - 8
    - 16
    is_training: true
  loader:
    shuffle: true
    batch_size_per_card: 16
    drop_last: true
    num_workers: 8
Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./Temp/train_data/images
    label_file_list:
    - ./Temp/train_data/rec/val.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - MultiLabelEncode:
        gtc_encode: NRTRLabelEncode
    - RecResizeImg:
        image_shape:
        - 3
        - 48
        - 320
    - KeepKeys:
        keep_keys:
        - image
        - label_ctc
        - label_gtc
        - length
        - valid_ratio
  loader:
    shuffle: false
    drop_last: false
    batch_size_per_card: 16
    num_workers: 4
profiler_options: null

启动训练

python tools/train.py -c .\\Temp\\configs\\en_PP-OCRv4_rec.yml

测试训练好的识别模型检测图片是否准确

python tools/infer_rec.py -c .\\Temp\\configs\\en_PP-OCRv4_rec.yml -o Global.pretrained_model=.\\output/en_PP-OCRv4_rec/best_accuracy.pdparams Global.infer_img="D:\\Files\\PaddleOCR-2.8.1\\Temp\\testImages\\577.png"

评估性能

在训练过程中可以及时对已经训练好的模型进行评估模型效果

在验证集上评估已经训练好的模型的性能表现，不继续训练，仅计算准确率、编辑距离和损失值等指标。

python .\\tools\\eval.py -c .\\Temp\\configs\\en_PP-OCRv4_rec.yml -o Global.checkpoints=.\\output\\en_PP-OCRv4_rec\\latest

2.4 导出推理模型

分别将上面训练好的检测、识别模型导出为推理模型

导出检测模型

python .\\tools\\export_model.py -c .\\Temp\\configs\\en_PP-OCRv3_det_cml.yml -o Global.pretrained_model=.\\Temp\\output\\en_PP-OCRv3_det\\latest Global.save_inference_dir=./inference_model/det_model/

导出识别模型

python .\\tools\\export_model.py -c .\\Temp\\configs\\en_PP-OCRv4_rec.yml -o Global.pretrained_model=.\\output/en_PP-OCRv4_rec/best_accuracy Global.save_inference_dir=./inference_model/rec_model/

3. 测试模型

import os
from paddleocr import PaddleOCR
from PIL import Image
from paddleocr.tools.infer.utility import draw_ocr

# 初始化OCR模型
ocr = PaddleOCR(
    lang="ch",
    use_gpu=False,
    det_model_dir="inference_model/v3det_model/Student",
    rec_model_dir="inference_model/rec_model",
    rec_char_dict_path='inference_model/en_dict.txt',
    # det_db_thresh=0.1,   # 文字检测阈值 默认0.3
    # det_db_box_thresh=0.1,  # 框后处理阈值
    # det_db_unclip_ratio=3.0,  # 扩展框比例，值太小可能导致框过小或断裂
)

# 输入文件夹路径
img_folder = 'testImages'
output_folder = 'results_with_boxes'
os.makedirs(output_folder, exist_ok=True)

img_extensions = ['.png', '.jpg', '.jpeg']

for filename in os.listdir(img_folder):
    if any(filename.lower().endswith(ext) for ext in img_extensions):
        img_path = os.path.join(img_folder, filename)
        print(f"Processing: {img_path}")

        img = Image.open(img_path).convert('RGB')
        result = ocr.ocr(img_path)

        # 输出识别结果和置信度
        for line in result[0]:
            text = line[1][0]
            score = line[1][1]
            print(f"{filename}\\t{text}\\t{score:.4f}")

        # 画图：检测框坐标和文字内容
        boxes = [line[0] for line in result[0]]  # 检测框坐标
        texts = [line[1][0] for line in result[0]]  # 识别文本
        scores = [line[1][1] for line in result[0]]  # 置信度

        # PaddleOCR自带的绘图工具画图
        img_with_boxes = draw_ocr(img, boxes, texts, scores)

        # 保存图片
        img_with_boxes = Image.fromarray(img_with_boxes)
        save_path = os.path.join(output_folder, f"{os.path.splitext(filename)[0]}_with_boxes.png")
        img_with_boxes.save(save_path)
        print(f"Saved result to: {save_path}\\n")

4. 虚拟环境

conda activate paddle38

５.模型使用

时间：2025/10/10

环境

paddleocr             2.8.1
paddlepaddle-gpu      2.6.2
paddlex               3.2.1

代码test Model.py

import os
import sys
import argparse
import json
from paddleocr import PaddleOCR

def main():
    parser = argparse.ArgumentParser(description="Single image OCR")
    parser.add_argument("image_path", help="输入图片路径")
    args = parser.parse_args()

    image_path = args.image_path
    if not os.path.isfile(image_path):
        print(json.dumps({
            "success": False,
            "error": f"文件不存在: {image_path}"
        }, ensure_ascii=False))
        sys.exit(1)

    try:
        # PaddleOCR 2.8.1 调用训练模型
        ocr = PaddleOCR(
            use_gpu=True,
            det_model_dir=r"D:\\Files\\PythonProject2\\src\\main\\resources\\models\\det_model_v3\\Student",
            rec_model_dir=r"D:\\Files\\PythonProject2\\src\\main\\resources\\models\\rec_model_v4",
            rec_char_dict_path=r"D:\\Files\\PythonProject2\\src\\main\\resources\\models\\en_dict.txt"
        )

        result = ocr.ocr(image_path, cls=True)

        lines = []
        full_text = ""
        for line in result[0]:
            box = line[0]
            text = line[1][0]
            score = line[1][1]
            full_text += text
            lines.append({
                "box": box,
                "text": text,
                "score": score
            })

        print(json.dumps({
            "success": True,
            "full_text": full_text.strip(),
            "lines": lines
        }, ensure_ascii=False))

    except Exception as e:
        print(json.dumps({
            "success": False,
            "error": f"OCR处理失败: {str(e)}"
        }, ensure_ascii=False))
        sys.exit(2)

if __name__ == "__main__":
    main()

执行

PS D:\\Files\\PythonProject2> python .\\testModel.py .\\img.png
[2025/10/10 13:37:41] ppocr DEBUG: Namespace(help='==SUPPRESS==', use_gpu=True, use_xpu=False, use_npu=False, use_mlu=False, ir_optim=True, use_tensorrt=False, min_subgraph_size=15, precision='fp32', gpu_mem=500, gpu_id=0, image_dir=None, page_num=0, det_algorithm='DB', det_model_dir='D:\\\\Files\\\\PythonProject2\\\\src\\\\main\\\\resources\\\\models\\\\det_model_v3\\\\Student', det_limit_side_len=960, det_limit_type='max', det_box_type='quad', det_db_thresh=0.3, det_db_box_thresh=0.6, det_db_unclip_ratio=1.5, max_batch_size=10, use_dilation=False, det_db_score_mode='fast', det_east_score_thresh=0.8, det_east_cover_thresh=0.1, det_east_nms_thresh=0.2, det_sast_score_thresh=0.5, det_sast_nms_thresh=0.2, det_pse_thresh=0, det_pse_box_thresh=0.85, det_pse_min_area=16, det_pse_scale=1, scales=[8, 16, 32], alpha=1.0, beta=1.0, fourier_degree=5, rec_algorithm='SVTR_LCNet', rec_model_dir='D:\\\\Files\\\\PythonProject2\\\\src\\\\main\\\\resources\\\\models\\\\rec_model_v4', rec_image_inverse=True, rec_image_shape='3, 48, 320', rec_batch_num=6, max_text_length=25, rec_char_dict_path='D:\\\\Files\\\\PythonProject2\\\\src\\\\main\\\\resources\\\\models\\\\en_dict.txt', use_space_char=True, vis_font_path='./doc/fonts/simfang.ttf', drop_score=0.5, e2e_algorithm='PGNet', e2e_model_dir=None, e2e_limit_side_len=768, e2e_limit_type='max', e2e_pgnet_score_thresh=0.5, e2e_char_dict_path='./ppocr/utils/ic15_dict.txt', e2e_pgnet_valid_set='totaltext', e2e_pgnet_mode='fast', use_angle_cls=False, cls_model_dir='C:\\\\Users\\\\33226/.paddleocr/whl\\\\cls\\\\ch_ppocr_mobile_v2.0_cls_infer', cls_image_shape='3, 48, 192', label_list=['0', '180'], cls_batch_num=6, cls_thresh=0.9, enable_mkldnn=False, cpu_threads=10, use_pdserving=False, warmup=False, sr_model_dir=None, sr_image_shape='3, 32, 128', sr_batch_num=1, draw_img_save_dir='./inference_results', save_crop_res=False, crop_res_save_dir='./output', use_mp=False, total_process_num=1, process_id=0, benchmark=False, save_log_path='./log_output/', show_log=True, use_onnx=False, return_word_box=False, output='./output', table_max_len=488, table_algorithm='TableAttn', table_model_dir=None, merge_no_span_structure=True, table_char_dict_path=None, layout_model_dir=None, layout_dict_path=None, layout_score_threshold=0.5, layout_nms_threshold=0.5, kie_algorithm='LayoutXLM', ser_model_dir=None, re_model_dir=None, use_visual_backbone=True, ser_dict_path='../train_data/XFUND/class_list_xfun.txt', ocr_order_method=None, mode='structure', image_orientation=False, layout=True, table=True, ocr=True, recovery=False, use_pdf2docx_api=False, invert=False, binarize=False, alphacolor=(255, 255, 255), lang='ch', det=True, rec=True, type='ocr', savefile=False, ocr_version='PP-OCRv4', structure_version='PP-StructureV2')
[2025/10/10 13:37:45] ppocr WARNING: Since the angle classifier is not initialized, it will not be used during the forward process
[2025/10/10 13:37:46] ppocr DEBUG: dt_boxes num : 5, elapsed : 1.3364863395690918
[2025/10/10 13:37:46] ppocr DEBUG: rec_res num  : 5, elapsed : 0.13319826126098633
{"success": true, "full_text": "0317WS2022D100012", "lines": [{"box": [[16.0, 4.0], [122.0, 4.0], [122.0, 83.0], [16.0, 83.0]], "text": "0317", "score": 0.8762916922569275}, {"box": [[148.0, 11.0], [248.0, 11.0], [248.0, 82.0], [148.0, 82.0]], "text": "WS", "score": 0.9998351335525513}, {"box": [[252.0, 17.0], [379.0, 12.0], [382.0, 81.0], [256.0, 86.0]], "text": "2022", "score": 0.9991670250892639}, {"box": [[432.0, 18.0], [524.0, 23.0], [521.0, 83.0], [429.0, 79.0]], "text": "D10", "score": 0.9650375247001648}, {"box": [[16.0, 90.0], [114.0, 90.0], [114.0, 151.0], [16.0, 151.0]], "text": "0012", "score": 0.9996055960655212}]}

6. 检测模型参数量

import paddle
import os

def count_inference_model_params(model_prefix: str, use_gpu=False):
    """
    统计 Paddle 推理模型 (.pdmodel + .pdiparams) 参数总量
    model_prefix: 文件前缀，例如 D:\\...\\rec_model_v4\\inference
    """
    if not os.path.exists(model_prefix + ".pdmodel") or not os.path.exists(model_prefix + ".pdiparams"):
        raise FileNotFoundError(f"找不到模型文件: {model_prefix}.pdmodel 或 {model_prefix}.pdiparams")

    paddle.enable_static()
    place = paddle.CUDAPlace(0) if use_gpu else paddle.CPUPlace()
    exe = paddle.static.Executor(place)

    inference_program, feed_target_names, fetch_targets = paddle.static.load_inference_model(
        path_prefix=model_prefix, executor=exe
    )

    total_params = 0
    for block in inference_program.blocks:
        for var in block.vars.values():
            if var.persistable and var.name not in feed_target_names:
                try:
                    shape = var.shape
                except RuntimeError:
                    continue
                param_count = 1
                for dim in shape:
                    param_count *= dim
                total_params += param_count
    return total_params

if __name__ == "__main__":
    det_model_prefix = r"D:\\Files\\PythonProject2\\src\\main\\resources\\models\\det_model_v3\\Student\\inference"
    rec_model_prefix = r"D:\\Files\\PythonProject2\\src\\main\\resources\\models\\rec_model_v4\\inference"

    det_params = count_inference_model_params(det_model_prefix, use_gpu=False)
    rec_params = count_inference_model_params(rec_model_prefix, use_gpu=False)

    print(f"检测模型参数量: {det_params}")
    print(f"识别模型参数量: {rec_params}")

输出

PS D:\\Files\\PythonProject2> python .\\testModelsCanShu.py
I1010 13:56:04.597359  3416 program_interpreter.cc:212] New Executor is Running.
检测模型参数量: 592625
识别模型参数量: 1890532

临时记录

start_epoch:188 验证集评估结果

[2025/06/25 16:15:36] ppocr INFO: resume from .\\output\\en_PP-OCRv4_rec\\latest
[2025/06/25 16:15:36] ppocr INFO: metric in ckpt ***************
[2025/06/25 16:15:36] ppocr INFO: acc:0
[2025/06/25 16:15:36] ppocr INFO: is_float16:False
[2025/06/25 16:15:36] ppocr INFO: start_epoch:188
[2025/06/25 16:15:39] ppocr INFO: metric eval ***************
[2025/06/25 16:15:39] ppocr INFO: acc:0.9583332834201415
[2025/06/25 16:15:39] ppocr INFO: norm_edit_dis:0.9878472228551793
[2025/06/25 16:15:39] ppocr INFO: fps:63.97288645081185

[2025/06/26 09:05:29] ppocr INFO: resume from .\\output\\en_PP-OCRv4_rec\\latest
[2025/06/26 09:05:29] ppocr INFO: metric in ckpt ***************
[2025/06/26 09:05:29] ppocr INFO: acc:0
[2025/06/26 09:05:29] ppocr INFO: is_float16:False
[2025/06/26 09:05:29] ppocr INFO: start_epoch:301
[2025/06/26 09:05:30] ppocr INFO: metric eval ***************
[2025/06/26 09:05:30] ppocr INFO: acc:0.9635416164822075
[2025/06/26 09:05:30] ppocr INFO: norm_edit_dis:0.9891493061206958
[2025/06/26 09:05:30] ppocr INFO: fps:265.29454544496195

[2025/06/25 18:24:53] ppocr INFO: epoch: [270/300], global_step: 300, lr: 0.000058, acc: 1.000000, norm_edit_dis: 1.0
00000, CTCLoss: 0.049039, NRTRLoss: 0.620189, loss: 0.669288, avg_reader_cost: 0.06627 s, avg_batch_cost: 8.51739 s, avg_samples: 70.4, ips: 8.26544 samples/s, eta: 0:39:29, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:25:28] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:25:28] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\iter_epoch_270
[2025/06/25 18:26:19] ppocr INFO: epoch: [271/300], global_step: 310, lr: 0.000057, acc: 1.000000, norm_edit_dis: 1.0
00000, CTCLoss: 0.048249, NRTRLoss: 0.620180, loss: 0.668467, avg_reader_cost: 0.11603 s, avg_batch_cost: 8.57698 s, avg_samples: 72.0, ips: 8.39456 samples/s, eta: 0:37:53, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:26:36] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:27:45] ppocr INFO: epoch: [272/300], global_step: 320, lr: 0.000055, acc: 0.989583, norm_edit_dis: 0.9
99132, CTCLoss: 0.059071, NRTRLoss: 0.620167, loss: 0.679213, avg_reader_cost: 0.07585 s, avg_batch_cost: 8.61765 s, avg_samples: 68.8, ips: 7.98361 samples/s, eta: 0:36:19, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:27:45] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:28:54] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:29:11] ppocr INFO: epoch: [274/300], global_step: 330, lr: 0.000053, acc: 0.994792, norm_edit_dis: 0.9
99566, CTCLoss: 0.057062, NRTRLoss: 0.620193, loss: 0.677321, avg_reader_cost: 0.13713 s, avg_batch_cost: 8.65757 s, avg_samples: 70.4, ips: 8.13162 samples/s, eta: 0:34:36, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:30:03] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:30:37] ppocr INFO: epoch: [275/300], global_step: 340, lr: 0.000051, acc: 1.000000, norm_edit_dis: 1.0
00000, CTCLoss: 0.058556, NRTRLoss: 0.620187, loss: 0.678740, avg_reader_cost: 0.07255 s, avg_batch_cost: 8.54815 s, avg_samples: 72.0, ips: 8.42288 samples/s, eta: 0:33:00, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:31:12] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:32:02] ppocr INFO: epoch: [276/300], global_step: 350, lr: 0.000049, acc: 0.994792, norm_edit_dis: 0.9
99566, CTCLoss: 0.062293, NRTRLoss: 0.620185, loss: 0.682432, avg_reader_cost: 0.06702 s, avg_batch_cost: 8.54973 s, avg_samples: 73.6, ips: 8.60846 samples/s, eta: 0:31:25, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:32:20] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:33:28] ppocr INFO: epoch: [277/300], global_step: 360, lr: 0.000048, acc: 0.994792, norm_edit_dis: 0.9
99566, CTCLoss: 0.053074, NRTRLoss: 0.620199, loss: 0.673219, avg_reader_cost: 0.06782 s, avg_batch_cost: 8.59742 s, avg_samples: 67.2, ips: 7.81629 samples/s, eta: 0:29:51, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:33:29] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:34:37] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:34:54] ppocr INFO: epoch: [279/300], global_step: 370, lr: 0.000046, acc: 0.994792, norm_edit_dis: 0.9
99566, CTCLoss: 0.045486, NRTRLoss: 0.620186, loss: 0.667856, avg_reader_cost: 0.12697 s, avg_batch_cost: 8.59989 s, avg_samples: 70.4, ips: 8.18615 samples/s, eta: 0:28:07, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:35:46] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:36:20] ppocr INFO: epoch: [280/300], global_step: 380, lr: 0.000044, acc: 0.989583, norm_edit_dis: 0.9
99092, CTCLoss: 0.057158, NRTRLoss: 0.620180, loss: 0.677308, avg_reader_cost: 0.07080 s, avg_batch_cost: 8.54077 s, avg_samples: 73.6, ips: 8.61749 samples/s, eta: 0:26:32, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:36:54] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:36:55] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\iter_epoch_280
[2025/06/25 18:37:46] ppocr INFO: epoch: [281/300], global_step: 390, lr: 0.000043, acc: 0.989583, norm_edit_dis: 0.9
99006, CTCLoss: 0.057158, NRTRLoss: 0.620190, loss: 0.677308, avg_reader_cost: 0.11939 s, avg_batch_cost: 8.60828 s, avg_samples: 72.0, ips: 8.36404 samples/s, eta: 0:24:57, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:38:04] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:39:11] ppocr INFO: epoch: [282/300], global_step: 400, lr: 0.000041, acc: 0.984375, norm_edit_dis: 0.9
98639, CTCLoss: 0.056775, NRTRLoss: 0.620197, loss: 0.677154, avg_reader_cost: 0.07201 s, avg_batch_cost: 8.55105 s, avg_samples: 67.2, ips: 7.85869 samples/s, eta: 0:23:22, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:39:12] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:40:21] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:40:38] ppocr INFO: epoch: [284/300], global_step: 410, lr: 0.000039, acc: 0.984375, norm_edit_dis: 0.9
98639, CTCLoss: 0.053467, NRTRLoss: 0.620176, loss: 0.673645, avg_reader_cost: 0.13797 s, avg_batch_cost: 8.66229 s, avg_samples: 73.6, ips: 8.49660 samples/s, eta: 0:21:39, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:41:30] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:42:04] ppocr INFO: epoch: [285/300], global_step: 420, lr: 0.000038, acc: 0.989583, norm_edit_dis: 0.9
98875, CTCLoss: 0.058270, NRTRLoss: 0.620152, loss: 0.678415, avg_reader_cost: 0.07106 s, avg_batch_cost: 8.63264 s, avg_samples: 72.0, ips: 8.34044 samples/s, eta: 0:20:05, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:42:39] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:43:58] ppocr INFO: epoch: [286/300], global_step: 430, lr: 0.000036, acc: 0.994792, norm_edit_dis: 0.9
99527, CTCLoss: 0.053610, NRTRLoss: 0.620160, loss: 0.676677, avg_reader_cost: 0.07294 s, avg_batch_cost: 11.40486 s, avg_samples: 70.4, ips: 6.17281 samples/s, eta: 0:18:38, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB    
[2025/06/25 18:44:18] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:45:32] ppocr INFO: epoch: [287/300], global_step: 440, lr: 0.000035, acc: 1.000000, norm_edit_dis: 1.0
00000, CTCLoss: 0.044965, NRTRLoss: 0.620161, loss: 0.665135, avg_reader_cost: 0.06975 s, avg_batch_cost: 9.39042 s, avg_samples: 68.8, ips: 7.32662 samples/s, eta: 0:17:05, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:45:33] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:46:48] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:47:07] ppocr INFO: epoch: [289/300], global_step: 450, lr: 0.000033, acc: 1.000000, norm_edit_dis: 1.0
00000, CTCLoss: 0.045696, NRTRLoss: 0.620164, loss: 0.665892, avg_reader_cost: 0.14613 s, avg_batch_cost: 9.47883 s, avg_samples: 73.6, ips: 7.76467 samples/s, eta: 0:15:23, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:48:03] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:48:40] ppocr INFO: epoch: [290/300], global_step: 460, lr: 0.000032, acc: 1.000000, norm_edit_dis: 1.0
00000, CTCLoss: 0.047539, NRTRLoss: 0.620172, loss: 0.667690, avg_reader_cost: 0.06608 s, avg_batch_cost: 9.31875 s, avg_samples: 70.4, ips: 7.55466 samples/s, eta: 0:13:48, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:49:18] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:49:19] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\iter_epoch_290
[2025/06/25 18:50:14] ppocr INFO: epoch: [291/300], global_step: 470, lr: 0.000031, acc: 1.000000, norm_edit_dis: 1.0
00000, CTCLoss: 0.049244, NRTRLoss: 0.620185, loss: 0.669500, avg_reader_cost: 0.11283 s, avg_batch_cost: 9.36815 s, avg_samples: 72.0, ips: 7.68562 samples/s, eta: 0:12:13, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB     
[2025/06/25 18:50:33] ppocr INFO: save model in ./output/en_PP-OCRv4_rec\\latest
[2025/06/25 18:51:48] ppocr INFO: epoch: [292/300], global_step: 480, lr: 0.000029, acc: 0.994792, norm_edit_dis: 0.9
99132, CTCLoss: 0.052151, NRTRLoss: 0.620176, loss: 0.672631, avg_reader_cost: 0.07306 s, avg_batch_cost: 9.35568 s, avg_samples: 68.8, ips: 7.35382 samples/s, eta: 0:10:38, max_mem_reserved: 12193 MB, max_mem_allocated: 11824 MB

安装依赖

python -m pip install scikit-image imgaug lmdb rapidfuzz -i <https://mirrors.aliyun.com/pypi/simple/>

 python -m pip install "numpy<2.0" --force-reinstall -i <https://mirrors.aliyun.com/pypi/simple/>

Wpprqi

Paddle训练模型