rec识别模型微调后的遗忘现象

[tph-CN]

系统：WIN10 22H2专业版，
python：3.13.5
CUDA:12.9
paddlepaddle版本：3.1.0
仓库版本：3.0.0
模型：ppocr-v5
显卡:英伟达RTX3090 24G显存 单卡
在官方预置模型上微调，第一次微调后能正确识别的字符，增加新数据后进行第二次训练，增加的第二次数据是新数据，与第一次不一样，第一次训练后能正确识别的数据，到第二次训练后，第一次训练的字符不能正确识别了，该调整哪些参数？配置文件如下，每次加入新数据都是重头在官方预置模型上开始训练的，并非在上一次自我训练的模型基础上进行的增量训练。图1为第一次训练的识别结果，图2为第二次训练的结果.图1是正确的结果，图2识别错误
Global:
model_name: PP-OCRv5_mobile_rec # To use static model for inference.
debug: false
use_gpu: true
epoch_num: 60
log_smooth_window: 20
print_batch_step: 1
save_model_dir: G:\PaddleOCR-3.3.0\output\PP-OCRv5_mobile\PP-OCRv5_mobile_rec
save_epoch_step: 1
eval_batch_step: [0, 323]
cal_metric_during_train: true
pretrained_model: G:\PaddleOCR-3.3.0\pretrained_model\PP-OCRv5_mobile\PP-OCRv5_mobile_rec_pretrained.pdparams
checkpoints:
save_inference_dir: null
use_visualdl: false
infer_img: doc/imgs_words/ch/word_1.jpg
character_dict_path: ./ppocr/utils/dict/ppocrv5_dict.txt
max_text_length: &max_text_length 60
infer_mode: false
use_space_char: true
distributed: true
save_res_path: ./output/rec/predicts_ppocrv5.txt
d2s_train_image_shape: [3, 48, 320]

Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.00001
warmup_epoch: 5
regularizer:
name: L2
factor: 3.0e-05

Architecture:
model_type: rec
algorithm: SVTR_LCNet
Transform:
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: 0.00001
- NRTRHead:
nrtr_dim: 384
max_text_length: *max_text_length

Loss:
name: MultiLoss
loss_config_list:
- CTCLoss:
- NRTRLoss:

PostProcess:
name: CTCLabelDecode

Metric:
name: RecMetric
main_indicator: acc

Train:
dataset:
name: MultiScaleDataSet
ds_width: false
data_dir: G:\PaddleOCR-3.3.0\train_data\rec
ext_op_transform_idx: 1
label_file_list:
- G:\PaddleOCR-3.3.0\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: *max_text_length
- RecAug:
- 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: &bs 96
fix_bs: false
divided_factor: [8, 16] # w, h
is_training: True
loader:
shuffle: true
batch_size_per_card: *bs
drop_last: true
num_workers: 8
Eval:
dataset:
name: SimpleDataSet
data_dir: G:\PaddleOCR-3.3.0\train_data\rec
label_file_list:
- G:\PaddleOCR-3.3.0\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: 96
num_workers: 4

图1.bmp
图2.bmp

[chenxz11]

请问你解决了吗？

[tph-CN]

没有