训练集acc接近1 数据样本3600
验证集 eval 的时候acc 超过0.987 数据样本600
eval model:: 91%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▍ | 10/11 [00:06<00:00, 1.59it/s]
[2021/11/10 12:12:55] root INFO: metric eval ***************
[2021/11/10 12:12:55] root INFO: acc:0.9875
但实际使用infer_rec.py对验证集数据进行推理,效果只有0.5
rec_chinese_common_v2.0模型
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
regularizer:
name: 'L2'
factor: 0.00004
Architecture:
model_type: rec
algorithm: CRNN
Transform:
Backbone:
name: ResNet
layers: 34
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 256
Head:
name: CTCHead
fc_decay: 0.00004
Loss:
name: CTCLoss
ohem_ratio: 2
PostProcess:
name: CTCLabelDecode
Metric:
name: RecMetric
main_indicator: acc
eval和infer有一些处理的diff,可以参考https://github.com/PaddlePaddle/PaddleOCR/issues/2080