bay1 / card-crnn-ctpn Goto Github PK

Test loss: 7.971412, accuray: 0.289766
Test loss: 8.877237, accuray: 0.210391
Test loss: 9.894743, accuray: 0.193203
Test loss: 7.524844, accuray: 0.291172
Test loss: 7.477661, accuray: 0.311875
Test loss: 7.481763, accuray: 0.277813
Test loss: 8.622279, accuray: 0.253047
Test loss: 8.895561, accuray: 0.241328
Test loss: 9.018443, accuray: 0.260391
Test loss: 7.462857, accuray: 0.316641
Test loss: 7.805124, accuray: 0.316484
Test loss: 8.346634, accuray: 0.337109
Test loss: 8.504839, accuray: 0.265000
Test loss: 8.384032, accuray: 0.333750
Test loss: 9.045285, accuray: 0.257344
Test loss: 8.064530, accuray: 0.269375
Test loss: 7.978924, accuray: 0.277578
Test loss: 7.978555, accuray: 0.268672
Test loss: 7.579483, accuray: 0.277187
Test loss: 8.058410, accuray: 0.291484
Test loss: 8.035892, accuray: 0.293516
Test loss: 9.988184, accuray: 0.183203
Test loss: 9.412756, accuray: 0.243906
Test loss: 9.066482, accuray: 0.278438
Test loss: 8.149712, accuray: 0.316875

训练了86轮，总共是106781张图片，Test的loss值很高，acc值很低。
但是训练的loss值很低
[88/210][462/545] Loss: 0.000032
[88/210][465/545] Loss: 0.000009
[88/210][468/545] Loss: 0.000142
[88/210][471/545] Loss: 0.000792
[88/210][474/545] Loss: 0.000005
[88/210][477/545] Loss: 0.000016
[88/210][480/545] Loss: 0.000024
[88/210][483/545] Loss: 0.000046
[88/210][486/545] Loss: 0.000013
[88/210][489/545] Loss: 0.000363

作者您好，在pytorch中CTCLoss的参数reduction默认是mean，他会对计算出来的向量先除以target lengths，然后再取平均，所以用法跟wrap_ctc并不是完全相同的，我觉得代码应该要改一下的。

你好，请问训练好的crnn model是否打算开源呢？

另想请教一个问题，“预估测试正确率达85%+”这个准确率是指单个字符的准确率还是全部银行卡号的准确率呢？
目前我用相同原始数据，并使用这里的数据增强脚本，train了一个tf版本的crnn，全部银行卡号准确率40%。

请问看明白你所follow的博主实现的yolo代码没？想请教一下