Replies: 3 comments 9 replies
-
|
试试main分支，2.8版本可能有训练精度的问题。 |
Beta Was this translation helpful? Give feedback.
-
|
好，我现在验证一下。 |
Beta Was this translation helpful? Give feedback.
-
|
或者试试在2.7指定一下numpy版本 |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
|
看现象像是v3和v4导致的acc=0，loss=null |
Beta Was this translation helpful? Give feedback.
-
这个我尝试过卸载numpy并重新安装，指定numpy版本为1.26.4，实际运行下来还是会报错numpy>2.0的错误。请问有推荐的numpy版本吗 |
Beta Was this translation helpful? Give feedback.
-
我用main分支训练和预测都没有问题。导出总是各种错误。一会是none无法拼接，一会是纬度不对。按照网上更新cuda版本也是会报错🤕 |
Beta Was this translation helpful? Give feedback.
-
|
试试回退paddle版本，paddle3.0目前支持可能不太完善。 |
Beta Was this translation helpful? Give feedback.
-
|
我看第一张图，使用的是C盘里python环境的numpy，这可能是导致numpy版本报错的原因 |
Beta Was this translation helpful? Give feedback.
-
检查了一下，C盘numpy是1.26.4 ，paddle环境numpy是1.24.4，满足numpy<2.0. |
Beta Was this translation helpful? Give feedback.
-
|
那就有点奇怪了 |
Beta Was this translation helpful? Give feedback.
-
|
我也有同样的问题 |
Beta Was this translation helpful? Give feedback.
-
AttributeError:
np.sctypeswas removed in the NumPy 2.0 release. Access dtypes explicitly instead.2.7.0版本训练效果还不错，acc可以到达0.9，python端推理顺利，但是导出模型一直会报错numpy版本问题。实际安装numpy<2.0。
2.7.0可以训练，无法导出。
2.7.0使用的预训练模型：en_PP-OCRv3_rec_train
2.8.0训练和导出都可以实现，但是训练过程中acc一直为0。
2.8.0使用的预训练模型：en_PP-OCRv4_rec_train
2.7.0和2.8.0数据集一样，2.7.0训练顺利，无法导出。2.8.0训练acc始终为0，会保存best_model且可以顺利导出。
2.7.0训练速度>2.8.0训练速度。
Beta Was this translation helpful? Give feedback.
All reactions