fastnlp / CPT

CPT: A Pre-Trained Unbalanced Transformer for Both Chinese Language Understanding and Generation
481 stars 70 forks source link

Pytorch model to ONNX, embedding layer parameters are misalignment? #76

Open smygreat opened 10 months ago

smygreat commented 10 months ago

使用 fast-Bart 把模型转成onnx格式,encoder的embedding层的参数就开始出现差异, image

考虑到bart和cpt encoder部分本身的差异,因为cpt的encoder本身就是bert,所以使用bert转onnx的方式将encoder转化,同样存在上述的问题,各位有遇到过类似问题吗?