zyg
zyg
这个应该不只是数据的问题,模型也有关系。我是先用垂直领域的语料训练albert模型,然后再做下游分类任务。同样的下游任务,如果预训练过程过拟合越严重,下游分类任务出现nan的概率就越高
我也发现这个问题了,没人回复啊
i have the same problem. create file directory for output file first.
I:BERT_VEC:[graph:opt: 42]:build graph... I:BERT_VEC:[graph:opt: 95]:load parameters from checkpoint... I:BERT_VEC:[graph:opt: 97]:freeze... INFO:tensorflow:Froze 181 variables. INFO:tensorflow:Converted 181 variables to const ops. I:BERT_VEC:[graph:opt:100]:optimize... I:BERT_VEC:[graph:opt:108]:write graph to a tmp file: D:\opensource\bert-utils-master\tmp\result\tmpr8i4x_4o WARNING:tensorflow:Using temporary folder...