FlagEmbedding icon indicating copy to clipboard operation
FlagEmbedding copied to clipboard

预训练问题

Open LLLiHaotian opened this issue 1 year ago • 2 comments

如果想基于RetroMAE预训练bart、t5系列的模型,应该如何解决呢?

bart-base-chinese-cluecorpussmall-retromae_batch256_max350.log

LLLiHaotian avatar May 14 '24 03:05 LLLiHaotian

Currently, this script doesn't support encoder-decoder architecture.

staoxiao avatar May 14 '24 04:05 staoxiao

好的谢谢 还有想请问,在预训练过程中的report是这样的,在您的预训练实验中是如何判断何时停止的呢?仅凭loss曲线的变化吗? {'loss': 2.8222, 'learning_rate': 1.1313075087080098e-05, 'step': 103000, 'epoch': 1.3} 另外我注意到,在训练过程中也会偶尔出现loss值变高(但不明显,很小的变化)的情况,请问你们在预训练的过程中是否遇到过这种情况,又是如何判断何时停止的呢?

LLLiHaotian avatar May 14 '24 05:05 LLLiHaotian