NEZHA: Neural Contextualized Representation for Chinese Language Understanding
262
stars
54
forks
source link
两个self.relative_positions_encoding[:to_seq_length, :to_seq_length, :].to(hidden_states.device)太影响性能了 #10
Open
huangyc0618 opened 3 years ago
占用了大量CPU资源和时间，建议init初始化后就直接to device