Skip to content

loss_scale 疑问 #3711

@Hello-Worldd

Description

@Hello-Worldd

请问为什么这里要把loss_scale键删除呢

encoded[k] = None

encoded[k] = None

比如我定义loss_scale函数实现token权重["学习", "好", "数学", "是", "重要", "的"][1.0, 0.5, 2.0, 0.5, 2.0, 0.1]
把loss_scale键删除这样后续计算loss不就会不按照token权重计算损失,而计算整句话的平均损失吗

if loss_scale is not None:

if loss_scale is not None:
encoded[k] = None

这里是指只有{'default', 'all', 'last_round'}这三种会删除键,而自己定义的loss_scale不需要执行这步吗

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions