asd9697
commented
1 month ago hello, this strategy is tested by ablation study. You can also fix the activation function to mish. But it will leads to more pre-training time and a little performance degradation. But of course, I believe you can find a better way to train the model with the single activation function. We just not take more time to test it due to it is not the focus of our method. … -----原始邮件----- 发件人:asd131 @.> 发送时间:2024-07-30 10:04:55 (星期二) 收件人: gallenszl/CFNet @.> 抄送: Subscribed @.> 主题: [gallenszl/CFNet] 作者您好，谢谢你做的工作，我想请问一下为什么只在第二轮训练时使用mish激活函数，第一轮为什么还是relu呢？ (Issue #40) — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.>
谢谢您的解答，我尝试了mish和relu，第二轮使用mish的确会使sceneflow的epe降低，但将生成的视差图上传到kitti，得到的指标没有使用relu生成的好，请问您得出过这个结论吗
gallenszl
hello, this strategy is tested by ablation study. You can also fix the activation function to mish. But it will leads to more pre-training time and a little performance degradation. But of course, I believe you can find a better way to train the model with the single activation function. We just not take more time to test it due to it is not the focus of our method.
-----原始邮件----- 发件人:asd131 @.> 发送时间:2024-07-30 10:04:55 (星期二) 收件人: gallenszl/CFNet @.> 抄送: Subscribed @.***> 主题: [gallenszl/CFNet] 作者您好，谢谢你做的工作，我想请问一下为什么只在第二轮训练时使用mish激活函数，第一轮为什么还是relu呢？ (Issue #40)
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>