KevinKune

Results 25 comments of KevinKune

Hi, The compressed file is partitioned into several parts due to the large size. You should uncompress these partitions together then you will find many .h5 files. For example, you...

感谢回答,最近发现调节softmax的温度会比较容易训练,在我的数据集上调节温度比attention_lambda更有效,能显著控制过拟合,建议代码中增加调节温度的参数

FM同样面临稀疏数据初始化的问题,据我所知另一种主流的FM初始化方法是采用方差为根号1/k的随机分布,k为embedding size,而且由于采用了adaptive optimizer,FM不能保证在不同的初始化条件下收敛到相同的解

SGD很慢,实验室环境中,数据量上亿之后很难做到上百轮的训练,用adaptive算法能显著加快收敛。 从比赛角度讲,比较出名的FFM就是用的adagrad算法,这可能也是libFFM的优秀训练速度的一个原因。 从工业界角度讲,end-to-end的策略也会比预训练更有优势,个人认为合理的初始化策略是值得探讨的一个问题。

感谢您耐心的回答

请问在validation的时候是否也用了drop out,这是合理情况吗?在test的时候是否会使用dropout?

貌似training和validation共用一套计算图,而且dropout在计算图中是写死的,没有发现training flag之类的分支

> This seems related to what you have found: https://github.com/cch1999/posecheck/blob/main/changelog.md#11---18-01-2024 So which version should I follow? The new version strain energy is quite different from PoseCheck paper results

Hi, I tried a quick fix (change to python=3.11.0) and now successfully built the enviroment. I can run test.py in this environment, but I'm not sure this will work with...