rudongyu

Results 23 comments of rudongyu

> Have you tried re-running the scripts for a few epochs to verify successful running? Yes, tested for a few epochs.

> @rudongyu If this is ready for review, you can request a review from Quan and me. Also perhaps we will want to move the sampler to the core codebase....

The dataloading of using dgl dataloader & sampler has a near 4x slower speed when i test it (almost the same computation for a batch). @BarclayII @mufeili Could either of...

> I run `main.py` on the ogbl-collab dataset, but got an exception at Line 387 `graph = graph.to_simple(copy_edata=True, aggregator='sum')`. The error message is `dgl._ffi.base.DGLError: [09:32:16] /opt/dgl/src/array/kernel.cc:399: Check failed: (feat->dtype).code ==...

The PR is ready for another round of review. @jermainewang. However, the dataloading and the sampler are implemented a little bit complicated to fit current interface of dataloader. I thinks...

It seems to be a bug of sampling. `to_block` here https://github.com/dmlc/dgl/blob/d41d07d0f6cbed17993644b58057e280a9e8f011/python/dgl/dataloading/neighbor_sampler.py#L115 wrongly uses `seed_nodes` as destination nodes when the `edge_dir` is 'out'. @BarclayII for awareness.

Hi, @aries-M. We'd like to know the scenario where you use the sampler with 'out' direction. Will the message passing follow the same direction? It would be helpful if some...

Hi @aries-M. I think in your case, using `dgl.reverse` to convert the edges to reversed ones will be a better choice. In our future plan, we will deprecate this option...

> 您好! > > 很抱歉又打扰了,最近在DWIE数据集上尝试复现BiLSTM一类模式时发现,按照docred默认的参数设置,模型在验证集上的f1基本只能收敛到40左右,离您论文中报道的验证集f1达到50还有着较大的差距,一方面我感觉和验证集数据量太少有关,一方面也和具体参数设置有关,不知道您怎么看,可以的话,能不能请教一下你们的一些具体训练设置,比如学习率,batch_size,最大长度截断,以及有无使用比如dropout或者lr_schedule等技巧。 > > 期待您的回复,非常感谢! > > 祝好 您好! 感谢反馈!不知道之前 ATLOP 那个结果您现在对齐了吗?后面我用他们在 docred 上的原始代码只做了必要的修改(关系数目,增加头尾实体相同的关系等)又试了一下,和之前给出的结果是接近的,应该和超参以及长度截断关系不是很大。想确认一下有没有可能是一类问题。

> 您好! > > 因为最近对你们在DWIE数据集上的实验十分感兴趣,我最近也在DWIE数据集上尝试复现一些模型结果,但结果好像都与您们论文里报道的有些差别,考虑原因可能是DWIE数据集上的很多文档长度会超过BERT的最大输入限制512,请问一下您们是用的ATLOP中的滑动窗口处理的吗?但ATLOP中的滑动窗口貌似也是扩充到了1024,实际还是不能满足很多文档长度需求,想请问一下您们是如何处理的呢? > > 期待您的回复,非常感谢! > > 祝好 您好!感谢关注! 请问您指的是 backbone 的结果差别比较大吗?具体差异有多少呢,下面给出一组 ATLOP 上收敛的 trace 给您做参考: ![image](https://user-images.githubusercontent.com/16982108/144346467-1b5cea58-2fa6-490e-91b3-36b9c81837f5.png) 关于 context 长度溢出的处理,目前我们遵循 ATLOP 的设定,用两个 512 长度的窗口叠加,受制于显存大小,大于 1024 的部分仍被截断。当然,超长文本上的抽取确实也还是一个值得探讨的问题,所以近期也有一些工作关注在关系抽取之前进行 model-based...