ZeroRin

Oh, I just noticed that maybe I can achieve this by simply passing `push_options: --dry-run` to the action

You're right. In my own repo I didn't rely on this action for the final push and simply call a `run: git push` at the end, so I didn't notice...

节点分类问题本来就是可见测试数据节点的啊，这也是与前面TextGCN等工作保持一致的基本设定。 当然，不同于TextGCN，引入了文本embedding之后理论上确实可以使得这个方法能扩展到训练时未见的graph上就是了，有兴趣看看效果的话可以自己做下。

原理上这个模型是可以用在使用同样方式建立的所有节点都没见过的新图上的。相对的，原版TextGCN使用onehot初始表征，相当于让模型给每个节点凭空学了一个embedding，原理上是无法扩展的。本身是有计划对这方面进一步探究的，但后来因为研究方向调整就没有再继续了。 我记得在中稿之后我曾经尝试过训练时只使用训练集构图，测试时使用训练集+测试集构图，性能指标上基本没有差异，但差别更大的图情况下效果就不清楚了。

细节已经记不清了，应该差不多吧

The default parameters were designed to utilize a V100 GPU. Try to reduce the batch size and it should be working.

seems that you failed to upload your figure. I tried the code on gpu with 12GB memory by reducing batchsize to 16, the model was trained succesfully. Probably something else...

> well so it seems to be impossible for a 6GB GPU to run this BERT - GCN model, :( this implementation send the full graph into GCN during training...

Consider using [wcwidth](https://github.com/jquast/wcwidth)? In my experience in other projects this is more accurate for control characters, non-asian characters, etc. (see [this](https://github.com/masaccio/compact-json/issues/4#issuecomment-1357049123)) Also wcwidth uses `lru_cache` which should be helpful for...

There also exist packages like [rich](https://github.com/Textualize/rich/blob/master/rich/cells.py) who generates a table with a tool script depend on `wcwidth` so that the package itself does not need it as dependency. Same strategy...