HongCheng comments

Results 62 comments of


                                            HongCheng

how can we use this tool with a model that have multiple parts in input?

> For example, my model takes two inputs like the following coding shows > > ` outputs = net(inputs, tr_angles)` I have tried `graph = hl.build_graph(net, inputs, tr_angles)` and `graph...

Update README.md

I find a good slide related to uncertainty and deep learning, here is the link "https://www.gatsby.ucl.ac.uk/~balaji/balaji-uncertainty-talk-cifar-dlrl.pdf" Best regards!

Sharing DeBERTa-v3 discriminator and generator with task-specific heads?

can I ask how to use this generator? I have no idea to use it with the old debertaV3

asking for video training and inference samples

Thank you ! We are thrilled to hear that Our MiniSora Community is planning to re-implement Sora, and we appreciate your recognition of OpenDiT as a strong baseline. Please keep...

spatial temporal embedding 的定义方式

> 更新了下新的实现方式，代码参考来源[Latte](https://github.com/Vchitect/Latte/blob/main/models/latte.py#L270)。本质是2d的xy与时间embedding分开。我觉得你这样也没问题，模型其实很快会学到2+1d的映射和3d的位置编码是差不多的。 `temp_embed = get_1d_sincos_temp_embed(self.temp_embed.shape[-1], self.temp_embed.shape[-2]) # line270 Latte` 现在很多实现方式都是(2+1)d, 但是时间维度真的可以这样和xy做相加么, 有没有别的更好的方法,

An open-source third-party training with 8 GPUs

thanks for sharing, I would like to try this with your future inference support

Community Integration: Making Sora cheaper, faster, and more efficient

Thanks for your work. Would you mind if I ask you if your models could provide some inference output samples ?

Update 2.使用 Embedding API.ipynb

这个还需要一定的验证，助教团会尽快核实，

更新了第一章文档中的一些时间。

之前有贡献者提交了关于“2024 年 1 月 29 日发布了 Baichuan 3”的pr : #76 和对应的notebook修改 #87

My Architecture Overhaul Practical Roadmap for Faster and Less Resources T2V Generation

> If you are going to make a paper, please add me to the co-authors > > My name is: Artem Khrapov email: [[email protected]](mailto:[email protected]) Affiliation: Huawei Noah's Ark Lab >...