zhuiyitechnology/roformer

Rotary Transformer

roformer's Issues

max_position_embeddings have restrict？

"max_position_embeddings": 1536 ？
Can it be bigger? thks

download中的这几个模型都有什么差别啊

请问字级别的roformer可以自己预训练吗？

Reference to prior work

Hey Everybody,

first of all congrats to the paper, it's really interesting and am looking forward to seeing its impact on the space!

I wanted to point out that a few colleagues and me followed a similar approach to introduce translational equivariance in kernelizable attention (as is implemented in the Performer) for image classification tasks and posted it on ArXiv in the beginning of February https://arxiv.org/abs/2102.07680.

While the approach proposed in your work is more generic, we would highly appreciate if you could also refer to our prior work in the publication.

Best,
Max

期待线性attention加unilm式的预训练模型

如题！这个预训练模型出来，别的不说，大部分nlp任务都可以用它先跑个baseline：）

Changing bases and sampling rate

Hi,

I am a hobbyist in AI, and I am interested to hear your thoughts on researching further developing RoPE.

Recent advancement in rotary token encoding with a focus on extending tokens by interpolation has merged with NTK theory, which necessarily requires a higher-order non-linear basis for encoding data so that NN can learn both low and high-frequency features instead of converging to low freq feat early.

The discussion below shows extending the context range by using a "simply non-linearly interpolated basis w/o fine tuning" could achieve comparable results with a "fine-tuned model on longer tokens with linearly interpolated basis":
https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/ntkaware_scaled_rope_allows_llama_models_to_have/
It's like zero-shot extending token length.

The following CoLab from the above discussion has demonstrated a non-linear power-series bases have achieved good results on 8k tokens inputs (which the meta paper on RoPE interpolation https://arxiv.org/abs/2306.15595)
https://colab.research.google.com/drive/1VI2nhlyKvd5cw4-zHvAIk00cAVj2lCCC#scrollTo=b80b3f37

Some typos on your paper

Hello @ZhuiyiTechnology

In the section Instruction on the first page.

The sequential order of words is of great value to natural language understanding. "Recurrent neural networks (RRNs)"
based models encode tokens’ order by recursively computing a hidden state along the time dimension.

It should be RNN, right ?

I have found that there are two parts in the first page that using RRN not RNN.

It's not a big deal to read this great work, just want to mention.

bert4keras 现在好像没有0.10.4这个版本