potential correction on train.py

[yilongzhao1122]

in train.py line61 'model = model.cuda()' should be corrected as 'model = model.to(device)'. otherwise, cpu is not allowed for training.

[zhoubay]

Thank you for your suggestion! If you have time, feel free to start a pull request to contribute to this code!

[lillian2015]

GPU must be much faster than CPU, right?

Anyway, the team did a great job. We all have to admit it.

The author team should have developed the program on a GPU platform themselves, without even considering whether your CPU can run it or not - after all, as researchers, they may also hope to leave some gap between them, and the readers, not hope anyone to be able to use it by a too simple way.

Moreover, unlike CLEAN, which provides pre-trained datasets directly, you can use them directly after downloading and installing, even no need for a training process;
Train by yourself, or the author TEAM, give the same effect. There is no real difference between the results calculated by GPU or CPU.

If you have to run the program on the CPU platform machine, then modify all CUDA to CPU, or “CUDA or CPU” mannually; After all, most of the basic and common bioinformatics toolkits accessible to the external network support GPUs or CPUs;

The Catapro tool is also very good, but it does not support CPU natively. Except from modifying the team's CUDA only, to CPU support, then you can run it smoothly on your CPU platform. there is no way out there.

Afterall, you can not day-dream the TEAM to cater these poor reders without a GPU platform....
Not even to mention that, as Chinese, to create a simplified Chinese protocol out there, to facilitate the Chinese readers to use this tool more convenient....

[zhoubay]

Hi @lillian2015 , thank you for your long note! For starter, the GPUs are much powerful than CPUs, especially on trianing deep learning models, and runing training on CPUs would take like 100x times longer time, making it almost impossible to train models on CPUs. Based on this idea, it's not recommended to train models on purely CPUs.

And changing model from GPU to CPU is as much easy as changing a line of code, so if you have your own requests, you could just modify that code of line then train your models on CPU.

For another solution, if you have time, we are more than welcome to see you modify the codes to make it support both the CPUs and GPUs using argparse package! Send us a PR, and we'll audit the codes! If we find it's cool, we could merge it to the repo!

For CLEAN, the tool is trained for zero-shot inference, where further training is not necessary. In contrast, ESM-Ezy is a tool that could help you adapt the model in your own data. Comparing the two model is not fair.

And for Chinese supports, as the same suggestion above, you could draft a Chinese version README or another supports that you think should have and then send us a PR. We would audit it as soon as possible once we saw that!

Thank you for your suggesions and thank you for your attention for the ESM-Ezy model!

[lillian2015]

很感谢开发团队的回复；我也就是出来冒个泡泡，而已；
GPU的强大，我自然是有所了解；
除了魔改我本地机子上的代码之外，我也没有多少闲余时间“大庇天下广寒人士尽欢颜”，毕竟我自己还没用上GPU平台的机子，才有这些有些抱怨的辞令【而csdn上面，已经有相关的魔改教程】；
另外，我都有点不敢相信，我一个没有学过生信，没有学过分子生物学的 临床兽医学 的底层科研工作者，有一天能够借助ai的风，走在修改代码的路上；

[zhoubay]

很感谢开发团队的回复；我也就是出来冒个泡泡，而已； GPU的强大，我自然是有所了解； 除了魔改我本地机子上的代码之外，我也没有多少闲余时间“大庇天下广寒人士尽欢颜”，毕竟我自己还没用上GPU平台的机子，才有这些有些抱怨的辞令【而csdn上面，已经有相关的魔改教程】； 另外，我都有点不敢相信，我一个没有学过生信，没有学过分子生物学的 临床兽医学 的底层科研工作者，有一天能够借助ai的风，走在修改代码的路上；

Thank you for your reply! If you have no experience on coding or have no access to your own GPUs, my lab developed another tool that could be useful for your research called SaprotHub, published in Nature Biotechonology. As the title of the paper indicate, it is designed to democratizing protein language model training, sharing and collaboration. The only thing you need is to register a google account, and you would have access to the free GPUs. And if you buy the Pro service from google, you could access to the advanced GPU resources. You could click this to go to the website.

The tutorial video is both available on Youtube and Bilibili. The only obstacles to use this website are basic understanding about Machine Learning and basic English listening and reading ability.

You could refer to the SaprotHub github repo to find the QR code and join in the wechat group and share your experience or issues.

Good to know your idea, if you have any other issues, don't hesitate to contact us. We would be more than welcome to hear from you!