几个训练相关问题

[zdragon2002]

您好，请问：

1. Readme中说：cd VideoWorld
   bash ./tools/calvin_train.sh
   但是calvin_train.sh文件没有找到，是要install之后才能看到吗？
2. 请问install VideoWorld的话，大概需要多少硬盘空间？
3. 请问如果训练的话，需要多大显存？能否提供一个训练时长数据供参考：卡型号+训练时长?
4. 请问能提供预训练的pth吗？比如./work_dirs/init/magvit/iter_332800_new.pth之类的
5. 能否提供下Go的Train步骤？数据集什么的，

谢谢

[zdragon2002]

For English:

In the Readme, it says:
cd VideoWorld
bash ./tools/calvin_train.sh

However, I couldn't find the calvin_train.sh file. Do I need to install it first to see this file?

If I need to install VideoWorld, approximately how much disk space is required?

For training, how much VRAM is needed? Could you also provide some reference data on training time, such as GPU model + training duration?

Could you provide a pre-trained .pth file, like ./work_dirs/init/magvit/iter_332800_new.pth or similar?

Could you also share the training steps for Go? For example, details about the dataset and other requirements?

Thank you!

[MaverickRen]

1. Apologies for missing the file. I’ll update the codebase right away.
   2&3. The training of the VideoWorld model requires relatively low GPU memory and disk space (only 300M parameters), but it does need a large number of training steps. For instance, training the Go scenario takes around 3 days on 8 A100 GPUs.
2. The file ''iter_332800_new.pth'' is a pretrained weight we obtained using magvit on natural images. It can serve as a good initialization, and I will upload it to the cloud.
3. The Go dataset is provided in the Video-GoBench link. https://huggingface.co/datasets/maverickrzw/VideoGo-Bench?clone=true