Inquiry Regarding GPU Memory and Number Requirements and Training Duration

[hero-y]

I have a few questions regarding the computational resources required to do so.

GPU Memory Requirement: Could you please specify the amount of GPU memory required for training the models discussed in the paper? I would appreciate knowing if there are any specific GPU models that are recommended.

Number of GPUs: How many GPUs are needed to achieve the results presented in the paper? If it's scalable, could you provide a guideline on how the number of GPUs might affect the training process?

Training Duration: What is the approximate training time for the models using the recommended setup? Is this time frame based on a specific GPU configuration?

[andvg3]

Hello @hero-y ,

We trained LGD on a cluster of three GPUs NVIDIA Quadro 4080 over 1000 epochs. Each GPU requires about 10-20GB RAM if I remember correctly. The training time was less than 3 days.

If it's scalable, could you provide a guideline on how the number of GPUs might affect the training process?

Yes, it is scalable, you can modify the number of GPUs by setting --nproc_per_node=n, where n is your desired number.