[Question] Clarification on Training Iterations and Pixel-Level Steps

[Ar0Kim]

Thank you for sharing this amazing research!

I have a couple of questions regarding the training iterations.

In the paper, it states that "The pixel-level and semantic-level LoRA modules undergo 4K and 8.5K training iterations."
However, in the training_utils.py file, I noticed the line:
parser.add_argument("--max_train_steps", type=int, default=100000)
which sets the default number of training steps to 100K.
Could you please clarify how we should configure the training iterations to match what's described in the paper?

Also, should the pixel-level training steps be adjusted depending on the batch size or the number of GPUs used?
For example, if I train with a batch size of 2 on a single GPU, does that mean the total number of training iterations would become 800K, and the pixel-level steps would be 32K?

Thanks in advance for your help!

[PhuTran1005]

Do you have answer for this question?

[Ar0Kim]

Do you have answer for this question?

Nope..