Open jiangroubao opened 3 weeks ago
In reality you have three options:
What you prefer?
You can disable reference image preloading by modifying the code. Just make sure to release preallocations from Torch.
This approach should only affect training speed by a little.
It seems that this version runs out of GPU memory more easily compared to the previous one. With the same quality and quantity of photos, the previous version could complete the training (30,000 steps), but now this version throws an out-of-memory exception. Is there any way to resolve this?