Now I’m researching for my project for this week and I found a great idea.
But the images I want to compare contains bit more details.
I feel 224x224 might not be enough for that.
If I want to increase the size what are the rule of thumbs. If I did so, what are cons and how I can mitigate them?
Like using a powerful GPU.
You may have to shrink the batch size. Though if you shrink too much the batchnorm performance will be affected. You also may want to use multiple gpu’s with model parallelism if the image sizes are really big There is also, gradient checkpointing which can tradeoff speed for memory but haven’t used it in pytorch/fastai
Ah okay. I’ll try to make is smaller as possible.