Resolution is an image’s pixel dimensions. In generated-image projects, it distinguishes smaller 64 × 64 images from higher-resolution 256 × 256 images. section 5.4.1
Higher resolution changes both the generated result and the training requirements:
- A 256 × 256 color image becomes a 3 × 256 × 256 channel-first tensor in PyTorch. section 5.4.1
- Larger images require more memory, so a batch size of 16 may need to be reduced to 10 or 8. section 5.4.1
- More sophisticated generators and discriminators can use 2D convolutional and transposed-convolutional layers for color images. section 4.4
import torchvision.transforms as T import torchvision batch_size=16 imgsz=256 transform=T.Compose([ T.Resize((imgsz,imgsz)), T.ToTensor(), T.Normalize([0.5,0.5,0.5],[0.5,0.5,0.5])]) data_set=torchvision.datasets.ImageFolder( root=r"files/glasses", transform=transform)
Resize training images to 256 × 256 pixels, convert them to tensors, and normalize their color values before loading them. section 5.4.1
