| import torch | |
| import torchvision | |
| from torch import nn | |
| def create_vit_model(num_classes: int = 3, | |
| seed: int = 42): | |
| # Create ViT_B_16 pre-trained weights, transforms and model | |
| weights = torchvision.models.ViT_B_16_Weights.DEFAULT | |
| transforms = weights.transforms() | |
| model = torchvision.models.vit_b_16(weights = weights) | |
| # Freeze all of the base layers | |
| for param in model.parameters(): | |
| param.requires_grad = False | |
| # Change classifier head to suit our needs | |
| torch.manual_seed(seed) | |
| model.heads = nn.Sequential(nn.Linear(in_features = 768, | |
| out_features = num_classes)) | |
| return model, transforms | |