vit_b_16

torchvision.models.vit_b_16(*, weights: Optional[torchvision.models.vision_transformer.ViT_B_16_Weights] = None, progress: bool = True, **kwargs: Any) → torchvision.models.vision_transformer.VisionTransformer

Constructs a vit_b_16 architecture from An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.

Parameters

• weights (ViT_B_16_Weights, optional) – The pretrained weights to use. See ViT_B_16_Weights below for more details and possible values. By default, no pre-trained weights are used.

• progress (bool, optional) – If True, displays a progress bar of the download to stderr. Default is True.

• **kwargs – parameters passed to the torchvision.models.vision_transformer.VisionTransformer base class. Please refer to the source code for more details about this class.

class torchvision.models.ViT_B_16_Weights(value)

The model builder above accepts the following values as the weights parameter. ViT_B_16_Weights.DEFAULT is equivalent to ViT_B_16_Weights.IMAGENET1K_V1. You can also use strings, e.g. weights='DEFAULT' or weights='IMAGENET1K_V1'.

ViT_B_16_Weights.IMAGENET1K_V1:

These weights were trained from scratch by using a modified version of DeIT’s training recipe. Also available as ViT_B_16_Weights.DEFAULT.

acc@1 (on ImageNet-1K)	81.072
acc@5 (on ImageNet-1K)	95.318
categories	tench, goldfish, great white shark, … (997 omitted)
num_params	86567656
min_size	height=224, width=224
recipe	link

The inference transforms are available at ViT_B_16_Weights.IMAGENET1K_V1.transforms and perform the following preprocessing operations: Accepts PIL.Image, batched (B, C, H, W) and single (C, H, W) image torch.Tensor objects. The images are resized to resize_size=[256] using interpolation=InterpolationMode.BILINEAR, followed by a central crop of crop_size=[224]. Finally the values are first rescaled to [0.0, 1.0] and then normalized using mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225].

ViT_B_16_Weights.IMAGENET1K_SWAG_E2E_V1:

These weights are learnt via transfer learning by end-to-end fine-tuning the original SWAG weights on ImageNet-1K data.

acc@1 (on ImageNet-1K)	85.304
acc@5 (on ImageNet-1K)	97.65
categories	tench, goldfish, great white shark, … (997 omitted)
recipe	link
license	link
num_params	86859496
min_size	height=384, width=384

The inference transforms are available at ViT_B_16_Weights.IMAGENET1K_SWAG_E2E_V1.transforms and perform the following preprocessing operations: Accepts PIL.Image, batched (B, C, H, W) and single (C, H, W) image torch.Tensor objects. The images are resized to resize_size=[384] using interpolation=InterpolationMode.BICUBIC, followed by a central crop of crop_size=[384]. Finally the values are first rescaled to [0.0, 1.0] and then normalized using mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225].

ViT_B_16_Weights.IMAGENET1K_SWAG_LINEAR_V1:

These weights are composed of the original frozen SWAG trunk weights and a linear classifier learnt on top of them trained on ImageNet-1K data.

acc@1 (on ImageNet-1K)	81.886
acc@5 (on ImageNet-1K)	96.18
categories	tench, goldfish, great white shark, … (997 omitted)
recipe	link
license	link
num_params	86567656
min_size	height=224, width=224

The inference transforms are available at ViT_B_16_Weights.IMAGENET1K_SWAG_LINEAR_V1.transforms and perform the following preprocessing operations: Accepts PIL.Image, batched (B, C, H, W) and single (C, H, W) image torch.Tensor objects. The images are resized to resize_size=[224] using interpolation=InterpolationMode.BICUBIC, followed by a central crop of crop_size=[224]. Finally the values are first rescaled to [0.0, 1.0] and then normalized using mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225].