swin3d_b

torchvision.models.video.swin3d_b(*, weights: Optional[Swin3D_B_Weights] = None, progress: bool = True, **kwargs: Any) → SwinTransformer3d

Constructs a swin_base architecture from Video Swin Transformer.

Parameters:

• weights (Swin3D_B_Weights, optional) – The pretrained weights to use. See Swin3D_B_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.video.swin_transformer.SwinTransformer base class. Please refer to the source code for more details about this class.

class torchvision.models.video.Swin3D_B_Weights(value)

The model builder above accepts the following values as the weights parameter. Swin3D_B_Weights.DEFAULT is equivalent to Swin3D_B_Weights.KINETICS400_V1. You can also use strings, e.g. weights='DEFAULT' or weights='KINETICS400_V1'.

Swin3D_B_Weights.KINETICS400_V1:

The weights were ported from the paper. The accuracies are estimated on video-level with parameters frame_rate=15, clips_per_video=12, and clip_len=32 Also available as Swin3D_B_Weights.DEFAULT.

acc@1 (on Kinetics-400)	79.427
acc@5 (on Kinetics-400)	94.386
categories	abseiling, air drumming, answering questions, … (397 omitted)
min_size	height=1, width=1
min_temporal_size	1
recipe	link
num_params	88048984
GFLOPS	140.67
File size	364.1 MB

The inference transforms are available at Swin3D_B_Weights.KINETICS400_V1.transforms and perform the following preprocessing operations: Accepts batched (B, T, C, H, W) and single (T, C, H, W) video frame torch.Tensor objects. The frames are resized to resize_size=[256] using interpolation=InterpolationMode.BILINEAR, followed by a central crop of crop_size=[224, 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]. Finally the output dimensions are permuted to (..., C, T, H, W) tensors.

Swin3D_B_Weights.KINETICS400_IMAGENET22K_V1:

The weights were ported from the paper. The accuracies are estimated on video-level with parameters frame_rate=15, clips_per_video=12, and clip_len=32

acc@1 (on Kinetics-400)	81.643
acc@5 (on Kinetics-400)	95.574
categories	abseiling, air drumming, answering questions, … (397 omitted)
min_size	height=1, width=1
min_temporal_size	1
recipe	link
num_params	88048984
GFLOPS	140.67
File size	364.1 MB

The inference transforms are available at Swin3D_B_Weights.KINETICS400_IMAGENET22K_V1.transforms and perform the following preprocessing operations: Accepts batched (B, T, C, H, W) and single (T, C, H, W) video frame torch.Tensor objects. The frames are resized to resize_size=[256] using interpolation=InterpolationMode.BILINEAR, followed by a central crop of crop_size=[224, 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]. Finally the output dimensions are permuted to (..., C, T, H, W) tensors.