r3d_18

torchvision.models.video.r3d_18(*, weights: Optional[R3D_18_Weights] = None, progress: bool = True, **kwargs: Any) → VideoResNet

Construct 18 layer Resnet3D model.

Warning

The video module is in Beta stage, and backward compatibility is not guaranteed.

Reference: A Closer Look at Spatiotemporal Convolutions for Action Recognition.

Parameters:

• weights (R3D_18_Weights, optional) – The pretrained weights to use. See R3D_18_Weights below for more details, and possible values. By default, no pre-trained weights are used.

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

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

class torchvision.models.video.R3D_18_Weights(value)

The model builder above accepts the following values as the weights parameter. R3D_18_Weights.DEFAULT is equivalent to R3D_18_Weights.KINETICS400_V1. You can also use strings, e.g. weights='DEFAULT' or weights='KINETICS400_V1'.

R3D_18_Weights.KINETICS400_V1:

The weights reproduce closely the accuracy of the paper. The accuracies are estimated on video-level with parameters frame_rate=15, clips_per_video=5, and clip_len=16. Also available as R3D_18_Weights.DEFAULT.

acc@1 (on Kinetics-400)	63.2
acc@5 (on Kinetics-400)	83.479
min_size	height=1, width=1
categories	abseiling, air drumming, answering questions, … (397 omitted)
recipe	link
num_params	33371472

The inference transforms are available at R3D_18_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=[128, 171] using interpolation=InterpolationMode.BILINEAR, followed by a central crop of crop_size=[112, 112]. Finally the values are first rescaled to [0.0, 1.0] and then normalized using mean=[0.43216, 0.394666, 0.37645] and std=[0.22803, 0.22145, 0.216989]. Finally the output dimensions are permuted to (..., C, T, H, W) tensors.