The Registry class provides a registry for all things To initialize: REGISTRY = Registry() To register a tracker: @R

Element constructing the S3D/S3DG. See models/base/backbone.py L99-186. Modifed from https://github.com/TengdaHan/CoCLR/blob/main/backbo

r""" LayerNorm that supports two data formats: channels_last (default) or channels_first. The ordering of the dimensions in the inputs. channels_

Constructs basic inception 3D conv. Modified from https://github.com/TengdaHan/CoCLR/blob/main/backbone/s3dg.py.

A scalar meter uses a deque to track a series of scaler values with a given window size. It supports calculating the median and average value

ResNet Stage containing several blocks.

From https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py. Drop paths (Stochastic Depth) per sample (whe

A timer which computes the time elapsed since the start/reset of the timer.

Self-attention module. Currently supports both full self-attention on all the input tokens, or only-spatial/only-temporal self-attention

Modified from https://github.com/TengdaHan/DPC/blob/master/utils/augmentation.py. Randomly change the brightness, contrast and saturation o

Global config object. It automatically loads from a hierarchy of config files and turns the keys to the class attributes.

Performs temporally adaptive 2D convolution. Currently, only application on 5D tensors is supported, which makes TAdaConv2d essentia

Constructs a base 3D block, composed of a shortcut and a conv branch.

Perform the multi-view ensemble for testing: each video with an unique index will be sampled with multiple clips, and the predictions of the

Video to Patch Embedding.

Randomly selects a rectangle region in an image and erases its pixels. 'Random Erasing Data Augmentation' by Zhong et al. See https:/

The routing function for generating the calibration weights.

Perform the multi-view ensemble for testing: each video with an unique index will be sampled with multiple clips, and the predictions of the

AugMix Transform Adapted and improved from impl here: https://github.com/google-research/augmix/blob/master/imagenet.py From paper: 'AugMix:

Standard video model. The model is divided into the backbone and the head, where the backbone extracts features and the head performs cla

r"""Implements layer-wise adaptive rate scaling for SGD. Args: params (iterable): iterable of parameters to optimize or dicts defining

Mixup/Cutmix that applies different params to each element or whole batch Args: mixup_alpha (float): mixup alpha value, mixup is active

Model Exponential Moving Average V2 Keep a moving average of everything in the model state_dict (parameters and buffers). V2 of this module i

Modified from DistributedSampler, which performs multi fold training for accelerating distributed training with large batches. Sampler t

Modified from DistributedSampler, which performs multi fold training for accelerating distributed training with large batches. Sampler t

The routing function for generating the calibration weights.

Performs temporally adaptive 2D convolution. Currently, only application on 5D tensors is supported, which makes TAdaConv2d essentia

Measure training stats.

Measures validation stats.

Constructs basic ResNet 2D Stem. A single 2D convolution is performed in the base 2D stem.

Constructs basic ResNet 3D Stem. A single 3D convolution is performed in the base 3D stem.

Head for predicting boundary matching map.

Constructs the base convolution branch for ResNet based approaches.

Constructs base head.

Constructs two base heads in parallel. This is specifically for EPIC-KITCHENS dataset, where 'noun' and 'verb' class are predicted.

Constructs base module that contains basic visualization function and corresponding hooks. Note: The visualization function has only tested i

The BaseVideoDataset object provides a base object for all the video/image/video-text datasets. Abstract methods are provided for completio

The ir-CSN branch. See Du Tran et al. Video Classification with Channel-Separated Convolutional Networks.

r""" ConvNeXt A PyTorch impl of : `A ConvNet for the 2020s` - https://arxiv.org/pdf/2201.03545.pdf Args: in_chans (int

Inherits base 3D stem and adds a maxpool as downsampling.

The factorized transformer. See Anurag Arnab et al. ViVIT: A Video Vision Transformer.

Backbone architecture for I3D/S3DG. Modifed from https://github.com/TengdaHan/CoCLR/blob/main/backbone/s3dg.py.

Generator for pseudo camera motions with static masks in MoSI. See paper "Self-supervised Motion Learning from Static Images", Huang

Head for predicting MoSI categories. This head performs prediction jointly on both axes.

Head for predicting MoSI categories. This head performs prediction on only x-axis.

Head for predicting MoSI categories. This head performs prediction on only y-axis.

Non-local block. See Xiaolong Wang et al. Non-local Neural Networks.

The R2D3D Branch. Essentially the MCx model in Du Tran et al. A Closer Look at Spatiotemporal Convoluitions for Action Recogn

The R(2+1)D Branch. See Du Tran et al. A Closer Look at Spatiotemporal Convoluitions for Action Recognition.

R(2+1)D Stem.

Meta architecture for 3D ResNet based models.

Element constructing the S3D/S3DG. See models/base/backbone.py L99-186. Modifed from https://github.com/TengdaHan/CoCLR/blob/main/backbo

Backbone architecture for temporal action localization, which only contains three simple convs.

Constructs head for the SlowFast Networks.

SlowFast Head for EPIC-KITCHENS dataset.

Constructs SlowFast model. Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. "SlowFast networks for video recognit

Constructs SlowFast conv branch. See Christoph Feichtenhofer et al. SlowFast Networks for Video Recognition.

The TAdaConv branch with average pooling as the feature aggregation scheme. For details, see Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiw

r""" TAdaFormer Block. Args: cfg (Config): the global config object.

Performs temporally adaptive 3D convolution.

r""" TAdaConvNeXtV2 Block. Args: cfg (Config): the global config object. drop_path (float): Stochastic depth rate. Default: 0.0

Transformer layer for (L, N, D) input. Modified from CLIP.

Construct head for video vision transformers.

The Transformer head for EPIC-KITCHENS dataset.

Video to Tubelet Embedding.