Transform for converting output from dict to a tuple following PySlowFast dataset output format.

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

Transform for converting video frames as a list of tensors. Each tensor corresponding to a unique pathway.

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

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

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

Subclass torch's LayerNorm to handle fp16.

Fuses the information from the Fast pathway to the Slow pathway. Given the tensors from Slow pathway and Fast pathway, fuse information from

Measure the AVA train, val, and test stats.

Subclass torch's LayerNorm to handle fp16.

Wrapper for PyTorchVideo datasets.

Image to Patch Embedding

Internal implementation of Pascal object detection metrics.

Swish activation function: x * sigmoid(x).

Measure training stats.

Measures validation stats.

Apply for video.

Charades video loader. Construct the Charades video loader, then sample clips from the videos. For training and validation, a single clip is

A timer which computes the epoch time.

A class used to get weights and activations from specified layers from a Pytorch model.

GradCAM class helps create localization maps using the Grad-CAM method for input videos and overlap the maps over the input videos as heatmap

Apply mixup and/or cutmix for videos at batch level. mixup: Beyond Empirical Risk Minimization (https://arxiv.org/abs/1710.09412) CutMix:

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

A ResNet class that is similar to torchvision's but contains the following changes: - There are now 3 "stem" convolutions as opposed to 1, wi

This class defines multigrid training schedule and update cfg accordingly.

Builds Non-local Neural Networks as a generic family of building blocks for capturing long-range dependencies. Non-local Network computes

Action Predictor for action recognition.

Crop the given PIL Image to random size and aspect ratio with random interpolation. A crop of random size (default: of 0.08 to 1.0) of the origina

Residual block.

Squeeze-and-Excitation (SE) block w/ Swish: AvgPool, FC, Swish, FC, Sigmoid.

Extend Sampler to support "short cycle" sampling. See paper "A Multigrid Method for Efficiently Training Video Models", Wu et al., 2019 (

Image to Patch Embedding

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

Measure the AVA train, val, and test stats.

Visualize action predictions for videos or folder of images with precomputed and ground-truth boxes in AVA format.

ANet video loader. Construct the ANet video loader, then sample clips from the videos. For training and validation, a single clip is rand

Asynchronous Action Prediction and Visualization pipeline with AsyncVis.

AVA Dataset

Basic transformation: Tx3x3, 1x3x3, where T is the size of temporal kernel.

Bottleneck transformation: Tx1x1, 1x3x3, 1x1x1, where T is the size of temporal kernel.

Box collection. BoxList represents a list of bounding boxes as numpy array, where each bounding box is represented as a row of 4 numbers,

Naming conventions for BoxLists. Attributes: boxes: bounding box coordinates. classes: classes per bounding box. scores: scores

Convenience wrapper for BoxList with masks. BoxMaskList extends the np_box_list.BoxList to contain masks as well. In particular, its construc

Interface for object detection evalution classes. Example usage of the Evaluator: ------------------------------ evaluator = DetectionEva

Naming conventions for storing the output of the detector. Attributes: source_id: source of the original image. key: unique key corre

A timer which computes the epoch time.

GroupGather performs all gather on each of the local process/ GPU groups.

ImageNet dataset.

Names for the input tensors. Holds the standard data field names to use for identifying input tensors. This should be used by the decoder to

Kinetics video loader. Construct the Kinetics video loader, then sample clips from the videos. For training and validation, a single clip is

Kinetics video loader. Construct the Kinetics video loader, then sample clips from the videos. For training and validation, a single clip is

Multiscale Vision Transformers Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, Christoph Feichtenhofer

MiT video loader. Construct the MiT video loader, then sample clips from the videos. For training and validation, a single clip is random

Model Exponential Moving Average (DEPRECATED) Keep a moving average of everything in the model state_dict (parameters and buffers). This vers

A class to evaluate detections.

A class to evaluate detections using Open Images V2 metrics. Open Images V2 introduce group_of type of bounding boxes and this metric handles

CSN models using PyTorchVideo model builder.

R(2+1)D models using PyTorchVideo model builder.

ResNet models using PyTorchVideo model builder.

A class to evaluate detections using PASCAL metrics.

A class to evaluate instance masks using PASCAL metrics.

PatchEmbed.

Evaluate detection result of a single image.

ResNet model builder. It builds a ResNet like network backbone without lateral connection (C2D, I3D, Slow). Christoph Feichtenhofer, Hao

ResNe(X)t 3D head. This layer performs a fully-connected projection during training, when the input size is 1x1x1. It performs a convolut

ResNe(X)t 3D stem module. Performs spatiotemporal Convolution, BN, and Relu following by a spatiotemporal pooling.

Stage of 3D ResNet. It expects to have one or more tensors as input for single pathway (C2D, I3D, Slow), and multi-pathway (SlowFast) cas

SlowFast model builder for SlowFast network. Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. "SlowFast networks for