Masked 1D convolution. Interface remains the same as Conv1d. Only support a sub set of 1d convs

LayerNorm that supports inputs of size B, C, T

Computes and stores the average and current value. Used to compute dataset stats from mini-batches

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks) with a per channel scaling factor (and zero init) See

A simple conv block similar to the basic block used in ResNet

A simple (post layer norm) Transformer block Modified from https://github.com/karpathy/minGPT/blob/master/mingpt/model.py

Adapted from https://github.com/activitynet/ActivityNet/blob/master/Evaluation/eval_detection.py

Similar to nn.ParameterList, but for buffers Taken from https://github.com/facebookresearch/detectron2/blob/master/detectron2/modeling/ancho

Sets the learning rate of each parameter group to follow a linear warmup schedule between warmup_start_lr and base_lr followed by a cosine an

Sets the learning rate of each parameter group to follow a linear warmup schedule between warmup_start_lr and base_lr followed by a multi-ste

Local Multi Head Conv Attention with mask Add a depthwise convolution within a standard MHA The extra conv op can be used to (1) enc

Multi Head Conv Attention with mask Add a depthwise convolution within a standard MHA The extra conv op can be used to (1) encode re

1D Conv heads for classification

Shared 1D Conv heads for regression Simlar logic as PtTransformerClsHead with separated implementation for clarity

Multiply the output regression range by a learnable constant value

A backbone that with only conv

A backbone that combines convolutions with transformers

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

Feature pyramid network

Multi Head Attention with mask Modified from https://github.com/karpathy/minGPT/blob/master/mingpt/model.py

A generator for temporal "points" max_seq_len can be much larger than the actual seq length

Transformer based model for single stage action localization