More advanced bbox label class, with transformation support.

A wrapper for batching the data in dataloader.

A Conv2d -> Batchnorm -> silu/leaky relu block

A container for backbone features in torch.Tensor.

C3 in yolov5, CSP Bottleneck with 3 convolutions

RandAugment applied on event detection data.

MLP using 1x1 convs that keeps spatial dims

A container for RNN states, support partial update/reset. We maintain a dict of {worker_id: states} for each data worker.

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

A container for event repr in torch.Tensor.

Simple linear online tracker.

Grid or Block partition + Attn + FFN. NxC 'channels last' tensor layout. According to RW, NHWC attention is a few percent faster on GPUs (bu

Data augmentor for SSOD training, with weak and strong augmentation.

A wrapper of a seq of events, each data sample is a sub-seq of `length`. Compared to `SequenceForRandomAccess`, where all the loaded event seq ar

Tensorflow like 'SAME' wrapper for 2D average pooling

Original CBAM channel attention module, currently avg + max pool variant only.

This Dataset avoids the sharding problem by instantiating randomized stream concatenation at the batch and worker level. Pros: - Every sin

Conditionally Parameterized Convolution Inspired by: https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/condconv/condconv

Tensorflow like 'SAME' convolution wrapper for 2D convolutions

Downsample with input in NCHW [channel-first] format. Output in NHWC [channel-last] format.

LSTM with (depthwise-separable) Conv option in NCHW [channel-first] format.

Visualize predicted and GT bbox on event-converted RGB frames.

Record the labels of an event sequence.

Aggregate model predictions from TTA (Test-Time-Augmentation).

Focal loss for foreground-background classification (objectness).

IoU loss for bbox regression.

Pad input to desired height and width.

LayerNorm w/ fast norm option

An experimental 'lightweight' that sums avg + max pool first

An experimental 'lightweight' variant that sums avg_pool and max_pool results.

Simple linear velocity bbox tracker. Inspired by paper: Towards Unsupervised Object Detection from LiDAR Point Clouds.

Tensorflow like 'SAME' wrapper for 2D max pooling

First Attn in fixed local windows, then Attn in sparse global grids.

Mixed Grouped Convolution Based on MDConv and GroupedConv in MixNet impl: https://github.com/tensorflow/tpu/blob/master/models/official/mn

A wrapper containing many frames and their bbox labels.

Relative Position Embedding As per: https://gist.github.com/aravindsrinivas/56359b79f0ce4449bcb04ab4b56a57a2 Originally from: `Attention Augm

Relative Position Embedding As per: https://gist.github.com/aravindsrinivas/56359b79f0ce4449bcb04ab4b56a57a2 Originally from: `Attention Augm

Generate pseudo labels on training data.

Operates with NCHW [channel-first] format as input and output.

Rotary position embedding NOTE: This is my initial attempt at impl rotary embedding for spatial use, it has not been well tested, and will l

Similar as SequenceDataset, but we also use the SSODAugmentorGenX for data aug.

Selectable global pooling layer with dynamic input kernel size

Record how many timesteps we have seen for each sequence.

A wrapper for RandomAccessSequence. We perform data aug here.

Load labeled frames and `L` event reprs before it. Each frame is only loaded once.

A wrapper for datapipes. This is the one used in eval/test. Each datapipe is a full seq of events, with diff num of frames. Similar to torch.u

Original CBAM spatial attention module

A wrapper for model inference with TTA.

Removed the direct dependency on the backbone.

RNN-based MaxViT backbone + YOLOX detection head.

Attention based 2D feature pooling w/ learned (absolute) pos embedding. This is a multi-head attention based replacement for (spatial) average po

BAT Adapted from: https://github.com/BA-Transform/BAT-Image-Classification

BatchNorm + Activation This module performs BatchNorm + Activation in a manner that will remain backwards compatible with weights trained wit

r"""Creates a module that computes blurs and downsample a given feature map. See :cite:`zhang2019shiftinvar` for more details. Corresponds to

Bottleneck Attention Paper: `Bottleneck Transformers for Visual Recognition` - https://arxiv.org/abs/2101.11605 The internal dimensions of t

Constructs a circular ECA module. ECA module where the conv uses circular padding rather than zero padding. Unlike the spatial dimension, the

Classifier head w/ configurable global pooling and dropout.

Depthwise Conv + Conv

Base data module for event detection dataset/dataloaders. There are two possible datasets, random access and streaming. In random access, the

DropBlock. See https://arxiv.org/pdf/1810.12890.pdf

Constructs an ECA module. Args: channels: Number of channels of the input feature map for use in adaptive kernel sizes for ac