Bottleneck-Transformers-for-Visual-Recognition

Introduction

PyTorch Implementation of Bottleneck Transformers for Visual Recognition.

Link to paper: https://arxiv.org/abs/2101.11605

Structure of Self-Attention layer in paper:

Dependency

External libraries: eniops, yacs

Use pip command to install:

$ pip install eniops
$ pip install yacs

Train the Network

To train BoTNet from scratch using the standard training process:

$ python3 train.py

To execute this code, you should have ImageNet downloaded in your local directory and correct path set in _C.TRAIN.DATASET and_C.TEST.DATASET in util/config.py.

You can use .yaml or pass in any parameters in command line to overwrite the configures in config.py.

For exmaple, to update cfg.TRAIN.BATCH_SIZE, you can either type --train-batch_size in command line or pass a .yaml file through --cfg using the below format:

TRAIN:
  BATCH_SIZE: 256

References

Part of the codes are cited from:

Bottleneck Transformer Backbone:

https://github.com/lucidrains/bottleneck-transformer-pytorch https://gist.github.com/aravindsrinivas/56359b79f0ce4449bcb04ab4b56a57a2

Training Process:

https://github.com/BIGBALLON/distribuuuu (Biggest credit to my mentor at Sensetime!!)

Features to come:

• Downsample at the first stage of attention layer with correct dimensions
• Object detection experiments for downstream task