Diverse Branch Block: Building a Convolution as an Inception-like Unit

Overview

Diverse Branch Block: Building a Convolution as an Inception-like Unit (PyTorch) (CVPR-2021)

DBB is a powerful ConvNet building block to replace regular conv. It improves the performance without any extra inference-time costs. This repo contains the code for building DBB and converting it into a single conv. You can also get the equivalent kernel and bias in a differentiable way at any time (get_equivalent_kernel_bias in diversebranchblock.py). This may help training-based pruning or quantization.

This is the PyTorch implementation. The MegEngine version is at https://github.com/megvii-model/DiverseBranchBlock

Paper: https://arxiv.org/abs/2103.13425

Update: released the code for building the block, transformations and verification.

Update: a more efficient implementation of BNAndPadLayer

Sometimes I call it ACNet v2 because 'DBB' is two bits larger than 'ACB' in ASCII. (lol)

We provide the trained models and a super simple PyTorch-official-example-style training script to reproduce the results.

Abstract

We propose a universal building block of Convolutional Neural Network (ConvNet) to improve the performance without any inference-time costs. The block is named Diverse Branch Block (DBB), which enhances the representational capacity of a single convolution by combining diverse branches of different scales and complexities to enrich the feature space, including sequences of convolutions, multi-scale convolutions, and average pooling. After training, a DBB can be equivalently converted into a single conv layer for deployment. Unlike the advancements of novel ConvNet architectures, DBB complicates the training-time microstructure while maintaining the macro architecture, so that it can be used as a drop-in replacement for regular conv layers of any architecture. In this way, the model can be trained to reach a higher level of performance and then transformed into the original inference-time structure for inference. DBB improves ConvNets on image classification (up to 1.9% higher top-1 accuracy on ImageNet), object detection and semantic segmentation.

image image image

Use our pretrained models

You may download the models reported in the paper from Google Drive (https://drive.google.com/drive/folders/1BPuqY_ktKz8LvHjFK5abD0qy3ESp8v6H?usp=sharing) or Baidu Cloud (https://pan.baidu.com/s/1wPaQnLKyNjF_bEMNRo4z6Q, the access code is "dbbk"). Currently only ResNet-18 models are available. The others will be released very soon. For the ease of transfer learning on other tasks, we provide both training-time and inference-time models. For ResNet-18 as an example, assume IMGNET_PATH is the path to your directory that contains the "train" and "val" directories of ImageNet, you may test the accuracy by running

python test.py IMGNET_PATH train ResNet-18_DBB_7101.pth -a ResNet-18 -t DBB

Here "train" indicates the training-time structure

Convert the training-time models into inference-time

You may convert a trained model into the inference-time structure with

python convert.py [weights file of the training-time model to load] [path to save] -a [architecture name]

For example,

python convert.py ResNet-18_DBB_7101.pth ResNet-18_DBB_7101_deploy.pth -a ResNet-18

Then you may test the inference-time model by

python test.py IMGNET_PATH deploy ResNet-18_DBB_7101_deploy.pth -a ResNet-18 -t DBB

Note that the argument "deploy" builds an inference-time model.

ImageNet training

The multi-processing training script in this repo is based on the official PyTorch example for the simplicity and better readability. The modifications include the model-building part and cosine learning rate scheduler. You may train and test like this:

python train.py -a ResNet-18 -t DBB --dist-url tcp://127.0.0.1:23333 --dist-backend nccl --multiprocessing-distributed --world-size 1 --rank 0 --workers 64 IMGNET_PATH
python test.py IMGNET_PATH train model_best.pth.tar -a ResNet-18

Use like this in your own code

Assume your model is like

class SomeModel(nn.Module):
    def __init__(self, ...):
        ...
        self.some_conv = nn.Conv2d(...)
        self.some_bn = nn.BatchNorm2d(...)
        ...
        
    def forward(self, inputs):
        out = ...
        out = self.some_bn(self.some_conv(out))
        ...

For training, just use DiverseBranchBlock to replace the conv-BN. Then SomeModel will be like

class SomeModel(nn.Module):
    def __init__(self, ...):
        ...
        self.some_dbb = DiverseBranchBlock(..., deploy=False)
        ...
        
    def forward(self, inputs):
        out = ...
        out = self.some_dbb(out)
        ...

Train the model just like you train the other regular models. Then call switch_to_deploy of every DiverseBranchBlock, test, and save.

model = SomeModel(...)
train(model)
for m in train_model.modules():
    if hasattr(m, 'switch_to_deploy'):
        m.switch_to_deploy()
test(model)
save(model)

FAQs

Q: Is the inference-time model's output the same as the training-time model?

A: Yes. You can verify that by

python dbb_verify.py

Q: What is the relationship between DBB and RepVGG?

A: RepVGG is a plain architecture, and the RepVGG-style structural re-param is designed for the plain architecture. On a non-plain architecture, a RepVGG block shows no superiority compared to a single 3x3 conv (it improves Res-50 by only 0.03%, as reported in the RepVGG paper). DBB is a universal building block that can be used on numerous architectures.

Q: How to quantize a model with DBB?

A1: Post-training quantization. After training and conversion, you may quantize the converted model with any post-training quantization method. Then you may insert a BN after the conv converted from a DBB and finetune to recover the accuracy just like you quantize and finetune the other models. This is the recommended solution.

A2: Quantization-aware training. During the quantization-aware training, instead of constraining the params in a single kernel (e.g., making every param in {-127, -126, .., 126, 127} for int8) for an ordinary conv, you should constrain the equivalent kernel of a DBB (get_equivalent_kernel_bias()).

Q: I tried to finetune your model with multiple GPUs but got an error. Why are the names of params like "xxxx.weight" in the downloaded weight file but sometimes like "module.xxxx.weight" (shown by nn.Module.named_parameters()) in my model?

A: DistributedDataParallel may prefix "module." to the name of params and cause a mismatch when loading weights by name. The simplest solution is to load the weights (model.load_state_dict(...)) before DistributedDataParallel(model). Otherwise, you may insert "module." before the names like this

checkpoint = torch.load(...)    # This is just a name-value dict
ckpt = {('module.' + k) : v for k, v in checkpoint.items()}
model.load_state_dict(ckpt)

Likewise, if the param names in the checkpoint file start with "module." but those in your model do not, you may strip the names like

ckpt = {k.replace('module.', ''):v for k,v in checkpoint.items()}   # strip the names
model.load_state_dict(ckpt)

Q: So a DBB derives the equivalent KxK kernels before each forwarding to save computations?

A: No! More precisely, we do the conversion only once right after training. Then the training-time model can be discarded, and every resultant block is just a KxK conv. We only save and use the resultant model.

Contact

[email protected]

Google Scholar Profile: https://scholar.google.com/citations?user=CIjw0KoAAAAJ&hl=en

My open-sourced papers and repos:

Simple and powerful VGG-style ConvNet architecture (preprint, 2021): RepVGG: Making VGG-style ConvNets Great Again (https://github.com/DingXiaoH/RepVGG)

State-of-the-art channel pruning (preprint, 2020): Lossless CNN Channel Pruning via Decoupling Remembering and Forgetting (https://github.com/DingXiaoH/ResRep)

CNN component (ICCV 2019): ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks (https://github.com/DingXiaoH/ACNet)

Channel pruning (CVPR 2019): Centripetal SGD for Pruning Very Deep Convolutional Networks with Complicated Structure (https://github.com/DingXiaoH/Centripetal-SGD)

Channel pruning (ICML 2019): Approximated Oracle Filter Pruning for Destructive CNN Width Optimization (https://github.com/DingXiaoH/AOFP)

Unstructured pruning (NeurIPS 2019): Global Sparse Momentum SGD for Pruning Very Deep Neural Networks (https://github.com/DingXiaoH/GSM-SGD)

Single-Shot Motion Completion with Transformer

Single-Shot Motion Completion with Transformer 👉 [Preprint] 👈 Abstract Motion completion is a challenging and long-discussed problem, which is of gr

FuxiCV 78 Dec 29, 2022
Codebase of deep learning models for inferring stability of mRNA molecules

Kaggle OpenVaccine Models Codebase of deep learning models for inferring stability of mRNA molecules, corresponding to the Kaggle Open Vaccine Challen

Eternagame 40 Dec 29, 2022
TensorFlow implementation of PHM (Parameterization of Hypercomplex Multiplication)

Parameterization of Hypercomplex Multiplications (PHM) This repository contains the TensorFlow implementation of PHM (Parameterization of Hypercomplex

Aston Zhang 9 Oct 26, 2022
Memory Efficient Attention (O(sqrt(n)) for Jax and PyTorch

Memory Efficient Attention This is unofficial implementation of Self-attention Does Not Need O(n^2) Memory for Jax and PyTorch. Implementation is almo

Amin Rezaei 126 Dec 27, 2022
Code for our paper Domain Adaptive Semantic Segmentation with Self-Supervised Depth Estimation

CorDA Code for our paper Domain Adaptive Semantic Segmentation with Self-Supervised Depth Estimation Prerequisite Please create and activate the follo

Qin Wang 60 Nov 30, 2022
Code for Referring Image Segmentation via Cross-Modal Progressive Comprehension, CVPR2020.

CMPC-Refseg Code of our CVPR 2020 paper Referring Image Segmentation via Cross-Modal Progressive Comprehension. Shaofei Huang*, Tianrui Hui*, Si Liu,

spyflying 55 Dec 01, 2022
OpenMMLab Text Detection, Recognition and Understanding Toolbox

Introduction English | 简体中文 MMOCR is an open-source toolbox based on PyTorch and mmdetection for text detection, text recognition, and the correspondi

OpenMMLab 3k Jan 07, 2023
PROJECT - Az Residential Real Estate Analysis

AZ RESIDENTIAL REAL ESTATE ANALYSIS -Decided on libraries to import. Includes pa

2 Jul 05, 2022
Code for our TKDE paper "Understanding WeChat User Preferences and “Wow” Diffusion"

wechat-wow-analysis Understanding WeChat User Preferences and “Wow” Diffusion. Fanjin Zhang, Jie Tang, Xueyi Liu, Zhenyu Hou, Yuxiao Dong, Jing Zhang,

18 Sep 16, 2022
Official code of ICCV2021 paper "Residual Attention: A Simple but Effective Method for Multi-Label Recognition"

CSRA This is the official code of ICCV 2021 paper: Residual Attention: A Simple But Effective Method for Multi-Label Recoginition Demo, Train and Vali

163 Dec 22, 2022
SSL_SLAM2: Lightweight 3-D Localization and Mapping for Solid-State LiDAR (mapping and localization separated) ICRA 2021

SSL_SLAM2 Lightweight 3-D Localization and Mapping for Solid-State LiDAR (Intel Realsense L515 as an example) This repo is an extension work of SSL_SL

Wang Han 王晗 1.3k Jan 08, 2023
[cvpr22] Perturbed and Strict Mean Teachers for Semi-supervised Semantic Segmentation

PS-MT [cvpr22] Perturbed and Strict Mean Teachers for Semi-supervised Semantic Segmentation by Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasile

Yuyuan Liu 132 Jan 03, 2023
Hybrid CenterNet - Hybrid-supervised object detection / Weakly semi-supervised object detection

Hybrid-Supervised Object Detection System Object detection system trained by hybrid-supervision/weakly semi-supervision (HSOD/WSSOD): This project is

5 Dec 10, 2022
Implementing yolov4 target detection and tracking based on nao robot

Implementing yolov4 target detection and tracking based on nao robot

6 Apr 19, 2022
Official code for article "Expression is enough: Improving traffic signal control with advanced traffic state representation"

1 Introduction Official code for article "Expression is enough: Improving traffic signal control with advanced traffic state representation". The code s

Liang Zhang 10 Dec 10, 2022
Quick program made to generate alpha and delta tables for Hidden Markov Models

HMM_Calc Functions for generating Alpha and Delta tables from a Hidden Markov Model. Parameters: a: Matrix of transition probabilities. a[i][j] = a_{i

Adem Odza 1 Dec 04, 2021
Deep Crop Rotation

Deep Crop Rotation Paper (to come very soon!) We propose a deep learning approach to modelling both inter- and intra-annual patterns for parcel classi

Félix Quinton 5 Sep 23, 2022
The official PyTorch code for NeurIPS 2021 ML4AD Paper, "Does Thermal data make the detection systems more reliable?"

MultiModal-Collaborative (MMC) Learning Framework for integrating RGB and Thermal spectral modalities This is the official code for NeurIPS 2021 Machi

NeurAI 12 Nov 02, 2022
tensorrt int8 量化yolov5 4.0 onnx模型

onnx模型转换为 int8 tensorrt引擎

123 Dec 28, 2022
Neural Turing Machine (NTM) & Differentiable Neural Computer (DNC) with pytorch & visdom

Neural Turing Machine (NTM) & Differentiable Neural Computer (DNC) with pytorch & visdom Sample on-line plotting while training(avg loss)/testing(writ

Jingwei Zhang 269 Nov 15, 2022