Update (Aug 17, 2021): refactored the code of ACB. The readability has been greatly improved. You may call switch_to_deploy of an ACB to convert it to the inference-time structure. If you use ACB in your own model, the conversion is as easy as
for m in your_model.modules():
if hasattr(m, 'switch_to_deploy'):
m.switch_to_deploy()
There is also some runnable code for testing the equivalence in the main function of acnet/acb.py. Just check it by
python acnet/acb.py
ACNet v2 (Diverse Branch Block, DBB): Diverse Branch Block: Building a Convolution as an Inception-like Unit.
DBB (CVPR 2021) is a CNN component with higher performance than ACB and still no inference-time costs. Sometimes I call it ACNet v2 because "DBB" is 2 bits larger than "ACB" in ASCII (lol).
I would suggest you check the repo of DBB (https://github.com/DingXiaoH/DiverseBranchBlock). It also has an implementation of ACNet.
News:
- Zhang et al. used our ACB in their model ACFD, which won the 1st place in IJCAI 2020 iCartoon Face Challenge (Detection Track). Congratulations!
- Liu et al. extended ACB to EACB (Enhanced Asym Conv Block) in their MMDM, which helped them won the 3rd place in NTIRE 2020 Challenge on Image Demoireing at CVPR 2020. Congratulations!
- MMDM also won the 4th place in NTIRE 2020 Challenge on Real Image Denoising at CVPR 2020. Congratulations again!
- ACNet has been used in several real business products.
- At ICCV 2019, I was told that ACNet improved the performance of some semantic segmentation tasks by 2%. So glad to hear that!
Update: Updated the whole repo, including ImageNet training (with Distributed Data Parallel). The default learning rate schedules were changed to cosine annealing, which performed better on ImageNet. Changed the behavior of ACB when k > 3. It used to add 1x3 and 3x1 kernels onto 5x5, but now it uses 1x5 and 5x1.
ICCV 2019 paper: ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks.
Other implementations:
- PaddlePaddle re-implementation for building ACNet and converting the weights has been accepted by PaddlePaddle official repo. Amazing work by @parap1uie-s!
- Tensorflow2: an easy plugin module (https://github.com/CXYCarson/TF_AcBlock)! Just use it to build your model and call deploy() to convert it into the inference-time structure! Amazing work by @CXYCarson!
This demo will show you how to
- Build an ACNet with Asymmetric Convolution Block. Just a few lines of code!
- Train the ACNet together with the regular CNN baseline with the same training configurations.
- Test the ACNet and the baseline, get the average accuracy.
- Convert the ACNet into exactly the same structure as the regular counterpart for deployment. Congratulations! The users of your model will be happy because they can enjoy higher accuracy with exactly the same computational burdens as the baseline trained with regular conv layers.
About the environment:
- We used torch==1.3.0, torchvision==0.4.1, CUDA==10.2, NVIDIA driver version==440.82, tensorboard==1.11.0 on a machine with eight 2080Ti GPUs.
- Our method does not rely on any new or deprecated features of any libraries, so there is no need to make an identical environment.
- If you get any errors regarding tensorboard or tensorflow, you may simply delete the code related to tensorboard or SummaryWriter.
Some results (Top-1 accuracy) reproduced on CIFAR-10 using the codes in this repository (note that we add batch norm for Cifar-quick and VGG baselines):
Model | Baseline | ACNet |
---|---|---|
Cifar-quick | 86.20 | 86.87 |
VGG | 93.99 | 94.54 |
ResNet-56 | 94.55 | 95.06 |
WRN-16-8 | 95.89 | 96.33 |
If it does not work on your specific model and dataset, based on my experience, I would suggest you
- try different learning rate schedules
- initialize the trained scaling factor of batch norm (e.g., gamma variable in Tensorflow and bn.weight in PyTorch) in the three branches of every ACB as 1/3. This improves the performance on CIFAR
The experiments reported in the paper were performed using Tensorflow. However, the backbone of the codes was refactored from the official Tensorflow benchmark (https://github.com/tensorflow/benchmarks/tree/master/scripts/tf_cnn_benchmarks), which was designed in the pursuit of extreme speed, not readability.
Citation:
@InProceedings{Ding_2019_ICCV,
author = {Ding, Xiaohan and Guo, Yuchen and Ding, Guiguang and Han, Jungong},
title = {ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}
As designing appropriate Convolutional Neural Network (CNN) architecture in the context of a given application usually involves heavy human works or numerous GPU hours, the research community is soliciting the architecture-neutral CNN structures, which can be easily plugged into multiple mature architectures to improve the performance on our real-world applications. We propose Asymmetric Convolution Block (ACB), an architecture-neutral structure as a CNN building block, which uses 1D asymmetric convolutions to strengthen the square convolution kernels. For an off-the-shelf architecture, we replace the standard square-kernel convolutional layers with ACBs to construct an Asymmetric Convolutional Network (ACNet), which can be trained to reach a higher level of accuracy. After training, we equivalently convert the ACNet into the same original architecture, thus requiring no extra computations anymore. We have observed that ACNet can improve the performance of various models on CIFAR and ImageNet by a clear margin. Through further experiments, we attribute the effectiveness of ACB to its capability of enhancing the model's robustness to rotational distortions and strengthening the central skeleton parts of square convolution kernels.
-
Enter this directory.
-
Make a soft link to your ImageNet directory, which contains "train" and "val" directories.
ln -s YOUR_PATH_TO_IMAGENET imagenet_data
- Set the environment variables. We use 8 GPUs with Distributed Data Parallel. Of course, 4 GPUs work as well.
export PYTHONPATH=.
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
- Train a ResNet-18 on ImageNet with Asymmetric Convolution Blocks. The code will automatically convert the trained weights to the original structure and test. The top-1 accuracy will be around 71.2%.
python -m torch.distributed.launch --nproc_per_node=8 acnet/do_acnet.py -a sres18 -b acb
- Check the shape of weights in the converted model.
python3 display_hdf5.py acnet_exps/sres18_acb_train/finish_deploy.hdf5
- Train a regular ResNet-18 on ImageNet as baseline for the comparison. The top-1 accuracy will be around 70.6%.
python -m torch.distributed.launch --nproc_per_node=8 acnet/do_acnet.py -a sres18 -b base
- ResNet-34 and ResNet-50 are also provided in acnet/do_acnet.py, please try as you wish.
python -m torch.distributed.launch --nproc_per_node=8 acnet/do_acnet.py -a sres34 -b acb
python -m torch.distributed.launch --nproc_per_node=8 acnet/do_acnet.py -a sres50 -b acb
-
Enter this directory.
-
Make a soft link to your CIFAR-10 directory. If the dataset is not found in the directory, it will be automatically downloaded.
ln -s YOUR_PATH_TO_CIFAR cifar10_data
- Set the environment variables.
export PYTHONPATH=.
export CUDA_VISIBLE_DEVICES=0
- Train the Cifar-quick ACNet. The code will automatically convert the trained weights to the original structure (acnet_exps/cfqkbnc_acb_train/finish_deploy.hdf5) and test. Then train a regular model as baseline for the comparison.
python acnet/do_acnet.py -a cfqkbnc -b acb
python acnet/do_acnet.py -a cfqkbnc -b base
- Do the same on VGG.
python acnet/do_acnet.py -a vc -b acb
python acnet/do_acnet.py -a vc -b base
- Do the same on ResNet-56.
python acnet/do_acnet.py -a src56 -b acb
python acnet/do_acnet.py -a src56 -b base
- Do the same on WRN-16-8.
python acnet/do_acnet.py -a wrnc16plain -b acb
python acnet/do_acnet.py -a wrnc16plain -b base
- Show the accuracy of all the models.
python show_log.py acnet_exps
[email protected] (The original Tsinghua mailbox [email protected] will expire in several months)
Google Scholar Profile: https://scholar.google.com/citations?user=CIjw0KoAAAAJ&hl=en
Homepage: https://dingxiaohan.xyz/
My open-sourced papers and repos:
The Structural Re-parameterization Universe:
-
RepLKNet (CVPR 2022) Powerful efficient architecture with very large kernels (31x31) and guidelines for using large kernels in model CNNs
Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs
code. -
RepOptimizer (ICLR 2023) uses Gradient Re-parameterization to train powerful models efficiently. The training-time RepOpt-VGG is as simple as the inference-time. It also addresses the problem of quantization.
Re-parameterizing Your Optimizers rather than Architectures
code. -
RepVGG (CVPR 2021) A super simple and powerful VGG-style ConvNet architecture. Up to 84.16% ImageNet top-1 accuracy!
RepVGG: Making VGG-style ConvNets Great Again
code. -
RepMLP (CVPR 2022) MLP-style building block and Architecture
RepMLPNet: Hierarchical Vision MLP with Re-parameterized Locality
code. -
ResRep (ICCV 2021) State-of-the-art channel pruning (Res50, 55% FLOPs reduction, 76.15% acc)
ResRep: Lossless CNN Pruning via Decoupling Remembering and Forgetting
code. -
ACB (ICCV 2019) is a CNN component without any inference-time costs. The first work of our Structural Re-parameterization Universe.
ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks.
code. -
DBB (CVPR 2021) is a CNN component with higher performance than ACB and still no inference-time costs. Sometimes I call it ACNet v2 because "DBB" is 2 bits larger than "ACB" in ASCII (lol).
Diverse Branch Block: Building a Convolution as an Inception-like Unit
code.
Model compression and acceleration:
-
(CVPR 2019) Channel pruning: Centripetal SGD for Pruning Very Deep Convolutional Networks with Complicated Structure
code -
(ICML 2019) Channel pruning: Approximated Oracle Filter Pruning for Destructive CNN Width Optimization
code -
(NeurIPS 2019) Unstructured pruning: Global Sparse Momentum SGD for Pruning Very Deep Neural Networks
code