ReImplementation of Residual Attention Network for Image Classification

This is a Gluon implementation of the residual attention network in the paper 1704.06904.

Requirement

python3.5+, mxnet-1.2.1+, MXBoard

Inspiration

The code is inspired by the gluon resnet implementation and https://github.com/liudaizong/Residual-Attention-Network.

Mixup data augmentation use GluonCV for reference.

Train

GPU is preferred.

Cifar10

To view the training process, tensorboard is required.

tensorboard --logdir=./log/board/cifar10_201808311834 --host=0.0.0.0 --port=8888

Results	Accuracy	Loss	Test Accuracy	Test Accuracy(using mixup)	Official Report
Attention56			0.9499	0.9581	0.9448
Attention92			0.9524	0.9608	0.9501

The author does not give the architecture of cifar10-AttentionNet, I follow the implementation of https://github.com/tengshaofeng/ResidualAttentionNetwork-pytorch.
In previous version, the feature map is down sampled to 16x16 before stage1, it can only achieve about 0.93 on test set. Following Teng's implementation, the feature map size is still 32x32, and it gets an accuracy improvement of 2%.

Tricks
I followed the tricks in 1812.01187, the pipeline is training res-att-net56 on 8 1080ti GPUs with nvidia-dali and BytePS, which allow to finish in 1 hour for cifar10.

The result below can be reproduced by running scripts/train_bps.sh.

Layers	batch_size	lr	warmup	mix_up	alpha	epsilon	max_accuracy
56	1024	2.0	5	False	0.2	-	0.956631
56	1024	1.5	5	False	0.2	-	0.955929
56	1024	1.0	5	False	0.2	-	0.954527
56	1024	2.0	5	True	0.2	-	0.959135
56	1024	2.0	5	True	0.4	-	0.962240
56	1024	2.0	5	True	1.0	-	0.962941
56	1024	2.0	5	True	0.2	0.1	0.961839
56	1024	2.0	5	True	0.2	0.01	0.961038
56	1024	2.0	10	True	0.2	0.1	0.959635
56	1024	2.0	10	True	1.0	0.1	0.963742
56	1024	1.6	10	True	1.0	-	0.965044
56	1024	2.0	10	True	0.2	0.01	0.959235
56	512	2.0	5	False	0.2	-	0.955729
56	512	1.5	5	False	0.2	-	0.956831
56	512	1.0	5	False	0.2	-	0.955812
56	512	2.0	10	True	0.2	0.1	0.958534
56	512	2.0	10	True	1.0	0.1	0.957833
56	512	2.0	10	True	0.2	0.01	0.955829
56	512	1.0	5	True	0.4	-	0.962182
56	512	1.0	5	True	1.0	-	0.964670
56	512	1.0	5	True	1.0	0.1	0.963642
56	512	1.0	5	True	1.0	0.01	0.963674
56	512	1.0	10	True	1.0	0.1	0.962139
56	512	0.8	10	True	1.0	0.1	0.964271
56	512	0.4	10	True	1.0	0.1	0.963442

Note:

• alpha control how we generate the mix up data and label.
• epsilon is related to label smoothing.

ImageNet

Emmmm....

TODO

• Training scripts for cifar10.
  Just use this command and you can get accuracy over 0.95 on cifar10:

python3 cifar10_train.py --num-layers 92 --num-gpus 1 --workers 2 --batch-size 64 --epochs 200 --lr-steps 80,120 --mix-up 1 --alpha 1.0

It can be easily applied to other tasks.

• Attention-56,92,128,164,236,452 support
  The hyper-parameters are based on paper section 4.1. The number of layers can be calculated by 36m+20 where m is the number of Attention Module in each stage when p, t, r = 1, 2, 1.

attention_net_spec = {56: ([1, 2, 1], [1, 1, 1]),
                      92: ([1, 2, 1], [1, 2, 3]),
                      128: ([1, 2, 1], [3, 3, 3]),
                      164: ([1, 2, 1], [4, 4, 4]),
                      236: ([1, 2, 1], [6, 6, 6]),
                      452: ([2, 4, 3], [6, 6, 6])}

• Visualization of soft attention mask
• Attention module with other basic network unit
  ...

References

1. Residual Attention Network for Image Classification 1704.06904
2. MXNet Documentation and Tutorials zh.gluon.ai/
3. GluonCV Classification gluon-cv.mxnet.io/model_zoo/classification.html