shinjungwoo / aresb Goto Github PK

AresB-Net: Accurate Residual Binarized Neural Networks using Shortcut Concatenation and Shuffled Grouped Convolution

In summary, AresB-Net is very simple and easy for stacking pyramid structure.
Besides, even though there is no 1x1 convolution in stride for low-hardware overhead,
we have achieved remarkable evaluation results on CIFAR and ImageNet datasets without any specific initial weights and training techniques; 
we only focus on the model structure for apple-to-apple comparisons.  
Therefore, we can assure that there is additional margin to enhance the classification results.


This foloder contains supplementary materials of the evaluations for proving AresB-Net on CIFAR-10, CIFAR-10, and Imagenet dataset.
The following folders contains as:

- cifar10: evaluation of AresB-Net (no dropout before final FC layer)
- cifar100: evaluation of AresB-Net (dropout before final FC layer)
- imageet: evaluation of AresB-Net (no dropout before final FC layer)

In each folder, "aresb.sh" file contains the command for training and inference including batch size and weight decay. 
You can select either of one by commenting out other command.

In "train/{batch_size}/pretrained" folder, the pretrained checkpoints are stored, which can be used in the inference. 
Note: due to the storage limitation, only AresB-18 pretrained file is included.
-- Due to the limitation of Github storage, 
the following link is used to download the pretrained models:
CIFAR10 : https://drive.google.com/drive/folders/16dDp2RmmaZW9Fdzc2AmgG16lC1G4Yu8g?usp=sharing

CIFAR100: https://drive.google.com/drive/folders/16puLT1Wre0XJ8L8tYmhdKSbXX0ftDdY8?usp=sharing

ImageNet: https://drive.google.com/drive/folders/173Dt0GPFbwRgzeTI66gURPDB7PmZcLIb?usp=sharing

In "output/{batch_size}/pretrained" folder, the training and inference outputs for each epoch are listed.
If you run any command, the outputs are shown in "output/{batch_size}" folder.
The model file is contained in "models" folder. 

The outputs are as follows:

CIFAR10 - AresB-10: 90.74@Top-1
          AresB-18: 91.80@Top-1
          AresB-34: 92.71@Top-1 

CIFAR100 - AresB-10: 69.45@Top-1 91.70@Top-5
           AresB-18: 73.01@Top-1 92.57@Top-5
           AresB-34: 74.73@Top-1 93.25@Top-5

ImageNet - AresB-10: 48.51@Top-1 71.72@Top-5
           AresB-18: 54.81@Top-1 78.15@Top-5
           AresB-34: 58.46@Top-1 81.22@Top-5

Due to the limitation of file size, we do not include CIFAR and ImageNet datasets.
If you run the experiments on CIFAR dataset, the datasets can be automatically downloaded.
On the other hand, on ImageNet dataset, lmdb-formatted library is required.