yanchunhui / simplenet Goto Github PK

This repository contains the architectures, Models, logs, etc pertaining to the SimpleNet Paper (Lets keep it simple: Using simple architectures to outperform deeper architectures ) : https://arxiv.org/abs/1608.06037
(Check the successor of this architecture at Towards Principled Design of Deep Convolutional Networks: Introducing SimpNet)

SimpleNet-V1 outperforms deeper and heavier architectures such as AlexNet, VGGNet,ResNet,GoogleNet,etc in a series of benchmark datasets, such as CIFAR10/100, MNIST, SVHN. It also achievs a higher accuracy (currently 67.17/87.44) in imagenet, more than AlexNet, NIN, Squeezenet, etc with only 5.7M parameters. Slimer versions of the architecture work very decently against more complex architectures such as ResNet and WRN as well.

If you find SimpleNet useful in your research, please consider citing:

@article{hasanpour2016lets,
  title={Lets keep it simple, Using simple architectures to outperform deeper and more complex architectures},
  author={Hasanpour, Seyyed Hossein and Rouhani, Mohammad and Fayyaz, Mohsen and Sabokrou, Mohammad},
  journal={arXiv preprint arXiv:1608.06037},
  year={2016}
}

Pytorch : For using Pytorch implemnetation click Pytorch implementation

ImageNet result was achieved using simple SGD without hyper parameter tuning for 100 epochs(single crop). no multicrop techniques were used. no dense evaluation or combinations of such techniques were used unlike all other architectures. the models will be uploaded when the training is finished.

• Achieved using Pytorch implementation

*Note that the Fractional max pooling[13] uses deeper architectures and also uses extreme data augmentation.۞ means No zero-padding or normalization with dropout and ۩ means Standard data-augmentation- with dropout. To our knowledge, our architecture has the state of the art result, without aforementioned data-augmentations.

*Note that we didn’t intend on achieving the state of the art performance here as we are using a single optimization policy without fine-tuning hyper parameters or data-augmentation for a specific task, and still we nearly achieved state-of-the-art on MNIST. **Results achieved using an ensemble or extreme data-augmentation

Table 6-Slimmed version Results on Different Datasets

*Since we presented their results in their respective sections, we avoided mentioning the results here again.

** Achieved using several data-augmentation tricks

Flops and Parameter Comparison: [tab:Flops_appndx]

Flops and Parameter Comparison of Models trained on ImageNet

*Inception v3, v4 did not have any Caffe model, so we reported their size related information from MXNet and Tensorflow respectively. Inception-ResNet-V2 would take 60 days of training with 2 Titan X to achieve the reported accuracy. Statistics are obtained using http://dgschwend.github.io/netscope

1# Data-augmentation method used by stochastic depth paper: https://github.com/Pastromhaug/caffe-stochastic-depth.

2# https://github.com/dmlc/mxnet-model-gallery/blob/master/imagenet-1k-inception-v3.md

3# https://github.com/tensorflow/models/tree/master/slim

4# https://github.com/revilokeb/inception\_resnetv2\_caffe