https://github.com/nutszebra/prelu_net

Implementation of PReLUNet by chainer (Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification: https://arxiv.org/abs/1502.01852)
https://github.com/nutszebra/prelu_net

Last synced: about 1 month ago
JSON representation

Implementation of PReLUNet by chainer (Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification: https://arxiv.org/abs/1502.01852)

• Host: GitHub
• URL: https://github.com/nutszebra/prelu_net
• Owner: nutszebra
• License: mit
• Created: 2016-12-16T15:00:59.000Z (almost 8 years ago)
• Default Branch: master
• Last Pushed: 2017-02-02T06:20:29.000Z (almost 8 years ago)
• Last Synced: 2024-08-01T22:49:53.768Z (4 months ago)
• Language: Python
• Homepage:
• Size: 54.7 KB
• Stars: 12
• Watchers: 2
• Forks: 2
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-image-classification - unofficial-chainer : https://github.com/nutszebra/prelu_net
• awesome-image-classification - unofficial-chainer : https://github.com/nutszebra/prelu_net

README

        
# What's this

Implementation of PReLUNet by chainer  

# Dependencies

    git clone https://github.com/nutszebra/prelu_net.git

    cd prelu_net

    git submodule init

    git submodule update

# How to run

    python main.py -g 0

# Details about my implementation

All hyperparameters and network architecture are the same as in [[1]][Paper] except for some parts.

* Data augmentation  

Train: Pictures are randomly resized in the range of [256, 512], then 224x224 patches are extracted randomly and are normalized locally. Horizontal flipping is applied with 0.5 probability.  

Test: Pictures are resized to 384x384, then they are normalized locally. Single image test is used to calculate total accuracy.  

* SPP net

Instead of spp, I use global average pooling.

* Learning rate schedule

Learning rate is divided by 10 at [150, 225] epoch. The total number of epochs is 300.

# Cifar10 result

| network                                                   | total accuracy (%) |

|:----------------------------------------------------------|-------------------:|

| my implementation(model A)                                | 94.98              |





# References

Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification [[1]][Paper]  

[paper]: https://arxiv.org/abs/1502.01852 "Paper"