Skip to content

Latest commit

 

History

History
24 lines (19 loc) · 1.18 KB

README.md

File metadata and controls

24 lines (19 loc) · 1.18 KB
Raw

SqueezeNet Keras Implementation

This is the Keras implementation of SqueezeNet using functional API (arXiv 1602.07360). SqueezeNet is a small model of AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size. The original model was implemented in caffe.

Reference

pysqueezenet by yhenon

Differences:

  • Switch from Graph model to Keras 1.0 functional API
  • Fix the bug of pooling layer
  • Many thanks to StefOe, the source can now support Keras 2.0 API.

Result

This repository contains only the Keras implementation of the model, for other parameters used, please see the demo script, squeezenet_demo.py in the simdat package.

The training process uses a total of 2,600 images with 1,300 images per class (so, total two classes only). There are a total 130 images used for validation. After 20 epochs, the model achieves the following:

loss: 0.6563 - acc: 0.7065 - val_loss: 0.6247 - val_acc: 0.8750

Model Visualization