SqueezeNet

Tensorflow implementation of SqueezeNet following this paper. Referenced Caffe code can be found here.

Prerequisites

• Python 3.3+
• Tensorflow
• pillow (PIL)
• (Optional) Tiny ImageNet Database: Tiny ImageNet Database (200 classes)

Implemented Models

SqueezeNet V0

• Use of 8 Fire Modules.
• Starting convolution filter of kernel size = 7 and output_filters = 96
• Pool Layers used after conv1, fire4 and fire8.

SqueezeNet V0 Residual (Using Bypass Connections)

• Uses bypass connections between layers.
• Bypass connections between fire2 and fire4, fire6 and pool2, fire8 and fire6, cxonv10 and pool3.

SqueezeNet V1

• Model defined in official repository
• Starting convolution filter of kernel size = 3 and output filters = 64
• Pool layers used after conv1, fire3 and fire5.

License

This project is licensed under the MIT License - see the LICENSE file for details

Author

Abhishek Tandon/ @Tandon-A