SqueezeNet

An implementation of SqueezeNet in chainer.

Below are some benchmarks on a ImageNet-like dataset (1 million 255x255 images with 128 batch size)

Details

• ZFNet instead of AlexNet (I already had a trained ZFNet model).

• This is the basic squeezenet without deep compression. I had to add a dense layer to the end of SqueezeNet to get correct shape for my labels. I also had to add batch normalisation to the fire modules so it would fit.

• The paper says images are 224x224 but the code and parameters suggests they use 227x227. I also added some padding, relu and initialisation that was used in the squeezenet code but not mentioned in the paper.

• Accuracy of squeezenet is lower in this test but there is a 25x size reduction over ZFNet (SqueezeNet size is the same as in the paper) and a 12x reduction in params.