nobodyinamerica/squeezenet-chainer

SqueezeNet

An implementation of SqueezeNet in chainer.

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

Model	Size (mb)	Parameters (million)	CPU prediction (ms)	GPU prediction (ms)	Accuracy
ZFNet	117	16.42
SqueezeNet	4.6	1.277

Details

• ZFNet instead of AlexNet (I already had a trained ZFNet model). Initial filters are 7x7 instead of 11x11
• I had to add a dense layer to the end of SqueezeNet to get correct shape for my labels.
• This is the basic squeezenet without deep compression
• Paper says images are 224x224 but code and parameters suggests they use 227x227
• Seeing an 18x size reduction over ZFNet (SqueezeNet size is the same as the paper) and an 13x reduction in params.