README.md

The Caffe-compatible files that you are probably looking for:

SqueezeNet_v1.0/train_val.prototxt          #model architecture
SqueezeNet_v1.0/solver.prototxt             #additional training details (learning rate schedule, etc.)
SqueezeNet_v1.0/squeezenet_v1.0.caffemodel  #pretrained model parameters

If you find SqueezeNet useful in your research, please consider citing the SqueezeNet paper:

@article{SqueezeNet,
    Author = {Forrest N. Iandola and Matthew W. Moskewicz and Khalid Ashraf and Song Han and William J. Dally and Kurt Keutzer},
    Title = {SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <1MB model size},
    Journal = {arXiv:1602.07360},
    Year = {2016}
}

Helpful hints:

1. Getting the SqueezeNet model: git clone <this repo>. In this repository, we include Caffe-compatible files for the model architecture, the solver configuration, and the pretrained model (4.8MB uncompressed).

2. Batch size. For the SqueezeNet model in our paper, we used a batch size of 1024. If implemented naively on a single GPU, this may result in running out of memory. An effective workaround is to use hierarchical batching (sometimes called "delayed batching"). Caffe supports hierarchical batching by doing train_val.prototxt>batch_size training samples concurrently in memory. After solver.prototxt>iter_size iterations, the gradients are summed and the model is updated. Mathematically, the batch size is batch_size * iter_size. In the included prototxt files, we have set (batch_size=32, iter_size=32), but any combination of batch_size and iter_size that multiply to 1024 will produce eqivalent results. In fact, with the same random number generator seed, the model will be fully reproducable if trained multiple times. Finally, note that in Caffe iter_size is applied while training on the training set but not while testing on the test set.

3. Implementing Fire modules. In the paper, we describe the expand portion of the Fire layer as a collection of 1x1 and 3x3 filters. Caffe does not natively support a convolution layer that has multiple filter sizes. To work around this, we implement expand1x1 and expand3x3 layers and concatenate the results together in the channel dimension.