My research experiments on speeding up the training stage of deep learning network SqueezeNet was successful. I improved SqueezeNet by simplifying its architecture Simplified_SqueezeNet.
The speedup gain was about 4 times faster than SqueezeNet. I conducted my experiments on two large scale image datasets. The used datasets are MIT Places205 and MIT Places2. Simplified_SqueezeNet got higher accuracy on both Top-1 and Top-5 on both datasets. Your suggestions and contributions to port the networks weights and architecture to PyTorch are highly appreciated. You can find the network details in this repo: https://github.com/NidabaSystems/Simplified_SqueezeNet