This repository contains the Caffe implementation of SqueezeNext. Each directory contains the network definition along with the solver parameters used for training. For visualizing each architecture you can use netscope ( you need to copy the train_val.prototxt files). For details of each architecture please see this paper.
The corresponding trained caffemodel files are also available here. All networks were trained using IntelCaffe on 32 Intel KNightsLanding (KNL) using the same hyper-parameters. Fine-tuning the hyper-parameters for each model may lead to better results. Our goal has been to show the general trend but please contact us if you got new results.
Pretrained SqueezeNext models on ImageNet have been added to imgclsmob repository. In Pytorch you can use the following command to load the pretrained network:
from pytorchcv.model_provider import get_model as ptcv_get_model
net = ptcv_get_model(“sqnxt23_w1”, pretrained=True)
Other possible options for the model are "sqnxt23_w3d2", "sqnxt23_w2", "sqnxt23v5_w1", "sqnxt23v5_w3d2", "sqnxt23v5_w2". And this is the link for the corresponding network definition(s).
For TensorFlow instructions please see this link.
- For TensorFlow implementation, please see Timen's implementation
- For results on these datasets, please see luuuyi's repository