Skip to content
A neat pytorch implementation of nasnet and the ported weights from tensorflow
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
nasnet add files Mar 20, 2018
LICENSE Initial commit Mar 20, 2018 Update Jul 4, 2018 update the code for pytorch 0.4 or higher May 14, 2019 add files Mar 20, 2018 add files Mar 20, 2018

A neat pytorch implementation of NASNet

The performance of the ported models on ImageNet (Accuracy):

Model Checkpoint Million Parameters Val Top-1 Val Top-5
NASNet-A_Mobile_224 5.3 70.2 89.4
NASNet-A_large_331 88.9 82.3 96.0

The slight performance drop may be caused by the different spatial padding methods between tensorflow and pytorch.

The porting process is done by and, modified from Cadene's project. Note that NASNets with the original performance can be found there.

You can evaluate the models by running, e.g. evaluate the NASNet-A_Mobile_224 ported model by

python --nas-type mobile --resume /path/to/modelfile --gpus 0 --data /path/to/imagenet_root_dir

The ported model files are provided: NASNet-A_Mobile_224, NASNet-A_large_331.

Future work:

  • add drop path for training
  • more nasnet model settings
You can’t perform that action at this time.