Awesome
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 tensorflow_dump.py and pytorch_load.py, modified from Cadene's project. Note that NASNets with the original performance can be found there.
You can evaluate the models by running imagenet_eval.py, e.g. evaluate the NASNet-A_Mobile_224 ported model by
python imagenet_eval.py --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