MariusAnje / LaMCTS

The release codes of LA-MCTS with its application to Neural Architecture Search.

Testing LaNet

1. Download pre-trained checkpoint from here, and place and unzip it in the same folder.

2. Run the following command to test.

python test.py  --checkpoint  ./lanas_128_99.03 --layers 24 --init_ch 128 --arch='[2, 2, 0, 2, 1, 2, 0, 2, 2, 3, 2, 1, 2, 0, 0, 1, 1, 1, 2, 1, 1, 0, 3, 4, 3, 0, 3, 1]'

[2, 2, 0, 2, 1, 2, 0, 2, 2, 3, 2, 1, 2, 0, 0, 1, 1, 1, 2, 1, 1, 0, 3, 4, 3, 0, 3, 1] is the best network found during the search. The snapshot below shows the top performing architectures (bottom to top) found during the distributed search. You can get the whole trace from here.

Training LaNet

1. Install cutmix

pip install git+https://github.com/ildoonet/cutmix

2. run training with the following command.

mkdir checkpoints
python train.py --auxiliary --batch_size=32 --init_ch=128 --layer=24 --arch='[2, 2, 0, 2, 1, 2, 0, 2, 2, 3, 2, 1, 2, 0, 0, 1, 1, 1, 2, 1, 1, 0, 3, 4, 3, 0, 3, 1]' --model_ema --model-ema-decay 0.9999 --auto_augment --epochs 1500

• Training on the ImageNet

Please use the training pipeline from Pytorch-Image-Models. Here we describe the procedures to do so:

1. get the network from train.py, line 121
2. go to Pytorch-Image-Models
3. find pytorch-image-models/blob/master/timm/models/factory.py, replace line 57 as follows

# model = create_fn(**model_args, **kwargs) 
model = our-network

Our ImageNet pipeline will be released soon, stay tuned.