Pytorch implementation of SNAS (Caution : This is not official version and was not written by the author of the paper)
Python >= 3.6.5, PyTorch == 1.0
Cifar-10
I followed hyperparameters that were given in the paper.
However, there are several parameters that were not given in the paper.
Ex) Softmax Temperature , annealiation rate of the softmax temperature, parameters regarding the levels of resource constraints
bash scripts/main_constraint_new.sh (WITH resource contraint)
bash scripts/retrain.sh (WITH resource contraint)
Figure1 : Search Validation Accuracy
(Note : the model was not fully trained(<==>converged) due to the limited resources (E.g., GPU and TIME!!)Figure2 : Search Validation Accuracy of ENAS
Figure3 : Network Architecture of normal cell (left) and reduction cell (right)
Architecture | Accuracy | Params |
---|---|---|
SNAS | 96.27% | 2.9M |
ENAS | 97.01% | 4.6M |