pytorch-storm
PyTorch implementation of a first-order stochastic trust-region method (STORM).
STORM1 exploits a specific optimizer, storm1.py, and an ad-hoc scheduling of the learning rate.
To train ResNet-18 on CIFAR-10 using STORM
cd pytorch-storm
python train.py -f 7800 --lr 1
Results
| Method | Test Acc. (%) |
|---|---|
| SGD + CosineAnnealingLR | 92.66 |
| STORM (ours) | 93.1 |