Used ResNet-18, pytorch on cifar10 dataset. Applied different Learning Rate Schedulers:
- Learning rate change on Platue.
- Cyclical LR, with small step size(2000).
- Cyclical LR, with large step size(10000 iterations = 25 epochs). Val Acc: 94.360% accuracy on validation data at 52nd epoch, taking a time of 35 min.
In the second case, it was possible to train the model to get 90% accuracy on validation dataset in 410 sec or 6.8 Min.
The aim eventually is to attain 94% which is human benchmark on cifar10, in under 2 min.