LR Schedulers
Added three learning rate scheduler:
- MultiStepLR
- ReduceLROnPlateau
- CosineAnnealingLR
Put test in test/extra, and can you add a couple integration tests that train models with these schedulers?
Done
Tests failing
Forgot to set the seed for tensors, should be good now
Why did it fail with SGD? Is there a reason or just flakiness?
There’s no reason in particular both optimizers works with the schedulers. Since for the testing we're training on MNIST it was easy for the model to reach convergence without a scheduler. So I tried multiple settings and Adam with those parameters worked more reliably and could pass the 3 categories of test (CPU / LLVM / Torch).
Less shaky testing could be done by training on ImageNet for example but I preferred to do something lightweight that didn't require lots of training time.
Cool, merged!