Skip to content

A Batch Size Scheduler library compatible with PyTorch DataLoaders.

License

Notifications You must be signed in to change notification settings

ancestor-mithril/bs-scheduler

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

42 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

bs-scheduler

A Batch Size Scheduler library compatible with PyTorch DataLoaders.


Documentation


Why use a Batch Size Scheduler?

  • Using a big batch size has several advantages:
    • Better hardware utilization.
    • Enhanced parallelism.
    • Faster training.
  • However, using a big batch size from the start may lead to a generalization gap.
  • Therefore, the solution is to gradually increase the batch size, similar to a learning rate decay policy.
  • See Don't Decay the Learning Rate, Increase the Batch Size.

Available Schedulers

Batch Size Schedulers

  1. LambdaBS - sets the batch size to the base batch size times a given lambda.
  2. MultiplicativeBS - sets the batch size to the current batch size times a given lambda.
  3. StepBS - multiplies the batch size with a given factor at a given number of steps.
  4. MultiStepBS - multiplies the batch size with a given factor each time a milestone is reached.
  5. ConstantBS - multiplies the batch size by a given factor once and decreases it again to its base value after a given number of steps.
  6. LinearBS - increases the batch size by a linearly changing multiplicative factor for a given number of steps.
  7. ExponentialBS - increases the batch size by a given $\gamma$ each step.
  8. PolynomialBS - increases the batch size using a polynomial function in a given number of steps.
  9. CosineAnnealingBS - increases the batch size to a maximum batch size and decreases it again following a cyclic cosine curve.
  10. IncreaseBSOnPlateau - increases the batch size each time a given metric has stopped improving for a given number of steps.
  11. CyclicBS - cycles the batch size between two boundaries with a constant frequency, while also scaling the distance between boundaries.
  12. CosineAnnealingBSWithWarmRestarts - increases the batch size to a maximum batch size following a cosine curve, then restarts while also scaling the number of iterations until the next restart.
  13. OneCycleBS - decreases the batch size to a minimum batch size then increases it to a given maximum batch size, following a linear or cosine annealing strategy.
  14. SequentialBS - calls a list of schedulers sequentially given a list of milestone points which reflect which scheduler should be called when.
  15. ChainedBSScheduler - chains a list of batch size schedulers and calls them together each step.

Installation

Please install PyTorch first before installing this repository.

pip install bs-scheduler

Licensing

The library is licensed under the BSD-3-Clause license.

Citation

To be added...