Conversation
|
Hi @DoubleNorth! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks! |
|
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks! |
|
Let's continue the discussion on the issue: pytorch/pytorch#160849 |
2 participants
Stochastic Gradient Descent (SGD) and SGD-like methods (e.g., Adam) are commonly used in PyTorch to train ML models. However, these methods rely on random data order to converge, which usually require a full data shuffle, leading to low I/O performance for disk-based storage.
We proposed a simple but novel two-level data shuffling strategy named CorgiPile (https://link.springer.com/article/10.1007/s00778-024-00845-0), which can avoid a full data shuffle while maintaining comparable convergence rate as if a full shuffle were performed. CorgiPile first samples and shuffles data at the block-level, and then shuffles data at the tuple-level within the sampled data blocks, i.e., firstly shuffling data blocks, and then merging sampled blocks in a small buffer, and finally shuffling tuples in the buffer for SGD. We have implemented CorgiPile inside PyTorch (https://github.com/DS3Lab/CorgiPile-PyTorch), and extensive experimental results show that our CorgiPile can achieve comparable convergence rate with the full-shuffle based SGD, and faster than PyTorch with full data shuffle.