composer

composer copied to clipboard

Dose composer support best checkpoint saver which can monitoring the checkpoint for best metrics or losses?

20

## 🚀 Feature Request From the Documentation in Checkpointing part, I didn't find there is a best checkpoint saver function supported. best checkpoint saver refers to the function that monitors...

mayujie

Composer container image doesn't show exceptions in logs when Trainer crashes

3

** Environment ** Collecting system information... --------------------------------- System Environment Report Created: 2023-05-24 19:24:21 UTC --------------------------------- PyTorch information ------------------- PyTorch version: 1.13.1+cu117 Is debug build: False CUDA used to build PyTorch:...

jwatte

Enhance `label_smoothing` algorithm

2

# What does this PR do? Enhance Composer's `label_smoothing` by implementing the following label smoothers: `UniformSmoother`, `CategoricalSmoother` (new), and `OnlineSmoother` (new). * `UniformSmoother` smooths labels according to a uniform distribution....

ez2rok

Slow startup and OOMs when using device_train_microbatch_size with torch.compile

1

When I combine `device_train_microbatch_size="auto"` with a model that is compiled with `torch.compile(...)`, I get bad behavior in a couple predictable dimensions: 1. Because each separate batch size requires re-compiling the...

tbenthompson

medical image segmentation ipynb sample fix

2

# What does this PR do? This PR change the location of the SIIM dataset and makes some edits to the preprocessing of the dataset that may have changed between...

ashishdatta

Medical Image Segmentation Example Not Working

1

** Environment ** PyTorch information ------------------- PyTorch version: 1.13.1+cu116 Is debug build: False CUDA used to build PyTorch: 11.6 ROCM used to build PyTorch: N/A OS: Ubuntu 20.04.5 LTS (x86_64)...

ashishdatta

num_checkpoints_to_keep of CheckpointSaver doesn't work when the training is resumed

4

** Environment ** PyTorch information ------------------- PyTorch version: 1.13.1+cu117 Is debug build: False CUDA used to build PyTorch: 11.7 ROCM used to build PyTorch: N/A OS: Ubuntu 18.04.6 LTS (x86_64)...

RolandGao

Distillation Algorithm

2

# What does this PR do? This PR adds a distillation algorithm to Composer. We found that distillation is a [consistent efficiency improvement for vision and MLM models](https://arxiv.org/abs/2211.00683). This implementation...

growlix

# What does this PR do? Previously we installed a new system Python and overrode the original. This caused issues with several packages such as `PyPi` and `setuptools` due to...

bandish-shah

Support for TorchEval for metrics usage

3

## 🚀 Feature Request ## Motivation TorchEval is a newly released library from PyTorch for common evaluation metrics and tools: https://pytorch.org/torcheval/main/ The MosaicML framework currently has a deep integration with...

ananthsub