MosaicML Composer

Open-source PyTorch library for efficient neural network training with built-in optimizations

AI Infrastructure & MLOpsFree

MosaicML Composer (now part of Databricks) is an open-source PyTorch library that makes neural network training dramatically more efficient by implementing dozens of state-of-the-art training methods as modular algorithms that can be combined and applied to any training run. These methods including gradient clipping, progressive resizing, and selective backpropagation reduce training time and cost without sacrificing model quality. ML researchers and engineers training deep learning models use Composer to apply best practices from the training optimization literature without implementing them from scratch.

Key Features

• ✓Training optimizations
• ✓Modular algorithms
• ✓PyTorch integration
• ✓Cost reduction
• ✓Open-source