Lyken17/Efficient-PyTorch
A collection of best practices for efficiently training large neural network models using PyTorch.
Velocity · 7d
+0.4
★ / day
Trend
→steady
star history
This repository shares optimization techniques for PyTorch training pipelines, focusing on maximizing training throughput. It demonstrates achieving approximately 730 images per second when training ResNet-50 on ImageNet through various tuning tips and best practices. The benchmarks compare favorably against TensorFlow and MXNet implementations.