Changelog

October 2022

• upgrade the PyTorch container to 22.10
• switched to using torchvision IMAGENET1K_V2 backbone weights
• added a flag to control for torchvision weight enums
• added a flag to control TF32 computations
• fixed various depreciation warnings
• set TORCH_CUDNN_V8_API_ENABLED environment variable which replaces CUDNN_V8_API_ENABLED from older containers
• updated nv-cocoapi from 0.6.0 to 0.7.3
• updated python dependencies

June 2022

• upgrade the PyTorch container to 22.05
• fixed DALI depreciation warnings

January 2022

• upgrade the PyTorch container to 22.01
• made AMP the default data precision
• added --data-layout option (channels_first is the recommended layout with --no-amp)
• updated README with new performance numbers

November 2021

• upgrade the PyTorch container to 21.11
• switched data layout from NCHW (channels first) to NHWC (channels last)
• replaced torch.distributed.launch with torchrun
• updated README with new performance numbers

May 2021

• upgrade the PyTorch container to 21.05
• replaced APEX AMP with native PyTorch AMP
• updated nv-cocoapi from 0.4.0 to 0.6.0
• code updated to use DALI 1.2.0

April 2021

• upgrade the PyTorch container to 21.04
• changed python package naming

March 2021

• upgrade the PyTorch container to 21.03
• code updated to use DALI 0.30.0
• use DALI BoxEncoder instead of a CUDA extension
• replaced cocoapi with nv-cocoapi

June 2020

• upgrade the PyTorch container to 20.06
• update performance tables to include A100 results
• update examples with A100 configs

August 2019

• upgrade the PyTorch container to 19.08
• update Results section in the README
• code updated to use DALI 0.12.0
• checkpoint loading fix
• fixed links in the README

July 2019

• script and notebook for inference
• use AMP instead of hand-crafted FP16 support
• README update
• introduced a parameter with a path to the custom backbone checkpoint
• minor enchantments of example/* scripts
• alignment to changes in PyTorch 19.06

March 2019

• Initial release