From cc870df7b8a4029ced7037500e9b253fec11cc10 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 4 Jun 2021 14:23:34 -0700 Subject: [PATCH] Update README.md --- timm/bits/README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/timm/bits/README.md b/timm/bits/README.md index eeb8297e..f17418ac 100644 --- a/timm/bits/README.md +++ b/timm/bits/README.md @@ -14,7 +14,7 @@ The current train.py and validate.py scripts are evolving to use the timm.bits c `timm` models will always be useable in pure PyTorch w/o `bits` or anything besides the utils / helpers for pretrained models, feature extraction, default data config. I may breakout bits into a diff project if there is any interest besides my own use for timm image and video model training. The layers: -* Device - DeviceEnv dataclass abstraction deals with PyTorch CPU, GPU and XLA device differences, incl distributed helpers, wrappers, etc. There is more than a passing similarity to HuggingFace Accelerate, but developed in parallel and with some difference in the detail. +* DeviceEnv - DeviceEnv dataclass abstraction deals with PyTorch CPU, GPU and XLA device differences, incl distributed helpers, wrappers, etc. There is more than a passing similarity to HuggingFace Accelerate, but developed in parallel and with some difference in the detail. * Updater - Dataclass that combines the backward pass, optimizer step, grad scaling, grad accumulation is a possibly device specific abstraction. * Currently basic single optimizer, single forward/backward Updaters are included for GPU, XLA. * Deepseed will need its own Updater(s) since its Engine is a monolith of epic proportions that breaks all separations of concern in PyTorch (UGH!). NOTE Deepspeed not working yet nor is it a priority. @@ -113,4 +113,4 @@ If you find bugs (there are likely many), feel free to file an issue with `[BITS # Acknowledgements -The TPU-VMs I've used for creating and testing this code, and that I hope to use for many future `timm` models were made available by the TPU Research Cloud (https://sites.research.google/trc/). \ No newline at end of file +The TPU-VMs I've used for creating and testing this code, and that I hope to use for many future `timm` models were made available by the TPU Research Cloud (https://sites.research.google/trc/).