From 0e8eabe89b03e520b2741956d99e180427bd1ec4 Mon Sep 17 00:00:00 2001 From: Less Wright Date: Sun, 16 May 2021 13:20:36 -0700 Subject: [PATCH] Update README.md --- README.md | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/README.md b/README.md index ce13920..52cdeb8 100644 --- a/README.md +++ b/README.md @@ -23,6 +23,16 @@ Net results: 14.15% greater accuracy with Ranger21 vs Adam, same training epochs ### Ranger21 Status:
+ May 16 - ImageNet training finished, finishing paper, updated Ranger21 with 1 off iteration fix and new show_schedule() feature:
+* ImageNet runs have finished and hope to have arxiv paper ready in next week or so. +* Big thanks to @zsgj-Xxx for finding that the warmup ends up with the lr being 1 iteration short. Have updated with fix. +* In order to make it easier to see the lr schedule, have added a new show_schedule() that will show a pyplot image directly, along with the start/max/min values for the schedule. This info was already there via the tracking_lr value, but now it's even easier to train and then make a single: + ~~~ + optimizer.show_schedule() + ~~~ + to quickly view the full schedule and key values. +![ranger21_show_schedule](https://user-images.githubusercontent.com/46302957/118411403-2c7a1c00-b649-11eb-9619-4e647320e682.JPG) + May 1 PM - Multiple ImageNet runs in progress, updated Ranger code checked in Have multiple ImageNet runs in progress to prep for a paper for Ranger21. The Base comparison is simply Adam on ImageNet and Ranger21 on ImageNet, with a ResNet50. Ranger21 started later but has already matched Adam with half the epochs...plan is to run to 60 epochs each.
![Ranger21_Adam_ImNet](https://user-images.githubusercontent.com/46302957/116797727-ded4af80-aa9d-11eb-81b8-317b1d638af2.JPG)