mirror of
https://github.com/wassname/Ranger21.git
synced 2026-09-09 11:15:43 +08:00
Update README.md
This commit is contained in:
@@ -29,10 +29,12 @@ Net results: 14.15% greater accuracy with Ranger21 vs Adam, same training epochs
|
||||
|
||||
</br>
|
||||
In addition, training a BN Free (no batch norm) ResNet50 as an additional comparison. Of interest, even after 4 restarts, Adam was unable to get more than 3 epochs in on the NormFree Resnet50. By comparison, Ranger21 is doing well so this already shows the improved resilience of training with Ranger21. </br>
|
||||

|
||||
|
||||

|
||||
|
||||
|
||||
* Ranger21 code updates - due to firsthand experience, have added in safety guards in the event that num_epochs does not match the actual epochs trained, as well as updated the linear warmdown code to be simpler and never go below the min_lr designated (defaults to 3e-5). This will start to spew a lot of text to alert you on each iteration, but the lr will be automatically guarded. </br>
|
||||

|
||||
 </br>
|
||||
|
||||
|
||||
<b> April 27 PM - Ranger21 now training on ImageNet!</b> Starting work on benchmarking Ranger21 on ImageNet. Due to cost, will train to 60 epochs on ImageNet and compare with same setup with 60 epochs using Adam to have a basic "gold standard" comparison. Training is underway now.
|
||||
|
||||
Reference in New Issue
Block a user