From 6a08df612feb44ba7b7bf1fd318a69e266f090e7 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Thu, 7 Nov 2024 21:45:46 -0800 Subject: [PATCH] Change eps defaults in adafactor_bv again after some checking --- timm/optim/adafactor_bv.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/timm/optim/adafactor_bv.py b/timm/optim/adafactor_bv.py index d603b528..bad982ee 100644 --- a/timm/optim/adafactor_bv.py +++ b/timm/optim/adafactor_bv.py @@ -214,8 +214,8 @@ def _single_tensor_adafactor( exp_avg = exp_avgs[i] step_t = state_steps[i] if eps is None: - # use square of machine eps for grad dtype if not set - eps = torch.finfo(grad.dtype).eps ** 2 + # default eps for avoiding div by zero, diff from float type eps + eps = 1e-7 if grad.dtype == torch.float16 else 1e-30 # Update step step_t += 1