From 58bdb908f99f2f0d2c1c3f5c6ef7e19a043008b4 Mon Sep 17 00:00:00 2001
From: PENG Bo <33809201+BlinkDL@users.noreply.github.com>
Date: Wed, 22 Sep 2021 20:32:41 +0800
Subject: [PATCH] Update README.md

---
 README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index ed556aa..d83c904 100644
--- a/README.md
+++ b/README.md
@@ -68,7 +68,7 @@ We also propose a new sampling method called top-a (as in src/utils.py):
 
 (2) Remove all entries whose probability is lower than 0.02 * pow(p_max, 2). So it's adaptive, hence "top-a".
 
-(3) Feel free to tune the 0.02 and 2 factor.
+(3) Feel free to tune the 0.02 and 2 factor. Tune 0.02 first.
 
 The idea of top-a:
 1. If max_prob=0.9, then remove all tokens with prob < 0.0162 (so, removing most alternatives)
@@ -78,7 +78,7 @@ The idea of top-a:
 ```
 probs = F.softmax(logits, dim=-1)
 
-limit = torch.pow(torch.max(probs), 2.0) * 0.02
+limit = torch.pow(torch.max(probs), 2) * 0.02
 logits[probs < limit] = -float('Inf')
 ```