From ad1012bcdaa404907a2fecc7de48ea8b5eaf6b88 Mon Sep 17 00:00:00 2001
From: songqun <songqun001@gmail.com>
Date: Wed, 28 Apr 2021 17:29:42 +0800
Subject: [PATCH] add comment for alpha beta in hardswish and hardsigmoid
 compared to tf/pytorch implementation (#2859)

---
 src/layer/hardsigmoid.cpp | 2 ++
 src/layer/hardswish.cpp   | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/src/layer/hardsigmoid.cpp b/src/layer/hardsigmoid.cpp
index fe02f30f8..060b08e63 100644
--- a/src/layer/hardsigmoid.cpp
+++ b/src/layer/hardsigmoid.cpp
@@ -24,6 +24,8 @@ HardSigmoid::HardSigmoid()
 
 int HardSigmoid::load_param(const ParamDict& pd)
 {
+    // tensorflow uses alpha,beta = 0.2, 0.5
+    // pytorch uses alpha,beta = 1/6, 0.5
     alpha = pd.get(0, 0.2f);
     beta = pd.get(1, 0.5f);
     lower = -beta / alpha;
diff --git a/src/layer/hardswish.cpp b/src/layer/hardswish.cpp
index fb064f43a..2bcc17d82 100644
--- a/src/layer/hardswish.cpp
+++ b/src/layer/hardswish.cpp
@@ -24,6 +24,8 @@ HardSwish::HardSwish()
 
 int HardSwish::load_param(const ParamDict& pd)
 {
+    // Note that tensorflow/pytorch use alpha,beta = 1/6, 0.5, not the default value here.
+    // You can setup them manually in .param file.
     alpha = pd.get(0, 0.2f);
     beta = pd.get(1, 0.5f);
     lower = -beta / alpha;