From 0b2ea43d5cfcd41b940358cd0500e43da7186041 Mon Sep 17 00:00:00 2001
From: Adrian Garcia Badaracco <1755071+adriangb@users.noreply.github.com>
Date: Wed, 2 Sep 2020 18:00:07 -0500
Subject: [PATCH] fix example

---
 rfcs/20200902-pickle-for-keras/README.md | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/rfcs/20200902-pickle-for-keras/README.md b/rfcs/20200902-pickle-for-keras/README.md
index 9e9d1c571..b25ef54ee 100644
--- a/rfcs/20200902-pickle-for-keras/README.md
+++ b/rfcs/20200902-pickle-for-keras/README.md
@@ -48,6 +48,8 @@ Examples that could be resolved using `Model.save` (but the user tried pickle fi
 
 * [SO #51878627](https://stackoverflow.com/questions/51878627/pickle-keras-ann)
 
+Here is an post describing how PyTorch ran into similar issues and how they resolved them: [link](https://matthewrocklin.com/blog/work/2018/07/23/protocols-pickle)
+
 ## User Benefit
 
 * Lessen the learning curve for new Keras/TF users since they will be able to
@@ -98,7 +100,9 @@ Solution (2) would look something like this (assuming `Model.save` worked with `
 class Model:
 
     def __reduce_ex__(self, protocol):
-        return tf.keras.models.load_model, (self.save(io.BytesIO()),)
+        b = io.BytesIO()
+        self.save(b)
+        return load_model, (b.get_value()),)  # where load_model is tf.keras.models.load_model
 ```
 
 By implementing this in all of Keras' base classes, things will automatically work