Add the documentation for simple-pbt

examples/v1beta1/trial-images/simple-pbt/README.md

@@ -0,0 +1,27 @@
+# Toy PBT problem for benchmarking adaptive learning rate
+
+The goal is to optimize this trainable's accuracy. The accuracy increases
+fastest at the optimal lr, which is a function of the current accuracy.
+The optimal lr schedule for this problem is the triangle wave as follows.
+Note that many lr schedules for real models also follow this shape:
+
+```
+ best lr
+  ^
+  |    /\
+  |   /  \
+  |  /    \
+  | /      \
+  ------------> accuracy
+```
+
+In this problem, using PBT with a population of 2-4 is sufficient to
+roughly approximate this lr schedule. Higher population sizes will yield
+faster convergence. Training will not converge without PBT.
+
+If you want read more about this example, vist the 
+[ray](https://github.com/ray-project/ray/blob/7f1bacc7dc9caf6d0ec042e39499bbf1d9a7d065/python/ray/tune/examples/README.rst) 
+documentation.
+
+Katib uses this training container in some Experiments, for instance in the
+[PBT example](../../hp-tuning/simple-pbt.yaml#L44-L52).

examples/v1beta1/trial-images/simple-pbt/pbt_test.py

@@ -1,3 +1,7 @@
+#!/usr/bin/env python
+# Implementation based on:
+#   https://github.com/ray-project/ray/blob/7f1bacc7dc9caf6d0ec042e39499bbf1d9a7d065/python/ray/tune/examples/pbt_example.py
+
 # Copyright 2022 The Kubeflow Authors.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -12,11 +16,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-#!/usr/bin/env python
-
-# Implementation based on:
-#   https://github.com/ray-project/ray/blob/master/python/ray/tune/examples/pbt_example.py
-
 import argparse
 import numpy as np
 import os