BOHB-template

• The template of Bayesian optimization Hyperband (BOHB)

How to Run

python main.py

in 2021-10-22

• tf keras version (tunecallback) completed.
• tf gradienttape version (train_iteration) completed.
• pytorch version completed.

Quick Start

1. Prerequisite

pip install ray[tune] hpbandster ConfigSpace

2. How to max_t setting in scheduler

• The epoch of keras or torch is terminated based on max_t value.

Running example

+-----------------------+----------+-------+--------------+-----------+-----------+--------------+
| Trial name            | status   | loc   | activation   |   neuron1 |   neuron2 | optimizers   |
|-----------------------+----------+-------+--------------+-----------+-----------+--------------|
| objective_dd8c7_00000 | PENDING  |       | tanh         |        43 |        50 | adam         |
| objective_dd8c7_00001 | PENDING  |       | relu         |        63 |        45 | adam         |
| objective_dd8c7_00002 | PENDING  |       | tanh         |        63 |        52 | rmsprop      |
+-----------------------+----------+-------+--------------+-----------+-----------+--------------+

3. How to set stop condition

• It can be stopped according to the value of t in the stop item of tune.run.

For example,

stop={
"mean_accuracy": 0.99, # Terminate if accuracy is greater than 0.99
"training_iteration": 1
},

In this case, no matter how long max_t is, when all trials are running (just training iteration one), it is considered as ends.

Number of trials: 3/3 (3 TERMINATED)
+-----------------------+------------+-------+--------------+-----------+-----------+--------------+----------+--------+------------------+
| Trial name            | status     | loc   | activation   |   neuron1 |   neuron2 | optimizers   |      acc |   iter |   total time (s) |
|-----------------------+------------+-------+--------------+-----------+-----------+--------------+----------+--------+------------------|
| objective_dd8c7_00000 | TERMINATED |       | tanh         |        43 |        50 | adam         | 0.938278 |      1 |          5.3518  |
| objective_dd8c7_00001 | TERMINATED |       | relu         |        63 |        45 | adam         | 0.942333 |      1 |          5.15036 |
| objective_dd8c7_00002 | TERMINATED |       | tanh         |        63 |        52 | rmsprop      | 0.94     |      1 |          4.28608 |
+-----------------------+------------+-------+--------------+-----------+-----------+--------------+----------+--------+------------------+

4. Visualization

• Records are accumulated in the specified path during execution, which can be checked with the tensorboard command.

tensorboard --logdir bohb_results/