HyperBoard is a simple visualization tool to facilitate hyperparameter tuning for Deep Learning players. It helps you to
- train on a remote server and visualize training curves on the local browser
- update curves in real time
- interactively compare hundreds of training curves for hyperparameter tuning, with filtering and visibility control
- save hundreds of training records on disk and re-load them when needed
HyperBoard resembles Tensorboard and Tensorboard for MXNet. However, HyperBoard is independent from any specific Deep Learning platform and provides extra functions.
Watch a 3-minute demonstration on Youtube or Bilibili.
Screenshot:
# install dependencies
pip install flask requests numpy
pip install flask-httpauth # depend on flask
git clone https://github.com/WarBean/hyperboard.git
cd hyperboard/
python setup.py installHyperBoard is easy to use. It is composed of three parts: Server, Agent and Dashboard. In general, when playing with deep neural networks on some remote server, you can
- launch HyperBoard Server on the same remote server
- set up your new hyperparameters, then call HyperBoard Agent to register some curves at the HyperBoard Server for this time
- start model training and for each K iteration, call HyperBoard Agent to send cross entropy, accuracy, BELU, etc to HyperBoard Server
- open HyperBoard Dashboard on your local browser, watch the curves growing
You can also run HyperBoard Server on your local PC. Currently, HyperBoard Server and Agent have been tested on Mac OS and Ubuntu. HyperBoard Dashboard has been tested on Firefox, Chrome and Safari.
Here are the details:
cd ~
mkdir my_recordscd my_records
hyperboard-run --user <your_user_name> --password <your_password>You can use HyperBoard without security authentication by simply using hyperboard-run without any arguments. Also use hyperboard-run -h for more help info.
For this part, you are recommended to run a simulation script example/run_agent.py first. Detailed usage of HyperBoard Agent will be explained below.
If hyperboard-run is run with argument --local, open http://127.0.0.1:5000, otherwise http://<your_remote_server_address>:5000.
The dashboard provides convenient interactive controls on visualization. Feel free to play with them - except the Delete from Server button :).
from hyperboard import Agent
agent = Agent(username = '', password = '', address = '127.0.0.1', port = 5000)username and password can be omitted if you have launched HyperBoard Server without security authentication. address and port can be omitted to use default value.
name = agent.register(hyperparameters, metric, overwrite = False)hyperparameters should be a dictionary.
metric is a string label indicating the numerical range of the values you send next. Curves sharing the same metric will share one y-axis on the dashboard. It helps you properly visualize cross entropy and accuracy in the same graph.
overwrite = False lets Agent to ask you for confirmation before removing existing records with the same hyperparameter setup.
agent.append(name, index, value)Each curve is saved as a file with a suffix .record, in the very directory where you launch HyperBoard Server. The content of record file is simple:
$ head my_records/fd8e3e27e4ef661488932e9a58197d90.record
{"batch size": 256, "corpus": "PennTreeBank", "criteria": "train accu", "learning rate": 0.001, "momentum": 0.9, "num hidden": 300, "optimizer": "Adam"}
accuracy
0 -0.22079159783278235
1 -0.15177436116678278
2 -0.0847468825330926
3 -0.009928149024110766
4 0.07511021349995883
5 0.16286792223048174
6 0.22981841687923243
7 0.25812625639630005The first line is hyperparameters. The second line is metric. Each line below contains the iteration index and the criteria value.
The next time you launch HyperBoard Server, it will reload all .record files (except those you delete manually) into memory.