Permalink
370774c Nov 29, 2017
4 contributors

Users who have contributed to this file

@richardartoul @andrewschreiber @spence @esc
154 lines (119 sloc) 7.16 KB

Hyperdash Python SDK

CircleCI

Hyperdash is a machine learning monitoring library capable of running alongside Tensorflow, Scikit-Learn, and other modeling libraries. It was developed with a focus on enabling fast knowledge gain.

Use Hyperdash if you're looking for cloud-based model monitoring that:

  • Is fast and easy-to-use with scripts and Jupyter.
  • Tracks your hyperparameters across different model experiments.
  • Graphs performance metrics (loss, reward, etc.) in real-time.
  • Can be viewed remotely on web, iOS, and Android without self-hosting (e.g. Tensorboard).
  • Saves your experiment's print output (standard out / error) as a local log file.
  • Notifies you when a long-running experiment has finished.

Hyperdash is compatible with: Python 2.7-3.6

Installation

Foreword: We care deeply about making Hyperdash fast and easy to install on Linux, Mac, and Windows. If you find a snag along the way, please let us know at support@hyperdash.io!

Install Hyperdash in terminal from pip. Installing within a python virtual environment such as virtualenv is recommended. If you are having trouble installing via pip, a virtual environment will usually fix the problem.

$ pip install --upgrade pip && pip install hyperdash

By installing via pip, you can call hyperdash from the command line via both hyperdash and hd.

# Login if you have an account
$ hyperdash login

# Or signup free with an email
$ hd signup

After login or signup, an API key is saved to your local machine for automatic authorization. If you'd rather manage your API key manually, then review the "API Key Storage" section below.

You're ready to use Hyperdash! Make sure Hyperdash works by running:

$ hd demo

Learn Hyperdash in 60 seconds

Pure logging

If all you need is logging and notifications, simply prefix any terminal command:

hd run -n "Hotdog CNN" python hotdog.py

Or use pipe:

./catsdogs.sh | hd pipe -n "Hotdog CNN"

In Jupyter:

screen shot 2017-09-24 at 7 27 37 pm

Experiment instrumentation

If you are interested in tracking hyperparameters and performance metrics, you'll want to use the Experiment api. Experiment objects are created with a model name, then auto-started and auto-incremented. By default, Experiment will record print logs. Here is an example of a simple Scikit Learn classifier instrumented:

# digits.py
from sklearn import svm, datasets
from hyperdash import Experiment

# Preprocess data
digits = datasets.load_digits()
test_cases = 50
X_train, y_train = digits.data[:-test_cases], digits.target[:-test_cases]
X_test, y_test = digits.data[-test_cases:], digits.target[-test_cases:]

# Create an experiment with a model name, then autostart
exp = Experiment("Digits Classifier")
# Record the value of hyperparameter gamma for this experiment
gamma = exp.param("gamma", 0.1)
# Param can record any basic type (Number, Boolean, String)

classifer = svm.SVC(gamma=gamma)
classifer.fit(X_train, y_train)

# Record a numerical performance metric
exp.metric("accuracy", classifer.score(X_test, y_test))

# Cleanup and mark that the experiment successfully completed
exp.end()

Hyperparameters and metrics are pretty printed for your logs and reference:

{ gamma     : 0.001 }
| accuracy  : 1.000 |
Experiment "digits-classifier_2017-09-20t18-50-55-258215" complete.
Logs are available locally at: /Users/username/.hyperdash/logs/digits-classifier/digits-classifier_2017-09-20t18-50-55-258215.log

You can also disable logging by setting capture_io to false:

exp = Experiment("Digits Classifier", capture_io=False)

You've learned Hyperdash!

Visualize your experiments in the Hyperdash web, iOS, and Android apps.


More information

  • Jupyter/IPython tips
  • Decorator Experiment API
  • API key management

IPython/Jupyter Notebook

Hyperdash in Jupyter

Note: by default all print statements will be redirected to the cell that creates the experiment object due to capturing Jupyter's stdio. Use exp = Experiment("model name", capture_io=False) for normal printing, but no logging.

The SDK currently doesn't support mid-experiment parameter redeclaration. Remember to end() your experiment before redeclaring exp.

Decorator experiment API

If your experiment is wrapped in a function, the decorator API saves you the trouble of having to remember to write exp.end(). The exact argument exp must be passed into the wrapped function to get access to the Experiment object.

from hyperdash import monitor

@monitor("dogs vs. cats")
def train_dogs_vs_cats(exp): # Get Experiment object as argument to function.
  lr = exp.param("learning rate", 0.005)
  model = Model(lr)
  model.fit()
  exp.metric(model.accuracy())

API Keys

Storage

If you signed up through the CLI, then your API key is already installed in hyperdash.json file in the home directory of your user.

You can alternatively override this API key with a hyperdash.json file in your local directory (so you can have different API keys for different projects) or with the HYPERDASH_API_KEY environment variable.

Finally, the monitor function accepts an api_key_getter argument that if passed in will be called everytime we need to retrieve your API key. Example:

# test_script.py

from hyperdash import monitor

# This function can do anything you want, as long as it returns a Hyperdash API key as a string
def get_api_key():
  return "super_secret_api_key"

@monitor("dogs vs. cats", api_key_getter=get_api_key)
def train_dogs_vs_cats():
  print("Epoch 1, accuracy: 50%")
  time.sleep(2)
  print("Epoch 2, accuracy: 75%")
  time.sleep(2)
  print("Epoch 3, accuracy: 100%")

Keep in mind that we will call your function on a regular basis while the job is running (currently about once every 5 minutes) so that we can support API key rotation for long-running jobs.

API Key Rotation

The Hyperdash library will try and load up your API key about once every 5 minutes. Generally speaking this isn't something you need to think about, but in the rare case that you need to rotate an API key without stopping a long-running job, you can just change the HYPERDASH_API_KEY environment variable or hyperdash.json file and the SDK will automatically pickup the new key within a few minutes.