Log, organize and parallelize hyperparameter search for Deep Learning experiments
Test tube is a python library to track and parallelize hyperparameter search for Deep Learning and ML experiments. It's framework agnostic and built on top of the python argparse API for ease of use.
pip install test_tube
Use Test Tube if you need to:
- Parallelize hyperparameter optimization (across multiple gpus or cpus).
- Parallelize hyperparameter optimization across HPC cluster using SLURM. Auto-starts continuation jobs when walltime approaches.
- Track multiple Experiments across models.
- Visualize experiments without uploading anywhere, logs store as csv files.
- Automatically track ALL parameters for a particular training run.
- Automatically snapshot your code for an experiment using git tags.
- Save progress images inline with training metrics.
- Not deal with allocating to the correct CPU or GPUs.
Compatible with:
- Python 2, 3
- Tensorflow
- Keras
- Pytorch
- Caffe, Caffe2
- Chainer
- MXNet
- Theano
- Scikit-learn
- Any python based ML or DL library
- Write tests for log.py
- Write tests for argparse_hopt.py
- Implement hyperband
- Implement SMAC
- Integrate tensorboardx
If you're a researcher, test-tube is highly encouraged as a way to post your paper's training logs to help add transparency and show others what you've tried that didn't work.
from test_tube.hpc import SlurmCluster
# hyperparameters is a test-tube hyper params object
hyperparams = args.parse()
# init cluster
cluster = SlurmCluster(
hyperparam_optimizer=hyperparams,
log_path='/path/to/log/results/to',
python_cmd='python3'
)
# let the cluster know where to email for a change in job status (ie: complete, fail, etc...)
cluster.notify_job_status(email='some@email.com', on_done=True, on_fail=True)
# set the job options. In this instance, we'll run 20 different models
# each with its own set of hyperparameters giving each one 1 GPU (ie: taking up 20 GPUs)
cluster.per_experiment_nb_gpus = 1
cluster.per_experiment_nb_nodes = 1
# run the models on the cluster
cluster.optimize_parallel_cluster_gpu(train, nb_trials=20, job_name='first_tt_batch', job_display_name='my_batch')
# we just ran 20 different hyperparameters on 20 GPUs in the HPC cluster!!
from test_tube import Experiment
exp = Experiment(name='dense_model', save_dir='../some/dir/')
exp.tag({'learning_rate': 0.002, 'nb_layers': 2})
for step in range(1, 10):
tng_err = 1.0 / step
exp.log({'tng_err': tng_err})
import pandas as pd
import matplotlib
# each experiment is saved to a metrics.csv file which can be imported anywhere
# images save to exp/version/images
df = pd.read_csv('../some/dir/test_tube_data/dense_model/version_0/metrics.csv')
df.tng_err.plot()
from test_tube import HyperOptArgumentParser
# subclass of argparse
parser = HyperOptArgumentParser(strategy='random_search')
parser.add_argument('--learning_rate', default=0.002, type=float, help='the learning rate')
# let's enable optimizing over the number of layers in the network
parser.opt_list('--nb_layers', default=2, type=int, tunable=True, options=[2, 4, 8])
# and tune the number of units in each layer
parser.opt_range('--neurons', default=50, type=int, tunable=True, low=100, high=800, nb_samples=10)
# compile (because it's argparse underneath)
hparams = parser.parse_args()
# optimize across 4 gpus
# use 2 gpus together and the other two separately
hparams.optimize_parallel_gpu(MyModel.fit, gpu_ids=['1', '2,3', '0'], nb_trials=192, nb_workers=4)
Or... across CPUs
hparams.optimize_parallel_cpu(MyModel.fit, nb_trials=192, nb_workers=12)
from test_tube import HyperOptArgumentParser
# subclass of argparse
parser = HyperOptArgumentParser(strategy='random_search')
parser.add_argument('--learning_rate', default=0.002, type=float, help='the learning rate')
# let's enable optimizing over the number of layers in the network
parser.opt_list('--nb_layers', default=2, type=int, tunable=True, options=[2, 4, 8])
# and tune the number of units in each layer
parser.opt_range('--neurons', default=50, type=int, tunable=True, low=100, high=800, nb_samples=10)
# compile (because it's argparse underneath)
hparams = parser.parse_args()
# run 20 trials of random search over the hyperparams
for hparam_trial in hparams.trials(20):
train_network(hparam_trial)
You can also optimize on a log scale to allow better search over magnitudes of hyperparameter values, with a chosen base (disabled by default). Keep in mind that the range you search over must be strictly positive.
from test_tube import HyperOptArgumentParser
# subclass of argparse
parser = HyperOptArgumentParser(strategy='random_search')
# Randomly searches over the (log-transformed) range [100,800).
parser.opt_range('--neurons', default=50, type=int, tunable=True, low=100, high=800, nb_samples=10, log_base=10)
# compile (because it's argparse underneath)
hparams = parser.parse_args()
# run 20 trials of random search over the hyperparams
for hparam_trial in hparams.trials(20):
train_network(hparam_trial)
import argparse
from test_tube import HyperOptArgumentParser
# these lines are equivalent
parser = argparse.ArgumentParser(description='Process some integers.')
parser = HyperOptArgumentParser(description='Process some integers.', strategy='grid_search')
# do normal argparse stuff
...
# name must have either jpg, png or jpeg in it
img = np.imread('a.jpg')
exp.log('test_jpg': img, 'val_err': 0.2)
# saves image to ../exp/version/media/test_0.jpg
# csv has file path to that image in that cell
- Hyperparameter optimization for PyTorch across 20 cluster GPUs
- Hyperparameter optimization across 20 cluster CPUs
- Experiments and hyperparameter optimization for tensorflow across 4 GPUs simultaneously
Feel free to fix bugs and make improvements! 1. Check out the current bugs here or feature requests. 2. To work on a bug or feature, head over to our project page and assign yourself the bug. 3. We'll add contributor names periodically as people improve the library!
To cite the framework use:
@misc{Falcon2017,
author = {Falcon, W.A.},
title = {Test Tube},
year = {2017},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/williamfalcon/test-tube}}
}
In addition to the terms outlined in the license, this software is U.S. Patent Pending.