NNI provides state-of-the-art tuning algorithm in builtin-tuners. NNI supports to build a tuner by yourself for tuning demand.
NNI provides state-of-the-art tuning algorithm in builtin-tuners. NNI supports to build a tuner by yourself for tuning demand.
...
@@ -123,3 +123,68 @@ Write a more advanced automl algorithm
...
@@ -123,3 +123,68 @@ Write a more advanced automl algorithm
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
The methods above are usually enough to write a general tuner. However, users may also want more methods, for example, intermediate results, trials' state (e.g., the methods in assessor), in order to have a more powerful automl algorithm. Therefore, we have another concept called ``advisor`` which directly inherits from ``MsgDispatcherBase`` in :githublink:`msg_dispatcher_base.py <nni/runtime/msg_dispatcher_base.py>`. Please refer to `here <CustomizeAdvisor.rst>`__ for how to write a customized advisor.
The methods above are usually enough to write a general tuner. However, users may also want more methods, for example, intermediate results, trials' state (e.g., the methods in assessor), in order to have a more powerful automl algorithm. Therefore, we have another concept called ``advisor`` which directly inherits from ``MsgDispatcherBase`` in :githublink:`msg_dispatcher_base.py <nni/runtime/msg_dispatcher_base.py>`. Please refer to `here <CustomizeAdvisor.rst>`__ for how to write a customized advisor.
Customize Assessor
==================
NNI supports to build an assessor by yourself for tuning demand.
If you want to implement a customized Assessor, there are three things to do:
#. Inherit the base Assessor class
#. Implement assess_trial function
#. Configure your customized Assessor in experiment YAML config file
**1. Inherit the base Assessor class**
.. code-block:: python
from nni.assessor import Assessor
class CustomizedAssessor(Assessor):
def __init__(self, ...):
...
**2. Implement assess trial function**
.. code-block:: python
from nni.assessor import Assessor, AssessResult
class CustomizedAssessor(Assessor):
def __init__(self, ...):
...
def assess_trial(self, trial_history):
"""
Determines whether a trial should be killed. Must override.
trial_history: a list of intermediate result objects.
Returns AssessResult.Good or AssessResult.Bad.
"""
# you code implement here.
...
**3. Configure your customized Assessor in experiment YAML config file**
NNI needs to locate your customized Assessor class and instantiate the class, so you need to specify the location of the customized Assessor class and pass literal values as parameters to the __init__ constructor.
.. code-block:: yaml
assessor:
codeDir: /home/abc/myassessor
classFileName: my_customized_assessor.py
className: CustomizedAssessor
# Any parameter need to pass to your Assessor class __init__ constructor
# can be specified in this optional classArgs field, for example
classArgs:
arg1: value1
Please noted in **2**. The object ``trial_history`` are exact the object that Trial send to Assessor by using SDK ``report_intermediate_result`` function.
The working directory of your assessor is ``<home>/nni-experiments/<experiment_id>/log``\ , which can be retrieved with environment variable ``NNI_LOG_DIRECTORY``\ ,
When the results are parsed, the tuners are also ranked based on their final performance. The following three tables show
When the results are parsed, the tuners are also ranked based on their final performance. The following three tables show
...
@@ -154,52 +154,52 @@ To view the same data in another way, for each tuner, we present the average ran
...
@@ -154,52 +154,52 @@ To view the same data in another way, for each tuner, we present the average ran
Besides these reports, our script also generates two graphs for each fold of each task: one graph presents the best score received by each tuner until trial x, and another graph shows the score that each tuner receives in trial x. These two graphs can give some information regarding how the tuners are "converging" to their final solution. We found that for "nnismall", tuners on the random forest model with search space defined in ``/examples/trials/benchmarking/automlbenchmark/nni/extensions/NNI/architectures/run_random_forest.py`` generally converge to the final solution after 40 to 60 trials. As there are too much graphs to incldue in a single report (96 graphs in total), we only present 10 graphs here.
Besides these reports, our script also generates two graphs for each fold of each task: one graph presents the best score received by each tuner until trial x, and another graph shows the score that each tuner receives in trial x. These two graphs can give some information regarding how the tuners are "converging" to their final solution. We found that for "nnismall", tuners on the random forest model with search space defined in ``/examples/trials/benchmarking/automlbenchmark/nni/extensions/NNI/architectures/run_random_forest.py`` generally converge to the final solution after 40 to 60 trials. As there are too much graphs to incldue in a single report (96 graphs in total), we only present 10 graphs here.
The previous two graphs are generated for fold 1 of the task "car". In the first graph, we observe that most tuners find a relatively good solution within 40 trials. In this experiment, among all tuners, the DNGOTuner converges fastest to the best solution (within 10 trials). Its best score improved for three times in the entire experiment. In the second graph, we observe that most tuners have their score flucturate between 0.8 and 1 throughout the experiment. However, it seems that the Anneal tuner (green line) is more unstable (having more fluctuations) while the GPTuner has a more stable pattern. This may be interpreted as the Anneal tuner explores more aggressively than the GPTuner and thus its scores for different trials vary a lot. Regardless, although this pattern can to some extent hint a tuner's position on the explore-exploit tradeoff, it is not a comprehensive evaluation of a tuner's effectiveness.
The previous two graphs are generated for fold 1 of the task "car". In the first graph, we observe that most tuners find a relatively good solution within 40 trials. In this experiment, among all tuners, the DNGOTuner converges fastest to the best solution (within 10 trials). Its best score improved for three times in the entire experiment. In the second graph, we observe that most tuners have their score flucturate between 0.8 and 1 throughout the experiment. However, it seems that the Anneal tuner (green line) is more unstable (having more fluctuations) while the GPTuner has a more stable pattern. This may be interpreted as the Anneal tuner explores more aggressively than the GPTuner and thus its scores for different trials vary a lot. Regardless, although this pattern can to some extent hint a tuner's position on the explore-exploit tradeoff, it is not a comprehensive evaluation of a tuner's effectiveness.
NNI provides an easy way to adopt an approach to set up parameter tuning algorithms, we call them **Tuner**.
The tuner decides which hyperparameter sets will be evaluated. It is a most important part of NNI HPO.
Tuner receives metrics from `Trial` to evaluate the performance of a specific parameters/architecture configuration. Tuner sends the next hyper-parameter or architecture configuration to Trial.
A tuner works in following steps:
The following table briefly describes the built-in tuners provided by NNI. Click the **Tuner's name** to get the Tuner's installation requirements, suggested scenario, and an example configuration. A link for a detailed description of each algorithm is located at the end of the suggested scenario for each tuner. Here is an `article <../CommunitySharings/HpoComparison.rst>`__ comparing different Tuners on several problems.
1. Initialize with a search space.
2. Generate hyperparameter sets from the search space.
3. Send hyperparameters to trials.
4. Receive evaluation results.
5. Update internal states according to the results.
6. Go to step 2, until experiment end.
NNI has out-of-the-box support for many popular tuning algorithms.
They should be sufficient to cover most typical machine learning scenarios.
However, if you have a very specific demand, or if you have designed an algorithm yourself,
you can also implement your own tuner: :doc:`custom_algorithm`
Common Usage
------------
All built-in tuners have similar usage.
To use a built-in tuner, you need to specify its name and arguments in experiment config,
and provides a standard :doc:`search_space`.
Some tuners, like SMAC and DNGO, have extra dependencies that need to be installed separately.
Please check each tuner's reference page for what arguments it supports and whether it needs extra dependencies.
For a general example, random tuner can be configured as follow:
.. code-block:: python
config.search_space = {
'x': {'_type': 'uniform', '_value': [0, 1]}
}
config.tuner.name = 'Random'
config.tuner.class_args = {'seed': 0}
Full List
---------
.. list-table::
.. list-table::
:header-rows: 1
:header-rows: 1
...
@@ -14,61 +49,52 @@ The following table briefly describes the built-in tuners provided by NNI. Click
...
@@ -14,61 +49,52 @@ The following table briefly describes the built-in tuners provided by NNI. Click
- The Tree-structured Parzen Estimator (TPE) is a sequential model-based optimization (SMBO) approach. SMBO methods sequentially construct models to approximate the performance of hyperparameters based on historical measurements, and then subsequently choose new hyperparameters to test based on this model. `Reference Paper <https://papers.nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization.pdf>`__
- The Tree-structured Parzen Estimator (TPE) is a sequential model-based optimization (SMBO) approach. SMBO methods sequentially construct models to approximate the performance of hyperparameters based on historical measurements, and then subsequently choose new hyperparameters to test based on this model. `Reference Paper <https://papers.nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization.pdf>`__
- In Random Search for Hyper-Parameter Optimization show that Random Search might be surprisingly simple and effective. We suggest that we could use Random Search as the baseline when we have no knowledge about the prior distribution of hyper-parameters. `Reference Paper <http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf>`__
- In Random Search for Hyper-Parameter Optimization show that Random Search might be surprisingly simple and effective. We suggest that we could use Random Search as the baseline when we have no knowledge about the prior distribution of hyper-parameters. `Reference Paper <http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf>`__
- This simple annealing algorithm begins by sampling from the prior, but tends over time to sample from points closer and closer to the best ones observed. This algorithm is a simple variation on the random search that leverages smoothness in the response surface. The annealing rate is not adaptive.
- This simple annealing algorithm begins by sampling from the prior, but tends over time to sample from points closer and closer to the best ones observed. This algorithm is a simple variation on the random search that leverages smoothness in the response surface. The annealing rate is not adaptive.
- Naïve Evolution comes from Large-Scale Evolution of Image Classifiers. It randomly initializes a population-based on search space. For each generation, it chooses better ones and does some mutation (e.g., change a hyperparameter, add/remove one layer) on them to get the next generation. Naïve Evolution requires many trials to work, but it's very simple and easy to expand new features. `Reference paper <https://arxiv.org/pdf/1703.01041.pdf>`__
- Naive Evolution comes from Large-Scale Evolution of Image Classifiers. It randomly initializes a population-based on search space. For each generation, it chooses better ones and does some mutation (e.g., change a hyperparameter, add/remove one layer) on them to get the next generation. Naïve Evolution requires many trials to work, but it's very simple and easy to expand new features. `Reference paper <https://arxiv.org/pdf/1703.01041.pdf>`__
- SMAC is based on Sequential Model-Based Optimization (SMBO). It adapts the most prominent previously used model class (Gaussian stochastic process models) and introduces the model class of random forests to SMBO, in order to handle categorical parameters. The SMAC supported by NNI is a wrapper on the SMAC3 GitHub repo.
- SMAC is based on Sequential Model-Based Optimization (SMBO). It adapts the most prominent previously used model class (Gaussian stochastic process models) and introduces the model class of random forests to SMBO, in order to handle categorical parameters. The SMAC supported by NNI is a wrapper on the SMAC3 GitHub repo.
Notice, SMAC needs to be installed by ``pip install nni[SMAC]`` command. `Reference Paper, <https://www.cs.ubc.ca/~hutter/papers/10-TR-SMAC.pdf>`__ `GitHub Repo <https://github.com/automl/SMAC3>`__
Notice, SMAC needs to be installed by ``pip install nni[SMAC]`` command. `Reference Paper, <https://www.cs.ubc.ca/~hutter/papers/10-TR-SMAC.pdf>`__ `GitHub Repo <https://github.com/automl/SMAC3>`__
- Batch tuner allows users to simply provide several configurations (i.e., choices of hyper-parameters) for their trial code. After finishing all the configurations, the experiment is done. Batch tuner only supports the type choice in search space spec.
- Batch tuner allows users to simply provide several configurations (i.e., choices of hyper-parameters) for their trial code. After finishing all the configurations, the experiment is done. Batch tuner only supports the type choice in search space spec.
- Hyperband tries to use limited resources to explore as many configurations as possible and returns the most promising ones as a final result. The basic idea is to generate many configurations and run them for a small number of trials. The half least-promising configurations are thrown out, the remaining are further trained along with a selection of new configurations. The size of these populations is sensitive to resource constraints (e.g. allotted search time). `Reference Paper <https://arxiv.org/pdf/1603.06560.pdf>`__
- Hyperband tries to use limited resources to explore as many configurations as possible and returns the most promising ones as a final result. The basic idea is to generate many configurations and run them for a small number of trials. The half least-promising configurations are thrown out, the remaining are further trained along with a selection of new configurations. The size of these populations is sensitive to resource constraints (e.g. allotted search time). `Reference Paper <https://arxiv.org/pdf/1603.06560.pdf>`__
- Metis offers the following benefits when it comes to tuning parameters: While most tools only predict the optimal configuration, Metis gives you two outputs: (a) current prediction of optimal configuration, and (b) suggestion for the next trial. No more guesswork. While most tools assume training datasets do not have noisy data, Metis actually tells you if you need to re-sample a particular hyper-parameter. `Reference Paper <https://www.microsoft.com/en-us/research/publication/metis-robustly-tuning-tail-latencies-cloud-systems/>`__
- Metis offers the following benefits when it comes to tuning parameters: While most tools only predict the optimal configuration, Metis gives you two outputs: (a) current prediction of optimal configuration, and (b) suggestion for the next trial. No more guesswork. While most tools assume training datasets do not have noisy data, Metis actually tells you if you need to re-sample a particular hyper-parameter. `Reference Paper <https://www.microsoft.com/en-us/research/publication/metis-robustly-tuning-tail-latencies-cloud-systems/>`__
- BOHB is a follow-up work to Hyperband. It targets the weakness of Hyperband that new configurations are generated randomly without leveraging finished trials. For the name BOHB, HB means Hyperband, BO means Bayesian Optimization. BOHB leverages finished trials by building multiple TPE models, a proportion of new configurations are generated through these models. `Reference Paper <https://arxiv.org/abs/1807.01774>`__
- BOHB is a follow-up work to Hyperband. It targets the weakness of Hyperband that new configurations are generated randomly without leveraging finished trials. For the name BOHB, HB means Hyperband, BO means Bayesian Optimization. BOHB leverages finished trials by building multiple TPE models, a proportion of new configurations are generated through these models. `Reference Paper <https://arxiv.org/abs/1807.01774>`__
- Gaussian Process Tuner is a sequential model-based optimization (SMBO) approach with Gaussian Process as the surrogate. `Reference Paper <https://papers.nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization.pdf>`__, `Github Repo <https://github.com/fmfn/BayesianOptimization>`__
- Gaussian Process Tuner is a sequential model-based optimization (SMBO) approach with Gaussian Process as the surrogate. `Reference Paper <https://papers.nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization.pdf>`__, `Github Repo <https://github.com/fmfn/BayesianOptimization>`__
* - `PBT Tuner <./PBTTuner.rst>`__
* - `PBT <../autotune_ref.html>`_
- PBT Tuner is a simple asynchronous optimization algorithm which effectively utilizes a fixed computational budget to jointly optimize a population of models and their hyperparameters to maximize performance. `Reference Paper <https://arxiv.org/abs/1711.09846v1>`__
- PBT Tuner is a simple asynchronous optimization algorithm which effectively utilizes a fixed computational budget to jointly optimize a population of models and their hyperparameters to maximize performance. `Reference Paper <https://arxiv.org/abs/1711.09846v1>`__
* - `DNGO Tuner <./DngoTuner.rst>`__
* - `DNGO <../autotune_ref.html>`_
- Use of neural networks as an alternative to GPs to model distributions over functions in bayesian optimization.
- Use of neural networks as an alternative to GPs to model distributions over functions in bayesian optimization.
.. toctree::
Comparison
:maxdepth: 1
----------
TPE <Tuner/TpeTuner>
These articles have compared built-in tuners' performance on some different tasks:
