Add Metis Tuner (#534)

* update readme in ga_squad

* update readme

* fix typo

* Update README.md

* Update README.md

* Update README.md

* update readme

* update

* fix path

* update reference

* fix bug in config file

* update nni_arch_overview.png

* update

* update

* update

* add metis tuner code

* 1. fix bug about import 2.update other sdk file

* add auto-gbdt-example and remove unused code

* add metis_tuner into README

* update the README

* update README | remove unused variable

* fix typo

* add sklearn into requirments

* Update src/sdk/pynni/nni/metis_tuner/metis_tuner.py

add default value in __init__

Co-Authored-By: xuehui1991 <[email protected]>

* Update docs/HowToChooseTuner.md

Co-Authored-By: xuehui1991 <[email protected]>

* Update docs/HowToChooseTuner.md

Co-Authored-By: xuehui1991 <[email protected]>

* fix typo | add more comments

docs/HowToChooseTuner.md

@@ -11,6 +11,8 @@ For now, NNI has supported the following tuner algorithms. Note that NNI install
  - [Grid Search](#Grid)
  - [Hyperband](#Hyperband)
  - [Network Morphism](#NetworkMorphism) (require pyTorch)
+ - [Metis Tuner](#MetisTuner) (require sklearn)
+
 
  ## Supported tuner algorithms
 
@@ -178,7 +180,7 @@ _Usage_:
 <a name="NetworkMorphism"></a>
 **Network Morphism**
 
-[Network Morphism](7) provides functions to automatically search for architecture of deep learning models. Every child network inherits the knowledge from its parent network and morphs into diverse types of networks, including changes of depth, width and skip-connection. Next, it estimates the value of child network using the history architecture and metric pairs. Then it selects the most promising one to train. More detail can be referred to [here](../src/sdk/pynni/nni/networkmorphism_tuner/README.md). 
+[Network Morphism][7] provides functions to automatically search for architecture of deep learning models. Every child network inherits the knowledge from its parent network and morphs into diverse types of networks, including changes of depth, width and skip-connection. Next, it estimates the value of child network using the history architecture and metric pairs. Then it selects the most promising one to train. More detail can be referred to [here](../src/sdk/pynni/nni/networkmorphism_tuner/README.md). 
 
 _Installation_: 
 NetworkMorphism requires [pyTorch](https://pytorch.org/get-started/locally), so users should install it first.
@@ -205,6 +207,43 @@ _Usage_:
 ```
 
 
+<a name="MetisTuner"></a>
+**Metis Tuner**
+
+[Metis][10] offers the following benefits when it comes to tuning parameters:
+While most tools only predicts the optimal configuration, Metis gives you two outputs: (a) current prediction of optimal configuration, and (b) suggestion for the next trial. No more guess work!
+
+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.
+
+While most tools have problems of being exploitation-heavy, Metis' search strategy balances exploration, exploitation, and (optional) re-sampling.
+
+Metis belongs to the class of sequential model-based optimization (SMBO), and it is based on the Bayesian Optimization framework. To model the parameter-vs-performance space, Metis uses both Gaussian Process and GMM. Since each trial can impose a high time cost, Metis heavily trades inference computations with naive trial. At each iteration, Metis does two tasks:
+*  It finds the global optimal point in the Gaussian Process space. This point represents the optimal configuration.
+* It identifies the next hyper-parameter candidate. This is achieved by inferring the potential information gain of exploration, exploitation, and re-sampling.
+
+Note that the only acceptable types of search space are `choice`, `quniform`, `uniform` and `randint`.
+
+More details can be found in our paper: https://www.microsoft.com/en-us/research/publication/metis-robustly-tuning-tail-latencies-cloud-systems/
+
+
+_Installation_: 
+Metis Tuner requires [sklearn](https://scikit-learn.org/), so users should install it first. User could use `pip3 install sklearn` to install it.
+
+
+_Suggested scenario_:
+Similar to TPE and SMAC, Metis is a black-box tuner. If your system takes a long time to finish each trial, Metis is more favorable than other approaches such as random search. Furthermore, Metis provides guidance on the subsequent trial. Here is an [example](../examples/trials/auto-gbdt/search_space_metis.json) about the use of Metis. User only need to send the final result like `accuracy` to tuner, by calling the nni SDK.
+
+_Usage_:
+```yaml
+  # config.yaml
+  tuner:
+    builtinTunerName: MetisTuner
+    classArgs:
+      #choice: maximize, minimize
+      optimize_mode: maximize
+```
+
+
 # How to use Assessor that NNI supports?
 
 For now, NNI has supported the following assessor algorithms.
@@ -273,3 +312,4 @@ _Usage_:
 [7]: https://arxiv.org/abs/1806.10282
 [8]: https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/46180.pdf
 [9]: http://aad.informatik.uni-freiburg.de/papers/15-IJCAI-Extrapolation_of_Learning_Curves.pdf
+[10]:https://www.microsoft.com/en-us/research/publication/metis-robustly-tuning-tail-latencies-cloud-systems/ 

examples/trials/auto-gbdt/config_metis.yml

@@ -0,0 +1,21 @@
+authorName: default
+experimentName: example_auto-gbdt-metis
+trialConcurrency: 1
+maxExecDuration: 10h
+maxTrialNum: 10
+#choice: local, remote, pai
+trainingServicePlatform: local
+searchSpacePath: search_space_metis.json
+#choice: true, false
+useAnnotation: false
+tuner:
+  #choice: TPE, Random, Anneal, Evolution, BatchTuner
+  #SMAC (SMAC should be installed through nnictl)
+  builtinTunerName: MetisTuner
+  classArgs:
+    #choice: maximize, minimize
+    optimize_mode: minimize
+trial:
+  command: python3 main.py
+  codeDir: .
+  gpuNum: 0

examples/trials/auto-gbdt/requirments.txt

@@ -1 +1,2 @@
-pip install lightgbm
+lightgbm
+pandas

examples/trials/auto-gbdt/search_space_metis.json

@@ -0,0 +1,5 @@
+{
+    "num_leaves":{"_type":"choice","_value":[31, 28, 24, 20]},
+    "learning_rate":{"_type":"choice","_value":[0.01, 0.05, 0.1, 0.2]},
+    "bagging_freq":{"_type":"choice","_value":[1, 2, 4, 8, 10]}
+}

src/nni_manager/rest_server/restValidationSchemas.ts

@@ -148,7 +148,7 @@ export namespace ValidationSchemas {
                 checkpointDir: joi.string().allow('')
             }),
             tuner: joi.object({
-                builtinTunerName: joi.string().valid('TPE', 'Random', 'Anneal', 'Evolution', 'SMAC', 'BatchTuner', 'GridSearch', 'NetworkMorphism'),
+                builtinTunerName: joi.string().valid('TPE', 'Random', 'Anneal', 'Evolution', 'SMAC', 'BatchTuner', 'GridSearch', 'NetworkMorphism', 'MetisTuner'),
                 codeDir: joi.string(),
                 classFileName: joi.string(),
                 className: joi.string(),

src/sdk/pynni/nni/constants.py

@@ -28,7 +28,8 @@
     'Medianstop': 'nni.medianstop_assessor.medianstop_assessor',
     'GridSearch': 'nni.gridsearch_tuner.gridsearch_tuner',
     'NetworkMorphism': 'nni.networkmorphism_tuner.networkmorphism_tuner',
-    'Curvefitting': 'nni.curvefitting_assessor.curvefitting_assessor'
+    'Curvefitting': 'nni.curvefitting_assessor.curvefitting_assessor',
+    'MetisTuner': 'nni.metis_tuner.metis_tuner'
 }
 
 ClassName = {
@@ -40,6 +41,7 @@
     'BatchTuner': 'BatchTuner',
     'GridSearch': 'GridSearchTuner',
     'NetworkMorphism':'NetworkMorphismTuner',
+    'MetisTuner':'MetisTuner',
 
     'Medianstop': 'MedianstopAssessor',
     'Curvefitting': 'CurvefittingAssessor'

src/sdk/pynni/nni/metis_tuner/Regression_GMM/CreateModel.py

@@ -0,0 +1,58 @@
+# Copyright (c) Microsoft Corporation
+# All rights reserved.
+#
+# MIT License
+#
+# Permission is hereby granted, free of charge,
+# to any person obtaining a copy of this software and associated
+# documentation files (the "Software"), to deal in the Software without restriction,
+# including without limitation the rights to use, copy, modify, merge, publish,
+# distribute, sublicense, and/or sell copies of the Software, and
+# to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+# The above copyright notice and this permission notice shall be included
+# in all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED *AS IS*, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING
+# BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
+# DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+import os
+import sys
+from operator import itemgetter
+
+import sklearn.mixture as mm
+
+sys.path.insert(1, os.path.join(sys.path[0], '..'))
+
+
+def create_model(samples_x, samples_y_aggregation, percentage_goodbatch=0.34):
+    '''
+    Create the Gaussian Mixture Model
+    '''
+    samples = [samples_x[i] + [samples_y_aggregation[i]] for i in range(0, len(samples_x))]
+
+    # Sorts so that we can get the top samples
+    samples = sorted(samples, key=itemgetter(-1))
+    samples_goodbatch_size = int(len(samples) * percentage_goodbatch)
+    samples_goodbatch = samples[0:samples_goodbatch_size]
+    samples_badbatch = samples[samples_goodbatch_size:]
+
+    samples_x_goodbatch = [sample_goodbatch[0:-1] for sample_goodbatch in samples_goodbatch]
+    #samples_y_goodbatch = [sample_goodbatch[-1] for sample_goodbatch in samples_goodbatch]
+    samples_x_badbatch = [sample_badbatch[0:-1] for sample_badbatch in samples_badbatch]
+
+    # === Trains GMM clustering models === #
+    #sys.stderr.write("[%s] Train GMM's GMM model\n" % (os.path.basename(__file__)))
+    bgmm_goodbatch = mm.BayesianGaussianMixture(n_components=max(1, samples_goodbatch_size - 1))
+    bad_n_components = max(1, len(samples_x) - samples_goodbatch_size - 1)
+    bgmm_badbatch = mm.BayesianGaussianMixture(n_components=bad_n_components)
+    bgmm_goodbatch.fit(samples_x_goodbatch)
+    bgmm_badbatch.fit(samples_x_badbatch)
+
+    model = {}
+    model['clusteringmodel_good'] = bgmm_goodbatch
+    model['clusteringmodel_bad'] = bgmm_badbatch
+    return model
+

src/sdk/pynni/nni/metis_tuner/Regression_GMM/Selection.py

@@ -0,0 +1,104 @@
+# Copyright (c) Microsoft Corporation
+# All rights reserved.
+#
+# MIT License
+#
+# Permission is hereby granted, free of charge,
+# to any person obtaining a copy of this software and associated
+# documentation files (the "Software"), to deal in the Software without restriction,
+# including without limitation the rights to use, copy, modify, merge, publish,
+# distribute, sublicense, and/or sell copies of the Software, and
+# to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+# The above copyright notice and this permission notice shall be included
+# in all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED *AS IS*, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING
+# BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
+# DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+import os
+import random
+import sys
+
+import nni.metis_tuner.lib_acquisition_function as lib_acquisition_function
+import nni.metis_tuner.lib_constraint_summation as lib_constraint_summation
+import nni.metis_tuner.lib_data as lib_data
+
+sys.path.insert(1, os.path.join(sys.path[0], '..'))
+
+
+CONSTRAINT_LOWERBOUND = None
+CONSTRAINT_UPPERBOUND = None
+CONSTRAINT_PARAMS_IDX = []
+
+
+def _ratio_scores(parameters_value, clusteringmodel_gmm_good, clusteringmodel_gmm_bad):
+    '''
+    The ratio is smaller the better
+    '''
+    ratio = clusteringmodel_gmm_good.score([parameters_value]) / clusteringmodel_gmm_bad.score([parameters_value])
+    sigma = 0
+    return ratio, sigma
+
+def selection_r(x_bounds,
+                x_types,
+                clusteringmodel_gmm_good,
+                clusteringmodel_gmm_bad,
+                num_starting_points=100,
+                minimize_constraints_fun=None):
+    '''
+    Call selection
+    '''
+    minimize_starting_points = [lib_data.rand(x_bounds, x_types)\
+                                    for i in range(0, num_starting_points)]
+    outputs = selection(x_bounds, x_types,
+                        clusteringmodel_gmm_good,
+                        clusteringmodel_gmm_bad,
+                        minimize_starting_points,
+                        minimize_constraints_fun)
+    return outputs
+
+def selection(x_bounds,
+              x_types,
+              clusteringmodel_gmm_good,
+              clusteringmodel_gmm_bad,
+              minimize_starting_points,
+              minimize_constraints_fun=None):
+    '''
+    Select the lowest mu value
+    '''
+    results = lib_acquisition_function.next_hyperparameter_lowest_mu(\
+                    _ratio_scores, [clusteringmodel_gmm_good, clusteringmodel_gmm_bad],\
+                    x_bounds, x_types, minimize_starting_points, \
+                    minimize_constraints_fun=minimize_constraints_fun)
+
+    return results
+
+def _rand_with_constraints(x_bounds, x_types):
+    '''
+    Random generate the variable with constraints
+    '''
+    outputs = None
+    x_bounds_withconstraints = [x_bounds[i] for i in CONSTRAINT_PARAMS_IDX]
+    x_types_withconstraints = [x_types[i] for i in CONSTRAINT_PARAMS_IDX]
+    x_val_withconstraints = lib_constraint_summation.rand(x_bounds_withconstraints,
+                                                          x_types_withconstraints,
+                                                          CONSTRAINT_LOWERBOUND,
+                                                          CONSTRAINT_UPPERBOUND)
+    if x_val_withconstraints is not None:
+        outputs = [None] * len(x_bounds)
+        for i, _ in enumerate(CONSTRAINT_PARAMS_IDX):
+            outputs[CONSTRAINT_PARAMS_IDX[i]] = x_val_withconstraints[i]
+        for i, _ in enumerate(outputs):
+            if outputs[i] is None:
+                outputs[i] = random.randint(x_bounds[i][0], x_bounds[i][1])
+    return outputs
+
+def _minimize_constraints_fun_summation(x):
+    '''
+    Minimize constraints fun summation
+    '''
+    summation = sum([x[i] for i in CONSTRAINT_PARAMS_IDX])
+    return CONSTRAINT_UPPERBOUND >= summation >= CONSTRAINT_LOWERBOUND

src/sdk/pynni/nni/metis_tuner/Regression_GP/CreateModel.py

@@ -0,0 +1,52 @@
+# Copyright (c) Microsoft Corporation
+# All rights reserved.
+#
+# MIT License
+#
+# Permission is hereby granted, free of charge,
+# to any person obtaining a copy of this software and associated
+# documentation files (the "Software"), to deal in the Software without restriction,
+# including without limitation the rights to use, copy, modify, merge, publish,
+# distribute, sublicense, and/or sell copies of the Software, and
+# to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+# The above copyright notice and this permission notice shall be included
+# in all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED *AS IS*, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING
+# BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
+# DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+import os
+import sys
+import numpy
+
+import sklearn.gaussian_process as gp
+
+sys.path.insert(1, os.path.join(sys.path[0], '..'))
+
+
+def create_model(samples_x, samples_y_aggregation,
+                 n_restarts_optimizer=250, is_white_kernel=False):
+    '''
+    Trains GP regression model
+    '''
+    kernel = gp.kernels.ConstantKernel(constant_value=1,
+                                       constant_value_bounds=(1e-12, 1e12)) * \
+                                                gp.kernels.Matern(nu=1.5)
+    if is_white_kernel is True:
+        kernel += gp.kernels.WhiteKernel(noise_level=1, noise_level_bounds=(1e-12, 1e12))
+    regressor = gp.GaussianProcessRegressor(kernel=kernel,
+                                            n_restarts_optimizer=n_restarts_optimizer,
+                                            normalize_y=True,
+                                            alpha=0)
+    regressor.fit(numpy.array(samples_x), numpy.array(samples_y_aggregation))
+
+    model = {}
+    model['model'] = regressor
+    model['kernel_prior'] = str(kernel)
+    model['kernel_posterior'] = str(regressor.kernel_)
+    model['model_loglikelihood'] = regressor.log_marginal_likelihood(regressor.kernel_.theta)
+
+    return model