DNGO tuner (#3479)

Co-authored-by: Yuge Zhang <Yuge.Zhang@microsoft.com>

dependencies/required_extra.txt

@@ -11,3 +11,6 @@ statsmodels==0.12.0
 
 # PPOTuner
 gym
+
+# DNGO
+pybnn

docs/en_US/Tuner/BuiltinTuner.rst

@@ -45,7 +45,8 @@ Currently, we support the following algorithms:
      - PPO Tuner is a Reinforcement Learning tuner based on PPO algorithm. `Reference Paper <https://arxiv.org/abs/1707.06347>`__
    * - `PBT Tuner <#PBTTuner>`__
      - 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>`__
+     - Use of neural networks as an alternative to GPs to model distributions over functions in bayesian optimization.
 
 Usage of Built-in Tuners
 ------------------------
@@ -574,6 +575,41 @@ Population Based Training (PBT) bridges and extends parallel search methods and
 
 Note that, to use this tuner, your trial code should be modified accordingly, please refer to `the document of PBTTuner <./PBTTuner.rst>`__ for details.
 
+DNGO Tuner
+^^^^^^^^^^
+
+..
+
+   Built-in Tuner Name: **DNGOTuner**
+
+DNGO advisor requires `pybnn`, which can be installed with the following command.
+
+.. code-block:: bash
+
+   pip install nni[DNGO]
+
+**Suggested scenario**
+
+Applicable to large scale hyperparameter optimization. Bayesian optimization that rapidly finds competitive models on benchmark object recognition tasks using convolutional networks, and image caption generation using neural language models.
+
+**classArgs requirements:**
+
+
+* **optimize_mode** (*'maximize' or 'minimize'*\ ) - If 'maximize', the tuner will target to maximize metrics. If 'minimize', the tuner will target to minimize metrics.
+* **sample_size** (*int, default = 1000*) - Number of samples to select in each iteration. The best one will be picked from the samples as the next trial.
+* **trials_per_update** (*int, default = 20*) - Number of trials to collect before updating the model.
+* **num_epochs_per_training** (*int, default = 500*) - Number of epochs to train DNGO model.
+
+**Usage example**
+
+.. code-block:: yaml
+
+   # config.yml
+   tuner:
+     builtinTunerName: DNGOTuner
+     classArgs:
+       optimize_mode: maximize
+
 **Reference and Feedback**
 ------------------------------
 

docs/en_US/Tutorial/SearchSpaceSpec.rst

@@ -243,13 +243,25 @@ Search Space Types Supported by Each Tuner
      - 
      - 
      - 
+   * - DNGO Tuner
+     - :raw-html:`✓`
+     - 
+     - :raw-html:`✓`
+     - :raw-html:`✓`
+     - :raw-html:`✓`
+     - :raw-html:`✓`
+     - :raw-html:`✓`
+     - 
+     - 
+     - 
+     - 
 
 
 Known Limitations:
 
 
 * 
-  GP Tuner and Metis Tuner support only **numerical values** in search space (\ ``choice`` type values can be no-numerical with other tuners, e.g. string values). Both GP Tuner and Metis Tuner use Gaussian Process Regressor(GPR). GPR make predictions based on a kernel function and the 'distance' between different points, it's hard to get the true distance between no-numerical values.
+  GP Tuner, Metis Tuner and DNGO tuner support only **numerical values** in search space (\ ``choice`` type values can be no-numerical with other tuners, e.g. string values). Both GP Tuner and Metis Tuner use Gaussian Process Regressor(GPR). GPR make predictions based on a kernel function and the 'distance' between different points, it's hard to get the true distance between no-numerical values.
 
 * 
   Note that for nested search space:

nni/algorithms/hpo/dngo_tuner.py

+import logging
+
+import numpy as np
+import torch
+import nni.parameter_expressions as parameter_expressions
+from nni import ClassArgsValidator
+from nni.tuner import Tuner
+from pybnn import DNGO
+from torch.distributions import Normal
+
+_logger = logging.getLogger(__name__)
+
+
+def _random_config(search_space, random_state):
+    chosen_config = {}
+    for key, val in search_space.items():
+        if val['_type'] == 'choice':
+            choices = val['_value']
+            index = random_state.randint(len(choices))
+            if all([isinstance(c, (int, float)) for c in choices]):
+                chosen_config[key] = choices[index]
+            else:
+                raise ValueError('Choices with type other than int and float is not supported.')
+        elif val['_type'] == 'uniform':
+            chosen_config[key] = random_state.uniform(val['_value'][0], val['_value'][1])
+        elif val['_type'] == 'randint':
+            chosen_config[key] = random_state.randint(
+                val['_value'][0], val['_value'][1])
+        elif val['_type'] == 'quniform':
+            chosen_config[key] = parameter_expressions.quniform(
+                val['_value'][0], val['_value'][1], val['_value'][2], random_state)
+        elif val['_type'] == 'loguniform':
+            chosen_config[key] = parameter_expressions.loguniform(
+                val['_value'][0], val['_value'][1], random_state)
+        elif val['_type'] == 'qloguniform':
+            chosen_config[key] = parameter_expressions.qloguniform(
+                val['_value'][0], val['_value'][1], val['_value'][2], random_state)
+        else:
+            raise ValueError('Unknown key %s and value %s' % (key, val))
+    return chosen_config
+
+
+class DngoTuner(Tuner):
+
+    def __init__(self, optimize_mode='maximize', sample_size=1000, trials_per_update=20, num_epochs_per_training=500):
+        self.searchspace_json = None
+        self.random_state = None
+        self.model = DNGO(do_mcmc=False, num_epochs=num_epochs_per_training)
+        self._model_initialized = False
+        self.sample_size = sample_size
+        self.trials_per_update = trials_per_update
+        self.optimize_mode = optimize_mode
+
+        self.x = []
+        self.y = []
+
+    def receive_trial_result(self, parameter_id, parameters, value, **kwargs):
+        self.x.append(parameters)
+        self.y.append(self._get_default_value(value))
+        if len(self.y) % self.trials_per_update == 0:
+            self._update_model()
+
+    def generate_parameters(self, parameter_id, **kwargs):
+        if not self._model_initialized:
+            return _random_config(self.searchspace_json, self.random_state)
+        else:
+            # random samples and pick best with model
+            candidate_x = [_random_config(self.searchspace_json, self.random_state) for _ in range(self.sample_size)]
+
+            x_test = np.array([np.array(list(xi.values())) for xi in candidate_x])
+            m, v = self.model.predict(x_test)
+            mean = torch.Tensor(m)
+            sigma = torch.Tensor(v)
+            u = (mean - torch.Tensor([0.95]).expand_as(mean)) / sigma
+            normal = Normal(torch.zeros_like(u), torch.ones_like(u))
+            ucdf = normal.cdf(u)
+            updf = torch.exp(normal.log_prob(u))
+            ei = sigma * (updf + u * ucdf)
+
+            if self.optimize_mode == 'maximize':
+                ind = torch.argmax(ei)
+            else:
+                ind = torch.argmin(ei)
+            new_x = candidate_x[ind]
+            return new_x
+
+    def update_search_space(self, search_space):
+        self.searchspace_json = search_space
+        self.random_state = np.random.RandomState()
+
+    def import_data(self, data):
+        for d in data:
+            self.x.append(d['parameter'])
+            self.y.append(self._get_default_value(d['value']))
+        self._update_model()
+
+    def _update_model(self):
+        _logger.info('Updating model on %d samples', len(self.x))
+        x_arr = []
+        for x in self.x:
+            x_arr.append([x[k] for k in sorted(x.keys())])
+        self.model.train(np.array(x_arr), np.array(self.y), do_optimize=True)
+        self._model_initialized = True
+
+    def _get_default_value(self, value):
+        if isinstance(value, dict) and 'default' in value:
+            return value['default']
+        elif isinstance(value, float):
+            return value
+        else:
+            raise ValueError(f'Unsupported value: {value}')
+
+
+class DNGOClassArgsValidator(ClassArgsValidator):
+    # DNGO tuner do not have much input arg, so the validation is actually hardly used
+    def validate_class_args(self, **kwargs):
+        pass

nni/runtime/default_config/registered_algorithms.yml

@@ -76,3 +76,7 @@ tuners:
   classArgsValidator: nni.algorithms.hpo.regularized_evolution_tuner.EvolutionClassArgsValidator
   className: nni.algorithms.hpo.regularized_evolution_tuner.RegularizedEvolutionTuner
   source: nni
+- builtinName: DNGOTuner
+  classArgsValidator: nni.algorithms.hpo.dngo_tuner.DNGOClassArgsValidator
+  className: nni.algorithms.hpo.dngo_tuner.DNGOTuner
+  source: nni

pipelines/fast-test.yml

@@ -161,7 +161,7 @@ stages:
     - script: |
         set -e
         python -m pip install -r dependencies/recommended.txt
-        python -m pip install -e .[SMAC,BOHB,PPOTuner]
+        python -m pip install -e .[SMAC,BOHB,PPOTuner,DNGO]
       displayName: Install extra dependencies
 
     # Need del later
@@ -246,7 +246,7 @@ stages:
     - script: |
         set -e
         python -m pip install -r dependencies/recommended_legacy.txt
-        python -m pip install -e .[SMAC,BOHB,PPOTuner]
+        python -m pip install -e .[SMAC,BOHB,PPOTuner,DNGO]
       displayName: Install extra dependencies
 
     # Need del later
@@ -335,7 +335,7 @@ stages:
     - script: |
         set -e
         python -m pip install -r dependencies/recommended.txt
-        python -m pip install -e .[SMAC,BOHB,PPOTuner]
+        python -m pip install -e .[SMAC,BOHB,PPOTuner,DNGO]
       displayName: Install extra dependencies
 
     # Need del later
@@ -398,6 +398,7 @@ stages:
 
     - script: |
         python -m pip install -r dependencies/recommended.txt
+        python -m pip install -e .[DNGO]
       displayName: Install extra dependencies
 
     # Need del later

pipelines/full-test-linux.yml

@@ -35,6 +35,7 @@ jobs:
       python3 -m pip install keras==2.1.6
       python3 -m pip install tensorflow==2.3.1 tensorflow-estimator==2.3.0
       python3 -m pip install thop
+      python3 -m pip install pybnn
       python3 -m pip install tianshou>=0.4.1 gym
       sudo apt-get install swig -y
     displayName: Install extra dependencies

pipelines/full-test-windows.yml

@@ -30,6 +30,7 @@ jobs:
       python -m pip install torch==1.6.0 torchvision==0.7.0 -f https://download.pytorch.org/whl/torch_stable.html
       python -m pip install 'pytorch-lightning>=1.1.1'
       python -m pip install tensorflow==2.3.1 tensorflow-estimator==2.3.0
+      python -m pip install pybnn
       python -m pip install tianshou>=0.4.1 gym
     displayName: Install extra dependencies
 

setup.py

@@ -88,7 +88,8 @@ def _setup():
         extras_require = {
             'SMAC': _read_requirements_txt('dependencies/required_extra.txt', 'SMAC'),
             'BOHB': _read_requirements_txt('dependencies/required_extra.txt', 'BOHB'),
-            'PPOTuner': _read_requirements_txt('dependencies/required_extra.txt', 'PPOTuner')
+            'PPOTuner': _read_requirements_txt('dependencies/required_extra.txt', 'PPOTuner'),
+            'DNGO': _read_requirements_txt('dependencies/required_extra.txt', 'DNGO'),
         },
         setup_requires = ['requests'],
 

test/ut/sdk/test_builtin_tuners.py

@@ -12,6 +12,7 @@ from collections import deque
 from unittest import TestCase, main
 
 from nni.algorithms.hpo.batch_tuner import BatchTuner
+from nni.algorithms.hpo.dngo_tuner import DngoTuner
 from nni.algorithms.hpo.evolution_tuner import EvolutionTuner
 from nni.algorithms.hpo.gp_tuner import GPTuner
 from nni.algorithms.hpo.gridsearch_tuner import GridSearchTuner
@@ -388,6 +389,12 @@ class BuiltinTunersTestCase(TestCase):
         ))
         self.import_data_test_for_pbt()
 
+    def test_dngo(self):
+        tuner_fn = lambda: DngoTuner(trials_per_update=100, num_epochs_per_training=1)
+        self.search_space_test_all(tuner_fn, fail_types=["choice_str", "choice_mixed",
+                                                         "normal", "lognormal", "qnormal", "qlognormal"])
+        self.import_data_test(tuner_fn, stype='choice_num')
+
     def tearDown(self):
         file_list = glob.glob("smac3*") + ["param_config_space.pcs", "scenario.txt", "model_path"]
         for file in file_list: