Fix a few bugs in Retiarii and upgrade Dockerfile (#3713)

Dockerfile

@@ -1,7 +1,7 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
-FROM nvidia/cuda:9.2-cudnn7-runtime-ubuntu18.04
+FROM nvidia/cuda:10.2-cudnn8-runtime-ubuntu18.04
 
 ARG NNI_RELEASE
 
@@ -44,7 +44,7 @@ RUN ln -s python3 /usr/bin/python
 RUN python3 -m pip install --upgrade pip==20.2.4 setuptools==50.3.2
 
 # numpy 1.14.3  scipy 1.1.0
-RUN python3 -m pip --no-cache-dir install numpy==1.14.3 scipy==1.1.0
+RUN python3 -m pip --no-cache-dir install numpy==1.19.5 scipy==1.6.3
 
 #
 # TensorFlow
@@ -52,15 +52,14 @@ RUN python3 -m pip --no-cache-dir install numpy==1.14.3 scipy==1.1.0
 RUN python3 -m pip --no-cache-dir install tensorflow==2.3.1
 
 #
-# Keras 2.1.6
+# Keras
 #
-RUN python3 -m pip --no-cache-dir install Keras==2.1.6
+RUN python3 -m pip --no-cache-dir install Keras==2.4.0
 
 #
 # PyTorch
 #
-RUN python3 -m pip --no-cache-dir install torch==1.6.0
-RUN python3 -m pip install torchvision==0.7.0
+RUN python3 -m pip --no-cache-dir install torch==1.7.1 torchvision==0.8.2 pytorch-lightning==1.3.3
 
 #
 # sklearn 0.24.1
@@ -70,7 +69,7 @@ RUN python3 -m pip --no-cache-dir install scikit-learn==0.24.1
 #
 # pandas==0.23.4 lightgbm==2.2.2
 #
-RUN python3 -m pip --no-cache-dir install pandas==0.23.4 lightgbm==2.2.2
+RUN python3 -m pip --no-cache-dir install pandas==1.1 lightgbm==2.2.2
 
 #
 # Install NNI

docs/en_US/NAS/retiarii/Advanced.rst

@@ -8,10 +8,13 @@ If you are experiencing issues with TorchScript, or the generated model code by
 
 This will come as the default execution engine in future version of Retiarii.
 
-Two steps are needed to enable this engine now.
+Three steps are needed to enable this engine now.
 
 1. Add ``@nni.retiarii.model_wrapper`` decorator outside the whole PyTorch model.
 2. Add ``config.execution_engine = 'py'`` to ``RetiariiExeConfig``.
+3. If you need to export top models, formatter needs to be set to ``dict``. Exporting ``code`` won't work with this engine.
+
+.. note:: You should always use ``super().__init__()` instead of ``super(MyNetwork, self).__init__()`` in the PyTorch model, because the latter one has issues with model wrapper.
 
 ``@basic_unit`` and ``serializer``
 ----------------------------------

examples/nas/multi-trial/mnist/search.py

@@ -4,22 +4,23 @@ import nni.retiarii.nn.pytorch as nn
 import nni.retiarii.strategy as strategy
 import nni.retiarii.evaluator.pytorch.lightning as pl
 import torch.nn.functional as F
-from nni.retiarii import serialize
+from nni.retiarii import serialize, model_wrapper
 from nni.retiarii.experiment.pytorch import RetiariiExeConfig, RetiariiExperiment, debug_mutated_model
 from torch.utils.data import DataLoader
 from torchvision import transforms
 from torchvision.datasets import MNIST
 
-
+# uncomment this for python execution engine
+# @model_wrapper
 class Net(nn.Module):
     def __init__(self, hidden_size):
-        super(Net, self).__init__()
+        super().__init__()
         self.conv1 = nn.Conv2d(1, 20, 5, 1)
         self.conv2 = nn.Conv2d(20, 50, 5, 1)
         self.fc1 = nn.LayerChoice([
             nn.Linear(4*4*50, hidden_size),
             nn.Linear(4*4*50, hidden_size, bias=False)
-        ])
+        ], label='fc1_choice')
         self.fc2 = nn.Linear(hidden_size, 10)
 
     def forward(self, x):
@@ -55,8 +56,13 @@ if __name__ == '__main__':
     exp_config.trial_concurrency = 2
     exp_config.max_trial_number = 2
     exp_config.training_service.use_active_gpu = False
+    export_formatter = 'code'
+
+    # uncomment this for python execution engine
+    # exp_config.execution_engine = 'py'
+    # export_formatter = 'dict'
 
     exp.run(exp_config, 8081 + random.randint(0, 100))
     print('Final model:')
-    for model_code in exp.export_top_models():
+    for model_code in exp.export_top_models(formatter=export_formatter):
         print(model_code)

nni/retiarii/codegen/pytorch.py

@@ -162,6 +162,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 import torch.optim as optim
 
+import nni.retiarii.nn.pytorch
+
 {}
 
 {}

nni/retiarii/execution/python.py

@@ -30,7 +30,7 @@ class PythonGraphData:
 class PurePythonExecutionEngine(BaseExecutionEngine):
     @classmethod
     def pack_model_data(cls, model: Model) -> Any:
-        mutation = {mut.mutator.label: _unpack_if_only_one(mut.samples) for mut in model.history}
+        mutation = get_mutation_dict(model)
         graph_data = PythonGraphData(get_importable_name(model.python_class, relocate_module=True),
                                      model.python_init_params, mutation, model.evaluator)
         return graph_data
@@ -51,3 +51,7 @@ def _unpack_if_only_one(ele: List[Any]):
     if len(ele) == 1:
         return ele[0]
     return ele
+
+
+def get_mutation_dict(model: Model):
+    return {mut.mutator.label: _unpack_if_only_one(mut.samples) for mut in model.history}

nni/retiarii/experiment/pytorch.py

@@ -29,6 +29,7 @@ from nni.tools.nnictl.command_utils import kill_command
 from ..codegen import model_to_pytorch_script
 from ..converter import convert_to_graph
 from ..execution import list_models, set_execution_engine
+from ..execution.python import get_mutation_dict
 from ..graph import Model, Evaluator
 from ..integration import RetiariiAdvisor
 from ..mutator import Mutator
@@ -317,7 +318,7 @@ class RetiariiExperiment(Experiment):
         """
         Export several top performing models.
 
-        For one-shot algorithms, only top-1 is supported. For others, ``optimize_mode`` asnd ``formater`` is
+        For one-shot algorithms, only top-1 is supported. For others, ``optimize_mode`` and ``formatter`` are
         available for customization.
 
         top_k : int
@@ -326,8 +327,12 @@ class RetiariiExperiment(Experiment):
             ``maximize`` or ``minimize``. Not supported by one-shot algorithms.
             ``optimize_mode`` is likely to be removed and defined in strategy in future.
         formatter : str
-            Only model code is supported for now. Not supported by one-shot algorithms.
+            Support ``code`` and ``dict``. Not supported by one-shot algorithms.
+            If ``code``, the python code of model will be returned.
+            If ``dict``, the mutation history will be returned.
         """
+        if formatter == 'code':
+            assert self.config.execution_engine != 'py', 'You should use `dict` formatter when using Python execution engine.'
         if isinstance(self.trainer, BaseOneShotTrainer):
             assert top_k == 1, 'Only support top_k is 1 for now.'
             return self.trainer.export()
@@ -335,9 +340,11 @@ class RetiariiExperiment(Experiment):
             all_models = filter(lambda m: m.metric is not None, list_models())
             assert optimize_mode in ['maximize', 'minimize']
             all_models = sorted(all_models, key=lambda m: m.metric, reverse=optimize_mode == 'maximize')
-            assert formatter == 'code', 'Export formatter other than "code" is not supported yet.'
+            assert formatter in ['code', 'dict'], 'Export formatter other than "code" and "dict" is not supported yet.'
             if formatter == 'code':
                 return [model_to_pytorch_script(model) for model in all_models[:top_k]]
+            elif formatter == 'dict':
+                return [get_mutation_dict(model) for model in all_models[:top_k]]
 
     def retrain_model(self, model):
         """

nni/retiarii/strategy/_rl_impl.py

@@ -1,6 +1,7 @@
 # This file might cause import error for those who didn't install RL-related dependencies
 
 import logging
+from multiprocessing.pool import ThreadPool
 
 import gym
 import numpy as np
@@ -9,6 +10,7 @@ import torch.nn as nn
 
 from gym import spaces
 from tianshou.data import to_torch
+from tianshou.env.worker import EnvWorker
 
 from .utils import get_targeted_model
 from ..graph import ModelStatus
@@ -18,6 +20,41 @@ from ..execution import submit_models, wait_models
 _logger = logging.getLogger(__name__)
 
 
+class MultiThreadEnvWorker(EnvWorker):
+    def __init__(self, env_fn):
+        self.env = env_fn()
+        self.pool = ThreadPool(processes=1)
+        super().__init__(env_fn)
+
+    def __getattr__(self, key):
+        return getattr(self.env, key)
+
+    def reset(self):
+        return self.env.reset()
+
+    @staticmethod
+    def wait(*args, **kwargs):
+        raise NotImplementedError('Async collect is not supported yet.')
+
+    def send_action(self, action) -> None:
+        # self.result is actually a handle
+        self.result = self.pool.apply_async(self.env.step, (action,))
+
+    def get_result(self):
+        return self.result.get()
+
+    def seed(self, seed):
+        super().seed(seed)
+        return self.env.seed(seed)
+
+    def render(self, **kwargs):
+        return self.env.render(**kwargs)
+
+    def close_env(self) -> None:
+        self.pool.terminate()
+        return self.env.close()
+
+
 class ModelEvaluationEnv(gym.Env):
     def __init__(self, base_model, mutators, search_space):
         self.base_model = base_model
@@ -107,7 +144,7 @@ class Actor(nn.Module):
         # to take care of choices with different number of options
         mask = torch.arange(self.action_dim).expand(len(out), self.action_dim) >= obs['action_dim'].unsqueeze(1)
         out[mask.to(out.device)] = float('-inf')
-        return nn.functional.softmax(out), kwargs.get('state', None)
+        return nn.functional.softmax(out, dim=-1), kwargs.get('state', None)
 
 
 class Critic(nn.Module):

nni/retiarii/strategy/rl.py

@@ -8,10 +8,10 @@ from ..execution import query_available_resources
 try:
     has_tianshou = True
     import torch
-    from tianshou.data import AsyncCollector, Collector, VectorReplayBuffer
-    from tianshou.env import SubprocVectorEnv
+    from tianshou.data import Collector, VectorReplayBuffer
+    from tianshou.env import BaseVectorEnv
     from tianshou.policy import BasePolicy, PPOPolicy  # pylint: disable=unused-import
-    from ._rl_impl import ModelEvaluationEnv, Preprocessor, Actor, Critic
+    from ._rl_impl import ModelEvaluationEnv, MultiThreadEnvWorker, Preprocessor, Actor, Critic
 except ImportError:
     has_tianshou = False
 
@@ -25,8 +25,6 @@ class PolicyBasedRL(BaseStrategy):
     This is a wrapper of algorithms provided in tianshou (PPO by default),
     and can be easily customized with other algorithms that inherit ``BasePolicy`` (e.g., REINFORCE [1]_).
 
-    Note that RL algorithms are known to have issues on Windows and MacOS. They will be supported in future.
-
     Parameters
     ----------
     max_collect : int
@@ -36,12 +34,6 @@ class PolicyBasedRL(BaseStrategy):
         After each collect, trainer will sample batch from replay buffer and do the update. Default: 20.
     policy_fn : function
         Takes ``ModelEvaluationEnv`` as input and return a policy. See ``_default_policy_fn`` for an example.
-    asynchronous : bool
-        If true, in each step, collector won't wait for all the envs to complete.
-        This should generally not affect the result, but might affect the efficiency. Note that a slightly more trials
-        than expected might be collected if this is enabled.
-        If asynchronous is false, collector will wait for all parallel environments to complete in each step.
-        See ``tianshou.data.AsyncCollector`` for more details.
 
     References
     ----------
@@ -51,7 +43,7 @@ class PolicyBasedRL(BaseStrategy):
     """
 
     def __init__(self, max_collect: int = 100, trial_per_collect = 20,
-                 policy_fn: Optional[Callable[['ModelEvaluationEnv'], 'BasePolicy']] = None, asynchronous: bool = True):
+                 policy_fn: Optional[Callable[['ModelEvaluationEnv'], 'BasePolicy']] = None):
         if not has_tianshou:
             raise ImportError('`tianshou` is required to run RL-based strategy. '
                               'Please use "pip install tianshou" to install it beforehand.')
@@ -59,7 +51,6 @@ class PolicyBasedRL(BaseStrategy):
         self.policy_fn = policy_fn or self._default_policy_fn
         self.max_collect = max_collect
         self.trial_per_collect = trial_per_collect
-        self.asynchronous = asynchronous
 
     @staticmethod
     def _default_policy_fn(env):
@@ -77,13 +68,8 @@ class PolicyBasedRL(BaseStrategy):
         env_fn = lambda: ModelEvaluationEnv(base_model, applied_mutators, search_space)
         policy = self.policy_fn(env_fn())
 
-        if self.asynchronous:
-            # wait for half of the env complete in each step
-            env = SubprocVectorEnv([env_fn for _ in range(concurrency)], wait_num=int(concurrency * 0.5))
-            collector = AsyncCollector(policy, env, VectorReplayBuffer(20000, len(env)))
-        else:
-            env = SubprocVectorEnv([env_fn for _ in range(concurrency)])
-            collector = Collector(policy, env, VectorReplayBuffer(20000, len(env)))
+        env = BaseVectorEnv([env_fn for _ in range(concurrency)], MultiThreadEnvWorker)
+        collector = Collector(policy, env, VectorReplayBuffer(20000, len(env)))
 
         for cur_collect in range(1, self.max_collect + 1):
             _logger.info('Collect [%d] Running...', cur_collect)

test/ut/retiarii/debug_mnist_pytorch.py

@@ -3,6 +3,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 import torch.optim as optim
 
+import nni.retiarii.nn.pytorch
+
 import torch
 
 

test/ut/retiarii/test_strategy.py

@@ -141,7 +141,6 @@ def test_evolution():
     _reset_execution_engine()
 
 
-@pytest.mark.skipif(sys.platform in ('win32', 'darwin'), reason='Does not run on Windows and MacOS')
 def test_rl():
     rl = strategy.PolicyBasedRL(max_collect=2, trial_per_collect=10)
     engine = MockExecutionEngine(failure_prob=0.2)
@@ -150,7 +149,7 @@ def test_rl():
     wait_models(*engine.models)
     _reset_execution_engine()
 
-    rl = strategy.PolicyBasedRL(max_collect=2, trial_per_collect=10, asynchronous=False)
+    rl = strategy.PolicyBasedRL(max_collect=2, trial_per_collect=10)
     engine = MockExecutionEngine(failure_prob=0.2)
     _reset_execution_engine(engine)
     rl.run(*_get_model_and_mutators())