setuptools-based plugin for StatsWriters

.github/workflows/pytest.yml

@@ -45,6 +45,7 @@ jobs:
         python -m pip install --progress-bar=off -e ./ml-agents
         python -m pip install --progress-bar=off -r test_requirements.txt
         python -m pip install --progress-bar=off -e ./gym-unity
+        python -m pip install --progress-bar=off -e ./ml-agents-plugin-examples
     - name: Save python dependencies
       run: |
         pip freeze > pip_versions-${{ matrix.python-version }}.txt

com.unity.ml-agents/CHANGELOG.md

@@ -12,6 +12,9 @@ and this project adheres to
 #### com.unity.ml-agents (C#)
 #### ml-agents / ml-agents-envs / gym-unity (Python)
 - TensorFlow trainers have been removed, please use the Torch trainers instead. (#4707)
+- A plugin system for `mlagents-learn` has been added. You can now define custom
+  `StatsWriter` implementations and register them to be called during training.
+  More types of plugins will be added in the future. (#4788)
 
 ### Minor Changes
 #### com.unity.ml-agents / com.unity.ml-agents.extensions (C#)

docs/Training-Plugins.md

@@ -0,0 +1,58 @@
+# Customizing Training via Plugins
+
+ML-Agents provides support for running your own python implementations of specific interfaces during the training
+process. These interfaces are currently fairly limited, but will be expanded in the future.
+
+** Note ** Plugin interfaces should currently be considered "in beta", and they may change in future releases.
+
+## How to Write Your Own Plugin
+[This video](https://www.youtube.com/watch?v=fY3Y_xPKWNA) explains the basics of how to create a plugin system using
+setuptools, and is the same approach that ML-Agents' plugin system is based on.
+
+The `ml-agents-plugin-examples` directory contains a reference implementation of each plugin interface, so it's a good
+starting point.
+
+### setup.py
+If you don't already have a `setup.py` file for your python code, you'll need to add one. `ml-agents-plugin-examples`
+has a [minimal example](../ml-agents-plugin-examples/setup.py) of this.
+
+In the call to `setup()`, you'll need to add to the `entry_points` dictionary for each plugin interface that you
+implement. The form of this is `{entry point name}={plugin module}:{plugin function}`. For example, in
+ `ml-agents-plugin-examples`:
+```python
+entry_points={
+    ML_AGENTS_STATS_WRITER: [
+        "example=mlagents_plugin_examples.example_stats_writer:get_example_stats_writer"
+    ]
+}
+```
+* `ML_AGENTS_STATS_WRITER` (which is a string constant, `mlagents.stats_writer`) is the name of the plugin interface.
+This must be one of the provided interfaces ([see below](#plugin-interfaces)).
+* `example` is the plugin implementation name. This can be anything.
+* `mlagents_plugin_examples.example_stats_writer` is the plugin module. This points to the module where the
+plugin registration function is defined.
+* `get_example_stats_writer` is the plugin registration function. This is called when running `mlagents-learn`. The
+arguments and expected return type for this are different for each plugin interface.
+
+### Local Installation
+Once you've defined `entry_points` in your `setup.py`, you will need to run
+```
+pip install -e [path to your plugin code]
+```
+in the same python virtual environment that you have `mlagents` installed.
+
+## Plugin Interfaces
+
+### StatsWriter
+The StatsWriter class receives various information from the training process, such as the average Agent reward in
+each summary period. By default, we log this information to the console and write it to
+[TensorBoard](Using-Tensorboard.md).
+
+#### Interface
+The `StatsWriter.write_stats()` method must be implemented in any derived classes. It takes a "category" parameter,
+which typically is the behavior name of the Agents being trained, and a dictionary of `StatSummary` values with
+string keys.
+
+#### Registration
+The `StatsWriter` registration function takes a `RunOptions` argument and returns a list of `StatsWriter`s. An
+example implementation is provided in [`mlagents_plugin_examples`](../ml-agents-plugin-examples/mlagents_plugin_examples/example_stats_writer.py)

ml-agents-plugin-examples/README.md

@@ -0,0 +1,3 @@
+# ML-Agents Plugins
+
+See the [Plugins documentation](../docs/Training-Plugins.md) for more information.

ml-agents-plugin-examples/mlagents_plugin_examples/example_stats_writer.py

@@ -0,0 +1,27 @@
+from typing import Dict, List
+from mlagents.trainers.settings import RunOptions
+from mlagents.trainers.stats import StatsWriter, StatsSummary
+
+
+class ExampleStatsWriter(StatsWriter):
+    """
+    Example implementation of the StatsWriter abstract class.
+    This doesn't do anything interesting, just prints the stats that it gets.
+    """
+
+    def write_stats(
+        self, category: str, values: Dict[str, StatsSummary], step: int
+    ) -> None:
+        print(f"ExampleStatsWriter category: {category} values: {values}")
+
+
+def get_example_stats_writer(run_options: RunOptions) -> List[StatsWriter]:
+    """
+    Registration function. This is referenced in setup.py and will
+    be called by mlagents-learn when it starts to determine the
+    list of StatsWriters to use.
+
+    It must return a list of StatsWriters.
+    """
+    print("Creating a new stats writer! This is so exciting!")
+    return [ExampleStatsWriter()]

ml-agents-plugin-examples/mlagents_plugin_examples/tests/test_stats_writer_plugin.py

@@ -0,0 +1,13 @@
+import pytest
+
+from mlagents.plugins.stats_writer import register_stats_writer_plugins
+from mlagents.trainers.settings import RunOptions
+
+from mlagents_plugin_examples.example_stats_writer import ExampleStatsWriter
+
+
+@pytest.mark.check_environment_trains
+def test_register_stats_writers():
+    # Make sure that the ExampleStatsWriter gets returned from the list of all StatsWriters
+    stats_writers = register_stats_writer_plugins(RunOptions())
+    assert any(isinstance(sw, ExampleStatsWriter) for sw in stats_writers)

ml-agents-plugin-examples/setup.py

@@ -0,0 +1,17 @@
+from setuptools import setup
+from mlagents.plugins import ML_AGENTS_STATS_WRITER
+
+setup(
+    name="mlagents_plugin_examples",
+    version="0.0.1",
+    # Example of how to add your own registration functions that will be called
+    # by mlagents-learn.
+    #
+    # Here, the get_example_stats_writer() function in mlagents_plugin_examples/example_stats_writer.py
+    # will get registered with the ML_AGENTS_STATS_WRITER plugin interface.
+    entry_points={
+        ML_AGENTS_STATS_WRITER: [
+            "example=mlagents_plugin_examples.example_stats_writer:get_example_stats_writer"
+        ]
+    },
+)

ml-agents/mlagents/plugins/__init__.py

@@ -0,0 +1 @@
+ML_AGENTS_STATS_WRITER = "mlagents.stats_writer"

ml-agents/mlagents/plugins/stats_writer.py

@@ -0,0 +1,63 @@
+import sys
+from typing import List
+
+# importlib.metadata is new in python3.8
+# We use the backport for older python versions.
+if sys.version_info < (3, 8):
+    import importlib_metadata
+else:
+    import importlib.metadata as importlib_metadata  # pylint: disable=E0611
+
+from mlagents.trainers.stats import StatsWriter
+
+from mlagents_envs import logging_util
+from mlagents.plugins import ML_AGENTS_STATS_WRITER
+from mlagents.trainers.settings import RunOptions
+from mlagents.trainers.stats import TensorboardWriter, GaugeWriter, ConsoleWriter
+
+
+logger = logging_util.get_logger(__name__)
+
+
+def get_default_stats_writers(run_options: RunOptions) -> List[StatsWriter]:
+    """
+    The StatsWriters that mlagents-learn always uses:
+    * A TensorboardWriter to write information to TensorBoard
+    * A GaugeWriter to record our internal stats
+    * A ConsoleWriter to output to stdout.
+    """
+    checkpoint_settings = run_options.checkpoint_settings
+    return [
+        TensorboardWriter(
+            checkpoint_settings.write_path,
+            clear_past_data=not checkpoint_settings.resume,
+        ),
+        GaugeWriter(),
+        ConsoleWriter(),
+    ]
+
+
+def register_stats_writer_plugins(run_options: RunOptions) -> List[StatsWriter]:
+    """
+    Registers all StatsWriter plugins (including the default one),
+    and evaluates them, and returns the list of all the StatsWriter implementations.
+    """
+    all_stats_writers: List[StatsWriter] = []
+    entry_points = importlib_metadata.entry_points()[ML_AGENTS_STATS_WRITER]
+
+    for entry_point in entry_points:
+
+        try:
+            logger.debug(f"Initializing StatsWriter plugins: {entry_point.name}")
+            plugin_func = entry_point.load()
+            plugin_stats_writers = plugin_func(run_options)
+            logger.debug(
+                f"Found {len(plugin_stats_writers)} StatsWriters for plugin {entry_point.name}"
+            )
+            all_stats_writers += plugin_stats_writers
+        except BaseException:
+            # Catch all exceptions from setting up the plugin, so that bad user code doesn't break things.
+            logger.exception(
+                f"Error initializing StatsWriter plugins for {entry_point.name}. This plugin will not be used."
+            )
+    return all_stats_writers

ml-agents/mlagents/trainers/cli_utils.py

@@ -189,6 +189,9 @@ def _create_parser() -> argparse.ArgumentParser:
         action=RaiseRemovedWarning,
         help="(Removed) Use the TensorFlow framework.",
     )
+    argparser.add_argument(
+        "--results-dir", default="results", help="Results base directory"
+    )
 
     eng_conf = argparser.add_argument_group(title="Engine Configuration")
     eng_conf.add_argument(

ml-agents/mlagents/trainers/learn.py

@@ -14,12 +14,7 @@
 from mlagents.trainers.environment_parameter_manager import EnvironmentParameterManager
 from mlagents.trainers.trainer import TrainerFactory
 from mlagents.trainers.directory_utils import validate_existing_directories
-from mlagents.trainers.stats import (
-    TensorboardWriter,
-    StatsReporter,
-    GaugeWriter,
-    ConsoleWriter,
-)
+from mlagents.trainers.stats import StatsReporter
 from mlagents.trainers.cli_utils import parser
 from mlagents_envs.environment import UnityEnvironment
 from mlagents.trainers.settings import RunOptions
@@ -34,6 +29,7 @@
     add_metadata as add_timer_metadata,
 )
 from mlagents_envs import logging_util
+from mlagents.plugins.stats_writer import register_stats_writer_plugins
 
 logger = logging_util.get_logger(__name__)
 
@@ -65,21 +61,15 @@ def run_training(run_seed: int, options: RunOptions) -> None:
         checkpoint_settings = options.checkpoint_settings
         env_settings = options.env_settings
         engine_settings = options.engine_settings
-        base_path = "results"
-        write_path = os.path.join(base_path, checkpoint_settings.run_id)
-        maybe_init_path = (
-            os.path.join(base_path, checkpoint_settings.initialize_from)
-            if checkpoint_settings.initialize_from is not None
-            else None
-        )
-        run_logs_dir = os.path.join(write_path, "run_logs")
+
+        run_logs_dir = checkpoint_settings.run_logs_dir
         port: Optional[int] = env_settings.base_port
         # Check if directory exists
         validate_existing_directories(
-            write_path,
+            checkpoint_settings.write_path,
             checkpoint_settings.resume,
             checkpoint_settings.force,
-            maybe_init_path,
+            checkpoint_settings.maybe_init_path,
         )
         # Make run logs directory
         os.makedirs(run_logs_dir, exist_ok=True)
@@ -90,14 +80,9 @@ def run_training(run_seed: int, options: RunOptions) -> None:
             )
 
         # Configure Tensorboard Writers and StatsReporter
-        tb_writer = TensorboardWriter(
-            write_path, clear_past_data=not checkpoint_settings.resume
-        )
-        gauge_write = GaugeWriter()
-        console_writer = ConsoleWriter()
-        StatsReporter.add_writer(tb_writer)
-        StatsReporter.add_writer(gauge_write)
-        StatsReporter.add_writer(console_writer)
+        stats_writers = register_stats_writer_plugins(options)
+        for sw in stats_writers:
+            StatsReporter.add_writer(sw)
 
         if env_settings.env_path is None:
             port = None
@@ -117,18 +102,18 @@ def run_training(run_seed: int, options: RunOptions) -> None:
 
         trainer_factory = TrainerFactory(
             trainer_config=options.behaviors,
-            output_path=write_path,
+            output_path=checkpoint_settings.write_path,
             train_model=not checkpoint_settings.inference,
             load_model=checkpoint_settings.resume,
             seed=run_seed,
             param_manager=env_parameter_manager,
-            init_path=maybe_init_path,
+            init_path=checkpoint_settings.maybe_init_path,
             multi_gpu=False,
         )
         # Create controller and begin training.
         tc = TrainerController(
             trainer_factory,
-            write_path,
+            checkpoint_settings.write_path,
             checkpoint_settings.run_id,
             env_parameter_manager,
             not checkpoint_settings.inference,
@@ -140,7 +125,7 @@ def run_training(run_seed: int, options: RunOptions) -> None:
         tc.start_learning(env_manager)
     finally:
         env_manager.close()
-        write_run_options(write_path, options)
+        write_run_options(checkpoint_settings.write_path, options)
         write_timing_tree(run_logs_dir)
         write_training_status(run_logs_dir)
 

ml-agents/mlagents/trainers/settings.py

@@ -1,3 +1,4 @@
+import os.path
 import warnings
 
 import attr
@@ -706,6 +707,23 @@ class CheckpointSettings:
     force: bool = parser.get_default("force")
     train_model: bool = parser.get_default("train_model")
     inference: bool = parser.get_default("inference")
+    results_dir: str = parser.get_default("results_dir")
+
+    @property
+    def write_path(self) -> str:
+        return os.path.join(self.results_dir, self.run_id)
+
+    @property
+    def maybe_init_path(self) -> Optional[str]:
+        return (
+            os.path.join(self.results_dir, self.initialize_from)
+            if self.initialize_from is not None
+            else None
+        )
+
+    @property
+    def run_logs_dir(self) -> str:
+        return os.path.join(self.write_path, "run_logs")
 
 
 @attr.s(auto_attribs=True)

ml-agents/mlagents/trainers/stats.py

@@ -87,6 +87,13 @@ class StatsWriter(abc.ABC):
     def write_stats(
         self, category: str, values: Dict[str, StatsSummary], step: int
     ) -> None:
+        """
+        Callback to record training information
+        :param category: Category of the statistics. Usually this is the behavior name.
+        :param values: Dictionary of statistics.
+        :param step: The current training step.
+        :return:
+        """
         pass
 
     def add_property(

ml-agents/setup.py

@@ -3,6 +3,7 @@
 
 from setuptools import setup, find_packages
 from setuptools.command.install import install
+from mlagents.plugins import ML_AGENTS_STATS_WRITER
 import mlagents.trainers
 
 VERSION = mlagents.trainers.__version__
@@ -71,13 +72,18 @@ def run(self):
         "cattrs>=1.0.0,<1.1.0",
         "attrs>=19.3.0",
         'pypiwin32==223;platform_system=="Windows"',
+        "importlib_metadata; python_version<'3.8'",
     ],
     python_requires=">=3.6.1",
     entry_points={
         "console_scripts": [
             "mlagents-learn=mlagents.trainers.learn:main",
             "mlagents-run-experiment=mlagents.trainers.run_experiment:main",
-        ]
+        ],
+        # Plugins - each plugin type should have an entry here for the default behavior
+        ML_AGENTS_STATS_WRITER: [
+            "default=mlagents.plugins.stats_writer:get_default_stats_writers"
+        ],
     },
     cmdclass={"verify": VerifyVersionCommand},
 )