from collections import defaultdict from enum import Enum from typing import List, Dict, NamedTuple, Any import numpy as np import abc import os import time from threading import RLock from mlagents_envs.side_channel.stats_side_channel import StatsAggregationMethod from mlagents_envs.logging_util import get_logger from mlagents_envs.timers import set_gauge from torch.utils.tensorboard import SummaryWriter from mlagents.torch_utils.globals import get_rank logger = get_logger(__name__) def _dict_to_str(param_dict: Dict[str, Any], num_tabs: int) -> str: """ Takes a parameter dictionary and converts it to a human-readable string. Recurses if there are multiple levels of dict. Used to print out hyperparameters. :param param_dict: A Dictionary of key, value parameters. :return: A string version of this dictionary. """ if not isinstance(param_dict, dict): return str(param_dict) else: append_newline = "\n" if num_tabs > 0 else "" return append_newline + "\n".join( [ "\t" + " " * num_tabs + "{}:\t{}".format(x, _dict_to_str(param_dict[x], num_tabs + 1)) for x in param_dict ] ) class StatsSummary(NamedTuple): mean: float std: float num: int sum: float aggregation_method: StatsAggregationMethod @staticmethod def empty() -> "StatsSummary": return StatsSummary(0.0, 0.0, 0, 0.0, StatsAggregationMethod.AVERAGE) @property def aggregated_value(self): if self.aggregation_method == StatsAggregationMethod.SUM: return self.sum else: return self.mean class StatsPropertyType(Enum): HYPERPARAMETERS = "hyperparameters" SELF_PLAY = "selfplay" class StatsWriter(abc.ABC): """ A StatsWriter abstract class. A StatsWriter takes in a category, key, scalar value, and step and writes it out by some method. """ @abc.abstractmethod def write_stats( self, category: str, values: Dict[str, StatsSummary], step: int ) -> None: pass def add_property( self, category: str, property_type: StatsPropertyType, value: Any ) -> None: """ Add a generic property to the StatsWriter. This could be e.g. a Dict of hyperparameters, a max step count, a trainer type, etc. Note that not all StatsWriters need to be compatible with all types of properties. For instance, a TB writer doesn't need a max step. :param category: The category that the property belongs to. :param property_type: The type of property. :param value: The property itself. """ pass class GaugeWriter(StatsWriter): """ Write all stats that we receive to the timer gauges, so we can track them offline easily """ @staticmethod def sanitize_string(s: str) -> str: """ Clean up special characters in the category and value names. """ return s.replace("/", ".").replace(" ", "") def write_stats( self, category: str, values: Dict[str, StatsSummary], step: int ) -> None: for val, stats_summary in values.items(): set_gauge( GaugeWriter.sanitize_string(f"{category}.{val}.mean"), float(stats_summary.mean), ) set_gauge( GaugeWriter.sanitize_string(f"{category}.{val}.sum"), float(stats_summary.sum), ) class ConsoleWriter(StatsWriter): def __init__(self): self.training_start_time = time.time() # If self-play, we want to print ELO as well as reward self.self_play = False self.self_play_team = -1 self.rank = get_rank() def write_stats( self, category: str, values: Dict[str, StatsSummary], step: int ) -> None: is_training = "Not Training" if "Is Training" in values: stats_summary = values["Is Training"] if stats_summary.aggregated_value > 0.0: is_training = "Training" elapsed_time = time.time() - self.training_start_time log_info: List[str] = [category] log_info.append(f"Step: {step}") log_info.append(f"Time Elapsed: {elapsed_time:0.3f} s") if "Environment/Cumulative Reward" in values: stats_summary = values["Environment/Cumulative Reward"] if self.rank is not None: log_info.append(f"Rank: {self.rank}") log_info.append(f"Mean Reward: {stats_summary.mean:0.3f}") log_info.append(f"Std of Reward: {stats_summary.std:0.3f}") log_info.append(is_training) if self.self_play and "Self-play/ELO" in values: elo_stats = values["Self-play/ELO"] log_info.append(f"ELO: {elo_stats.mean:0.3f}") else: log_info.append("No episode was completed since last summary") log_info.append(is_training) logger.info(". ".join(log_info) + ".") def add_property( self, category: str, property_type: StatsPropertyType, value: Any ) -> None: if property_type == StatsPropertyType.HYPERPARAMETERS: logger.info( """Hyperparameters for behavior name {}: \n{}""".format( category, _dict_to_str(value, 0) ) ) elif property_type == StatsPropertyType.SELF_PLAY: assert isinstance(value, bool) self.self_play = value class TensorboardWriter(StatsWriter): def __init__(self, base_dir: str, clear_past_data: bool = False): """ A StatsWriter that writes to a Tensorboard summary. :param base_dir: The directory within which to place all the summaries. Tensorboard files will be written to a {base_dir}/{category} directory. :param clear_past_data: Whether or not to clean up existing Tensorboard files associated with the base_dir and category. """ self.summary_writers: Dict[str, SummaryWriter] = {} self.base_dir: str = base_dir self._clear_past_data = clear_past_data def write_stats( self, category: str, values: Dict[str, StatsSummary], step: int ) -> None: self._maybe_create_summary_writer(category) for key, value in values.items(): self.summary_writers[category].add_scalar( f"{key}", value.aggregated_value, step ) self.summary_writers[category].flush() def _maybe_create_summary_writer(self, category: str) -> None: if category not in self.summary_writers: filewriter_dir = "{basedir}/{category}".format( basedir=self.base_dir, category=category ) os.makedirs(filewriter_dir, exist_ok=True) if self._clear_past_data: self._delete_all_events_files(filewriter_dir) self.summary_writers[category] = SummaryWriter(filewriter_dir) def _delete_all_events_files(self, directory_name: str) -> None: for file_name in os.listdir(directory_name): if file_name.startswith("events.out"): logger.warning( f"{file_name} was left over from a previous run. Deleting." ) full_fname = os.path.join(directory_name, file_name) try: os.remove(full_fname) except OSError: logger.warning( "{} was left over from a previous run and " "not deleted.".format(full_fname) ) def add_property( self, category: str, property_type: StatsPropertyType, value: Any ) -> None: if property_type == StatsPropertyType.HYPERPARAMETERS: assert isinstance(value, dict) summary = _dict_to_str(value, 0) self._maybe_create_summary_writer(category) if summary is not None: self.summary_writers[category].add_text("Hyperparameters", summary) self.summary_writers[category].flush() class StatsReporter: writers: List[StatsWriter] = [] stats_dict: Dict[str, Dict[str, List]] = defaultdict(lambda: defaultdict(list)) lock = RLock() stats_aggregation: Dict[str, Dict[str, StatsAggregationMethod]] = defaultdict( lambda: defaultdict(lambda: StatsAggregationMethod.AVERAGE) ) def __init__(self, category: str): """ Generic StatsReporter. A category is the broadest type of storage (would correspond the run name and trainer name, e.g. 3DBalltest_3DBall. A key is the type of stat it is (e.g. Environment/Reward). Finally the Value is the float value attached to this stat. """ self.category: str = category @staticmethod def add_writer(writer: StatsWriter) -> None: with StatsReporter.lock: StatsReporter.writers.append(writer) def add_property(self, property_type: StatsPropertyType, value: Any) -> None: """ Add a generic property to the StatsReporter. This could be e.g. a Dict of hyperparameters, a max step count, a trainer type, etc. Note that not all StatsWriters need to be compatible with all types of properties. For instance, a TB writer doesn't need a max step. :param property_type: The type of property. :param value: The property itself. """ with StatsReporter.lock: for writer in StatsReporter.writers: writer.add_property(self.category, property_type, value) def add_stat( self, key: str, value: float, aggregation: StatsAggregationMethod = StatsAggregationMethod.AVERAGE, ) -> None: """ Add a float value stat to the StatsReporter. :param key: The type of statistic, e.g. Environment/Reward. :param value: the value of the statistic. :param aggregation: the aggregation method for the statistic, default StatsAggregationMethod.AVERAGE. """ with StatsReporter.lock: StatsReporter.stats_dict[self.category][key].append(value) StatsReporter.stats_aggregation[self.category][key] = aggregation def set_stat(self, key: str, value: float) -> None: """ Sets a stat value to a float. This is for values that we don't want to average, and just want the latest. :param key: The type of statistic, e.g. Environment/Reward. :param value: the value of the statistic. """ with StatsReporter.lock: StatsReporter.stats_dict[self.category][key] = [value] StatsReporter.stats_aggregation[self.category][ key ] = StatsAggregationMethod.MOST_RECENT def write_stats(self, step: int) -> None: """ Write out all stored statistics that fall under the category specified. The currently stored values will be averaged, written out as a single value, and the buffer cleared. :param step: Training step which to write these stats as. """ with StatsReporter.lock: values: Dict[str, StatsSummary] = {} for key in StatsReporter.stats_dict[self.category]: if len(StatsReporter.stats_dict[self.category][key]) > 0: stat_summary = self.get_stats_summaries(key) values[key] = stat_summary for writer in StatsReporter.writers: writer.write_stats(self.category, values, step) del StatsReporter.stats_dict[self.category] def get_stats_summaries(self, key: str) -> StatsSummary: """ Get the mean, std, count, sum and aggregation method of a particular statistic, since last write. :param key: The type of statistic, e.g. Environment/Reward. :returns: A StatsSummary containing summary statistics. """ stat_values = StatsReporter.stats_dict[self.category][key] if len(stat_values) == 0: return StatsSummary.empty() return StatsSummary( mean=np.mean(stat_values), std=np.std(stat_values), num=len(stat_values), sum=np.sum(stat_values), aggregation_method=StatsReporter.stats_aggregation[self.category][key], )