Files
Gymnasium/gym/scoreboard/scoring.py
Greg Brockman 934b2acbb7 Add benchmark support (#338)
* Warn if seed doesn't return a list

* Add preliminary BenchmarkRun support

* Add experimental benchmark registration

* Flesh out interface

* Add preliminary BenchmarkRun support

* Warn if seed doesn't return a list

* Add experimental benchmark registration

* Flesh out interface

* Make benchmarkrun upload recursive

* Add evaluation episodes

* Add benchmark scoring

* Tweak reward locations

* Tweak scoring

* Clear default metadata in Wrapper

* Improve scoring

* Expose registry; fix test

* Add initial_reset_timestamp

* Add back algorithm; fix tests
2016-09-23 01:04:26 -07:00

169 lines
6.9 KiB
Python

"""This is the actual code we use to score people's solutions
server-side. The interfaces here are not yet stable, but we include
them so that people can reproduce our scoring calculations
independently.
We correspondly do not currently import this module.
"""
import numpy as np
import requests
import gym
def score_from_remote(url):
result = requests.get(url)
parsed = result.json()
episode_lengths = parsed['episode_lengths']
episode_rewards = parsed['episode_rewards']
episode_types = parsed.get('episode_types')
timestamps = parsed['timestamps']
# Handle legacy entries where initial_reset_timestamp wasn't set
initial_reset_timestamp = parsed.get('initial_reset_timestamp', timestamps[0])
env_id = parsed['env_id']
spec = gym.spec(env_id)
return score_from_merged(episode_lengths, episode_rewards, episode_types, timestamps, initial_reset_timestamp, spec.trials, spec.reward_threshold)
def score_from_local(directory):
"""Calculate score from a local results directory"""
results = gym.monitoring.monitor.load_results(directory)
# No scores yet saved
if results is None:
return None
episode_lengths = results['episode_lengths']
episode_rewards = results['episode_rewards']
episode_types = results['episode_types']
timestamps = results['timestamps']
initial_reset_timestamp = results['initial_reset_timestamp']
spec = gym.spec(results['env_info']['env_id'])
return score_from_merged(episode_lengths, episode_rewards, episode_types, timestamps, initial_reset_timestamp, spec.trials, spec.reward_threshold)
def score_from_merged(episode_lengths, episode_rewards, episode_types, timestamps, initial_reset_timestamp, trials, reward_threshold):
"""Method to calculate the score from merged monitor files. Scores
only a single environment; mostly legacy.
"""
if episode_types is not None:
# Select only the training episodes
t_idx = np.where(e == 't' for e in episode_types)
episode_lengths = np.array(episode_lengths)[t_idx]
episode_rewards = np.array(episode_rewards)[t_idx]
timestamps = np.array(timestamps)[t_idx]
# Make sure everything is a float -- no pesky ints.
episode_rewards = np.array(episode_rewards, dtype='float64')
episode_t_value = timestep_t_value = mean = error = None
seconds_to_solve = seconds_in_total = None
if len(timestamps) > 0:
# This is: time from the first reset to the end of the last episode
seconds_in_total = timestamps[-1] - initial_reset_timestamp
if len(episode_rewards) >= trials:
means = running_mean(episode_rewards, trials)
if reward_threshold is not None:
# Compute t-value by finding the first index at or above
# the threshold. It comes out as a singleton tuple.
(indexes_above_threshold, ) = np.where(means >= reward_threshold)
if len(indexes_above_threshold) > 0:
# Grab the first episode index that is above the threshold value
episode_t_value = indexes_above_threshold[0]
# Find timestep corresponding to this episode
cumulative_timesteps = np.cumsum(np.insert(episode_lengths, 0, 0))
# Convert that into timesteps
timestep_t_value = cumulative_timesteps[episode_t_value]
# This is: time from the first reset to the end of the first solving episode
seconds_to_solve = timestamps[episode_t_value] - initial_reset_timestamp
# Find the window with the best mean
best_idx = np.argmax(means)
best_rewards = episode_rewards[best_idx:best_idx+trials]
mean = np.mean(best_rewards)
if trials == 1: # avoid NaN
error = 0.
else:
error = np.std(best_rewards) / (np.sqrt(trials) - 1)
return {
'episode_t_value': episode_t_value,
'timestep_t_value': timestep_t_value,
'mean': mean,
'error': error,
'number_episodes': len(episode_rewards),
'number_timesteps': sum(episode_lengths),
'seconds_to_solve': seconds_to_solve,
'seconds_in_total': seconds_in_total,
}
def benchmark_score_from_merged(benchmark, env_id, episode_lengths, episode_rewards, episode_types):
"""Method to calculate an environment's benchmark score from merged
monitor files.
"""
return benchmark.score(benchmark, env_id, episode_lengths, episode_rewards, episode_types)
def running_mean(x, N):
x = np.array(x, dtype='float64')
cumsum = np.cumsum(np.insert(x, 0, 0))
return (cumsum[N:] - cumsum[:-N]) / N
def compute_graph_stats(episode_lengths, episode_rewards, timestamps, initial_reset_timestamp, buckets):
"""Method to compute the aggregates for the graphs."""
# Not a dependency of OpenAI Gym generally.
import scipy.stats
num_episodes = len(episode_lengths)
# Catch for if no files written which causes error with scipy.stats.binned_statistic
if num_episodes == 0:
return None
episode_rewards = np.array(episode_rewards)
episode_lengths = np.array(episode_lengths)
# The index of the start of each episode
x_timestep = np.cumsum(np.insert(episode_lengths, 0, 0))[:-1]
assert len(x_timestep) == num_episodes
# Delta since the beginning of time
x_seconds = [timestamp - initial_reset_timestamp for timestamp in timestamps]
# The index of each episode
x_episode = range(num_episodes)
# Calculate the appropriate x/y statistics
x_timestep_y_reward = scipy.stats.binned_statistic(x_timestep, episode_rewards, 'median', buckets)
x_timestep_y_length = scipy.stats.binned_statistic(x_timestep, episode_lengths, 'median', buckets)
x_episode_y_reward = scipy.stats.binned_statistic(x_episode, episode_rewards, 'median', buckets)
x_episode_y_length = scipy.stats.binned_statistic(x_episode, episode_lengths, 'median', buckets)
x_seconds_y_reward = scipy.stats.binned_statistic(x_seconds, episode_rewards, 'median', buckets)
x_seconds_y_length = scipy.stats.binned_statistic(x_seconds, episode_lengths, 'median', buckets)
return {
'initial_reset_timestamp': initial_reset_timestamp,
'x_timestep_y_reward': graphable_binned_statistic(x_timestep_y_reward),
'x_timestep_y_length': graphable_binned_statistic(x_timestep_y_length),
'x_episode_y_reward': graphable_binned_statistic(x_episode_y_reward),
'x_episode_y_length': graphable_binned_statistic(x_episode_y_length),
'x_seconds_y_length': graphable_binned_statistic(x_seconds_y_length),
'x_seconds_y_reward': graphable_binned_statistic(x_seconds_y_reward),
}
def graphable_binned_statistic(binned):
x = running_mean(binned.bin_edges, 2)
y = binned.statistic
assert len(x) == len(y)
# Get rid of nasty NaNs
valid = np.logical_not(np.isnan(x)) & np.logical_not(np.isnan(y))
x = x[valid]
y = y[valid]
return {
'x': x,
'y': y,
}