def compare(self, runs_directories_a, runs_directories_b):
runs_a = {
self.suffix(directory): MonitorV2.load_results(directory)
for directory in runs_directories_a
}
runs_b = {
self.suffix(directory): MonitorV2.load_results(directory)
for directory in runs_directories_b
}
f, (ax1, ax2) = plt.subplots(1, 2, sharey=True)
self.plot_all(runs_a,
field='episode_rewards',
title='rewards',
axes=ax1)
self.plot_all(runs_b,
field='episode_rewards',
title='rewards',
axes=ax2)
plt.show()
def get_run_dataframe(self,
directory,
agent_name='',
gamma=0.95,
subsample=10):
run_data = MonitorV2.load_results(directory)
if not run_data:
return pd.DataFrame()
# Common fields
data = {
"episode":
np.arange(np.size(run_data["episode_rewards"])),
"total reward":
run_data["episode_rewards"],
"discounted rewards": [
np.sum([episode[t] * gamma**t for t in range(len(episode))])
for episode in run_data["episode_rewards_"]
],
"length":
run_data["episode_lengths"],
}
# Additional highway-env fields
try:
dt = 1.0
data.update({
"crashed":
[np.any(episode) for episode in run_data["episode_crashed"]],
"velocity":
[np.mean(episode) for episode in run_data["episode_velocity"]],
"distance": [
np.sum(episode) * dt
for episode in run_data["episode_velocity"]
],
})
except KeyError as e:
print(e)
# Tags
df = pd.DataFrame(data)
df["run"] = str(directory.name)
df["agent"] = agent_name
# Filtering
for field in [
"total reward", "discounted rewards", "length", "crashed",
"velocity", "distance"
]:
df[field] = df[field].rolling(subsample).mean()
# Subsample
df = df.iloc[self.episodes_range[0]:self.episodes_range[1]:subsample]
return df
def analyze(self, run_directories):
runs = {self.suffix(directory): MonitorV2.load_results(directory) for directory in run_directories}
runs = {key: value for (key, value) in runs.items() if value is not None}
self.plot_all(runs, field='episode_rewards', title='rewards')
self.histogram_all(runs, field='episode_rewards', title='rewards')
self.describe_all(runs, field='episode_rewards', title='rewards')
self.histogram_all(runs, field='episode_lengths', title='lengths')
self.describe_all(runs, field='episode_lengths', title='lengths')
self.histogram_all(runs, field='episode_costs', title='costs', preprocess=lambda c: [sum(e) for e in c])
self.describe_all(runs, field='episode_costs', title='costs', preprocess=lambda c: [sum(e) for e in c])
self.compare(runs)
plt.show()