def GatherExperiments(
folders,
algo,
window=40,
title="",
min_num_x=-1,
timesteps=False,
output_file="",
):
"""
Compute mean and standard error for several experiments and plot the learning curve
:param folders: ([str]) Log folders, where the monitor.csv are stored
:param window: (int) Smoothing window
:param algo: (str) name of the RL algo
:param title: (str) plot title
:param min_num_x: (int) Minimum number of episode/timesteps to keep an experiment (default: -1, no minimum)
:param timesteps: (bool) Plot timesteps instead of episodes
:param y_limits: ([float]) y-limits for the plot
:param output_file: (str) Path to a file where the plot data will be saved
:param no_display: (bool) Set to true, the plot won't be displayed (useful when only saving plot)
"""
y_list = []
x_list = []
ok = False
for folder in folders:
if timesteps:
x, y = loadData(folder, smooth=1, bin_size=100)
if x is not None:
x, y = np.array(x), np.array(y)
else:
x, y = loadEpisodesData(folder)
if x is None or (min_num_x > 0 and y.shape[0] < min_num_x):
printRed("Skipping {}".format(folder))
continue
if y.shape[0] <= window:
printYellow("Folder {}".format(folder))
printRed("Not enough episodes for current window size = {}".format(
window))
continue
ok = True
y = movingAverage(y, window)
y_list.append(y)
print(len(x))
# Truncate x
x = x[len(x) - len(y):]
x_list.append(x)
if not ok:
printRed("Not enough data to plot anything with current config." +
" Consider decreasing --min-x")
return
lengths = list(map(len, x_list))
min_x, max_x = np.min(lengths), np.max(lengths)
print("Min x: {}".format(min_x))
print("Max x: {}".format(max_x))
for i in range(len(x_list)):
x_list[i] = x_list[i][:min_x]
y_list[i] = y_list[i][:min_x]
x = np.array(x_list)[0]
y = np.array(y_list)
# if output_file != "":
# printGreen("Saving aggregated data to {}.npz".format(output_file))
# np.savez(output_file, x=x, y=y)
return x, y
def plotGatheredExperiments(folders,
algo,
y_limits,
window=40,
title="",
min_num_x=-1,
timesteps=False,
output_file="",
no_display=False):
"""
Compute mean and standard error for several experiments and plot the learning curve
:param folders: ([str]) Log folders, where the monitor.csv are stored
:param window: (int) Smoothing window
:param algo: (str) name of the RL algo
:param title: (str) plot title
:param min_num_x: (int) Minimum number of episode/timesteps to keep an experiment (default: -1, no minimum)
:param timesteps: (bool) Plot timesteps instead of episodes
:param y_limits: ([float]) y-limits for the plot
:param output_file: (str) Path to a file where the plot data will be saved
:param no_display: (bool) Set to true, the plot won't be displayed (useful when only saving plot)
"""
y_list = []
x_list = []
ok = False
for folder in folders:
if timesteps:
x, y = loadData(folder, smooth=1, bin_size=100)
if x is not None:
x, y = np.array(x), np.array(y)
else:
x, y = loadEpisodesData(folder)
if x is None or (min_num_x > 0 and y.shape[0] < min_num_x):
printYellow("Skipping {}".format(folder))
continue
if y.shape[0] <= window:
printYellow("Folder {}".format(folder))
printYellow(
"Not enough episodes for current window size = {}".format(
window))
continue
ok = True
y = movingAverage(y, window)
y_list.append(y)
# Truncate x
x = x[len(x) - len(y):]
x_list.append(x)
if not ok:
printRed("Not enough data to plot anything with current config." +
" Consider decreasing --min-x")
return
lengths = list(map(len, x_list))
min_x, max_x = np.min(lengths), np.max(lengths)
print("Min x: {}".format(min_x))
print("Max x: {}".format(max_x))
for i in range(len(x_list)):
x_list[i] = x_list[i][:min_x]
y_list[i] = y_list[i][:min_x]
x = np.array(x_list)[0]
y = np.array(y_list)
printGreen("{} Experiments".format(y.shape[0]))
print("Min, Max rewards:", np.min(y), np.max(y))
fig = plt.figure(title)
# Compute mean for different seeds
m = np.mean(y, axis=0)
# Compute standard error
s = np.squeeze(np.asarray(np.std(y, axis=0)))
n = y.shape[0]
plt.fill_between(x,
m - s / np.sqrt(n),
m + s / np.sqrt(n),
color=lightcolors[0])
plt.plot(x, m, color=darkcolors[0], label=algo, linewidth=1)
if timesteps:
formatter = FuncFormatter(millions)
plt.xlabel('Number of Timesteps')
fig.axes[0].xaxis.set_major_formatter(formatter)
else:
plt.xlabel('Number of Episodes')
plt.ylabel('Rewards')
plt.title(title, **fontstyle)
plt.ylim(y_limits)
plt.legend(framealpha=0.5,
labelspacing=0.01,
loc='lower right',
fontsize=16)
if output_file != "":
printGreen("Saving aggregated data to {}.npz".format(output_file))
np.savez(output_file, x=x, y=y)
if not no_display:
plt.show()