1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309 | """Argument parser and logic for the ``all_plots`` sub-command."""
from __future__ import annotations
import argparse
from copy import deepcopy
from pathlib import Path
import pickle
import warnings
import numpy as np
try:
from matplotlib import pyplot as plt
import pytablewriter # ty:ignore[unresolved-import]
import seaborn # ty:ignore[unresolved-import]
_PLOT_DEPS_AVAILABLE = True
except ImportError:
_PLOT_DEPS_AVAILABLE = False
try:
import pandas as pd
from scipy.spatial import distance_matrix # ty:ignore[unresolved-import]
_SCIPY_AVAILABLE = True
except ImportError:
_SCIPY_AVAILABLE = False
try:
from rliable import library as rly # ty:ignore[unresolved-import]
from rliable import metrics, plot_utils # ty:ignore[unresolved-import]
_RLIABLE_AVAILABLE = True
except ImportError:
_RLIABLE_AVAILABLE = False
rly = None
try:
from lit_sb3.core.plots.score_normalization import normalize_score
_SCORE_NORM_AVAILABLE = True
except ImportError:
_SCORE_NORM_AVAILABLE = False
def _add_all_plots_args(parser: argparse.ArgumentParser) -> None:
parser.add_argument("-a", "--algos", help="Algorithms to include", nargs="+", type=str)
parser.add_argument("-e", "--env", help="Environments to include", nargs="+", type=str)
parser.add_argument("-f", "--exp-folders", help="Folders to include", nargs="+", type=str)
parser.add_argument("-l", "--labels", help="Label for each folder", nargs="+", type=str)
parser.add_argument(
"-k",
"--key",
help=(
"Key from the `evaluations.npz` file to use to aggregate results "
"(e.g. reward, success rate, ...), it is 'results' by default (i.e., the episode reward)"
),
default="results",
type=str,
)
parser.add_argument(
"-max", "--max-timesteps", help="Max number of timesteps to display", type=int, default=int(2e6)
)
parser.add_argument(
"-min",
"--min-timesteps",
help="Min number of timesteps to keep a trial",
type=int,
default=-1,
)
parser.add_argument(
"-o",
"--output",
help="Output filename (pickle file), where to save the post-processed data",
type=str,
)
parser.add_argument(
"-median", "--median", action="store_true", default=False, help="Display median instead of mean in the table"
)
parser.add_argument("--no-million", action="store_true", default=False, help="Do not convert x-axis to million")
parser.add_argument("--no-display", action="store_true", default=False, help="Do not show the plots")
parser.add_argument(
"-print",
"--print-n-trials",
action="store_true",
default=False,
help="Print the number of trial for each result",
)
def run_all_plots(args: argparse.Namespace) -> None: # noqa: C901
"""Execute the all_plots command with pre-parsed args."""
if not _PLOT_DEPS_AVAILABLE:
raise ImportError(
"matplotlib, seaborn and pytablewriter are required for plotting. "
"Install them with: pip install matplotlib seaborn pytablewriter"
)
# Activate seaborn
seaborn.set()
results: dict = {}
post_processed_results: dict = {}
args.algos = [algo.upper() for algo in args.algos]
if args.labels is None:
args.labels = args.exp_folders
for env in args.env:
plt.figure(f"Results {env}")
plt.title(f"{env}", fontsize=14)
x_label_suffix = "" if args.no_million else "(in Million)"
plt.xlabel(f"Timesteps {x_label_suffix}", fontsize=14)
plt.ylabel("Score", fontsize=14)
results[env] = {}
post_processed_results[env] = {}
for algo in args.algos:
for folder_idx, exp_folder in enumerate(args.exp_folders):
log_path = Path(exp_folder) / algo.lower()
if not Path(log_path).is_dir():
continue
results[env][f"{args.labels[folder_idx]}-{algo}"] = 0.0
dirs = [
Path(log_path) / d
for d in list(Path(log_path).iterdir())
if (env in d.name and (Path(log_path) / d).is_dir())
]
max_len = 0
merged_timesteps, merged_results = [], []
last_eval = []
timesteps = np.empty(0)
for _, dir_ in enumerate(dirs):
try:
log = np.load(Path(dir_) / "evaluations.npz")
except FileNotFoundError:
print("Eval not found for", dir_)
continue
mean_ = np.squeeze(log["results"].mean(axis=1))
if mean_.shape == ():
continue
max_len = max(max_len, len(mean_))
if len(log["timesteps"]) >= max_len:
timesteps = log["timesteps"]
# For post-processing
merged_timesteps.append(log["timesteps"])
merged_results.append(log[args.key])
# Truncate the plots
while timesteps[max_len - 1] > args.max_timesteps:
max_len -= 1
timesteps = timesteps[:max_len]
if len(log[args.key]) >= max_len:
last_eval.append(log[args.key][max_len - 1])
else:
last_eval.append(log[args.key][-1])
# Merge runs with different eval freq
if args.min_timesteps > 0 and _SCIPY_AVAILABLE:
min_ = np.inf
for n_timesteps in merged_timesteps:
if n_timesteps[-1] >= args.min_timesteps:
min_ = min(min_, len(n_timesteps))
if len(n_timesteps) == min_:
max_len = len(n_timesteps)
# Truncate the plots
while n_timesteps[max_len - 1] > args.max_timesteps:
max_len -= 1
timesteps = n_timesteps[:max_len]
# Avoid modifying original aggregated results
merged_results_ = deepcopy(merged_results)
# Downsample if needed
for trial_idx, n_timesteps in enumerate(merged_timesteps):
if len(n_timesteps) == min_ or n_timesteps[-1] < args.min_timesteps:
pass
else:
new_merged_results = []
# Nearest neighbour
distance_mat = distance_matrix(n_timesteps.reshape(-1, 1), timesteps.reshape(-1, 1))
closest_indices = distance_mat.argmin(axis=0)
for closest_idx in closest_indices:
new_merged_results.append(merged_results_[trial_idx][closest_idx])
merged_results[trial_idx] = new_merged_results
last_eval[trial_idx] = merged_results_[trial_idx][closest_indices[-1]]
# Remove incomplete runs
merged_results_tmp, last_eval_tmp = [], []
for idx in range(len(merged_results)):
if len(merged_results[idx]) >= max_len:
merged_results_tmp.append(merged_results[idx][:max_len])
last_eval_tmp.append(last_eval[idx])
merged_results = merged_results_tmp
last_eval = last_eval_tmp
# Post-process
if len(merged_results) > 0:
# shape: (n_trials, n_eval * n_eval_episodes)
merged_results = np.array(merged_results)
n_trials = len(merged_results)
n_eval = len(timesteps)
if args.print_n_trials:
print(f"{env}-{algo}-{args.labels[folder_idx]}: {n_trials}")
# reshape to (n_trials, n_eval, n_eval_episodes)
evaluations = merged_results.reshape((n_trials, n_eval, -1))
# re-arrange to (n_eval, n_trials, n_eval_episodes)
evaluations = np.swapaxes(evaluations, 0, 1)
# (n_eval,)
mean_ = np.mean(evaluations, axis=(1, 2))
# (n_eval, n_trials)
mean_per_eval = np.mean(evaluations, axis=-1)
# (n_eval,)
std_ = np.std(mean_per_eval, axis=-1)
# std: error
std_error = std_ / np.sqrt(n_trials)
# Take last evaluation
# shape: (n_trials, n_eval_episodes) to (n_trials,)
last_evals = np.array(last_eval).squeeze().mean(axis=-1)
# Standard deviation of the mean performance for the last eval
std_last_eval = np.std(last_evals)
# Compute standard error
std_error_last_eval = std_last_eval / np.sqrt(n_trials)
if args.median:
results[env][f"{algo}-{args.labels[folder_idx]}"] = f"{np.median(last_evals):.0f}"
else:
results[env][f"{algo}-{args.labels[folder_idx]}"] = (
f"{np.mean(last_evals):.0f} +/- {std_error_last_eval:.0f}"
)
# x axis in Millions of timesteps
divider = 1e6
if args.no_million:
divider = 1.0
post_processed_results[env][f"{algo}-{args.labels[folder_idx]}"] = {
"timesteps": timesteps,
"mean": mean_,
"std_error": std_error,
"last_evals": last_evals,
"std_error_last_eval": std_error_last_eval,
"mean_per_eval": mean_per_eval,
}
plt.plot(timesteps / divider, mean_, label=f"{algo}-{args.labels[folder_idx]}", linewidth=3)
plt.fill_between(timesteps / divider, mean_ + std_error, mean_ - std_error, alpha=0.5)
plt.legend()
# Markdown Table
writer = pytablewriter.MarkdownTableWriter(max_precision=3)
writer.table_name = "results_table"
headers = ["Environments"]
# One additional row for the subheader
value_matrix = [[] for _ in range(len(args.env) + 1)]
headers = ["Environments"]
# Header and sub-header
value_matrix[0].append("")
for algo in args.algos:
for label in args.labels:
value_matrix[0].append(label)
headers.append(algo)
writer.headers = headers
for i, env in enumerate(args.env, start=1):
value_matrix[i].append(env)
for algo in args.algos:
for label in args.labels:
key = f"{algo}-{label}"
value_matrix[i].append(f"{results[env].get(key, '0.0 +/- 0.0')}")
writer.value_matrix = value_matrix
writer.write_table()
post_processed_results["results_table"] = {"headers": headers, "value_matrix": value_matrix}
if args.output is not None:
print(f"Saving to {args.output}.pkl")
with Path(f"{args.output}.pkl").open("wb") as file_handler:
pickle.dump(post_processed_results, file_handler)
if not args.no_display:
plt.show()
def all_plots() -> None:
"""Gather, aggregate and plot results from multiple experiments (parse args then run)."""
parser = argparse.ArgumentParser("Gather results, plot them and create table")
_add_all_plots_args(parser)
args = parser.parse_args()
run_all_plots(args)
|