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 | """Argument parser and logic for the ``benchmark`` sub-command."""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import shutil
import subprocess
import numpy as np
import lit_sb3.core.import_envs # noqa: F401
from lit_sb3.utils.model import get_saved_hyperparams, get_trained_models
from lit_sb3.utils.system import get_latest_run_id
try:
import pandas as pd
import pytablewriter # ty:ignore[unresolved-import]
_TABLE_DEPS_AVAILABLE = True
except ImportError:
_TABLE_DEPS_AVAILABLE = False
try:
from lit_sb3.core.results_plotter import load_results, ts2xy
_RESULTS_PLOTTER_AVAILABLE = True
except ImportError:
_RESULTS_PLOTTER_AVAILABLE = False
try:
from lit_sb3.huggingface.load_from_hub import download_from_hub
_HUB_DOWNLOAD_AVAILABLE = True
except ImportError:
_HUB_DOWNLOAD_AVAILABLE = False
def _add_benchmark_args(parser: argparse.ArgumentParser) -> None:
"""Add benchmark arguments to *parser*."""
parser.add_argument("--log-dir", help="Root log folder", default="rl-trained-agents/", type=str)
parser.add_argument("--benchmark-dir", help="Benchmark log folder", default="logs/benchmark/", type=str)
parser.add_argument("-n", "--n-timesteps", help="number of timesteps", default=150000, type=int)
parser.add_argument("--n-envs", help="number of environments", default=1, type=int)
parser.add_argument("--verbose", help="Verbose mode (0: no output, 1: INFO)", default=1, type=int)
parser.add_argument("--seed", help="Random generator seed", type=int, default=0)
parser.add_argument(
"--test-mode",
action="store_true",
default=False,
help="Do only one experiment (useful for testing)",
)
parser.add_argument("--with-mujoco", action="store_true", default=False, help="Run also MuJoCo envs")
parser.add_argument("--no-hub", action="store_true", default=False, help="Do not download models from hub")
parser.add_argument("--num-threads", help="Number of threads for PyTorch", default=2, type=int)
def run_benchmark(args: argparse.Namespace) -> None: # noqa: C901
"""Run benchmark evaluation over all trained models and generate a report."""
trained_models = get_trained_models(args.log_dir)
if not args.no_hub:
try:
from lit_sb3.utils.model import get_hf_trained_models
trained_models.update(get_hf_trained_models())
except (ImportError, Exception) as e:
print(f"Warning: Could not fetch HuggingFace models: {e}")
n_experiments = len(trained_models)
results: dict[str, list] = {
"algo": [],
"env_id": [],
"mean_reward": [],
"std_reward": [],
"n_timesteps": [],
"eval_timesteps": [],
"eval_episodes": [],
}
for idx, trained_model in enumerate(trained_models.keys()):
algo, env_id = trained_models[trained_model]
n_envs = args.n_envs
n_timesteps = args.n_timesteps
# HER is now a replay buffer class
if algo == "her":
continue
if algo in ["dqn", "qrdqn", "ddpg", "sac", "td3", "tqc", "crossq"]:
n_envs = 1
n_timesteps *= args.n_envs
reward_log = Path(args.benchmark_dir) / trained_model
arguments = [
"-n",
str(n_timesteps),
"--n-envs",
str(n_envs),
"-f",
args.log_dir,
"--algo",
algo,
"--env",
env_id,
"--no-render",
"--num-threads",
str(args.num_threads),
"--seed",
str(args.seed),
"--verbose",
"0",
"--reward-log",
str(reward_log),
]
if args.verbose >= 1:
print(f"{idx + 1}/{n_experiments}")
print(f"Evaluating {algo} on {env_id}...")
skip_eval = False
if Path(reward_log).is_dir() and _RESULTS_PLOTTER_AVAILABLE:
try:
x, y = ts2xy(load_results(reward_log), "timesteps")
skip_eval = len(x) > 0
except (json.JSONDecodeError, Exception): # noqa: S110
pass
# Comment out to benchmark HER robotics env — requires a mujoco licence
if "Fetch" in env_id and not args.with_mujoco:
print(f"Skipping mujoco env: {env_id}")
if not skip_eval:
continue
if skip_eval:
print("Skipping eval...")
else:
return_code = subprocess.call(["python", "-m", "lit_sb3", "enjoy", *arguments]) # noqa: S603, S607
if return_code != 0:
print("Error during evaluation, skipping...")
continue
if not _RESULTS_PLOTTER_AVAILABLE:
print("Warning: lit_sb3.core.results_plotter not available, skipping result loading")
continue
x, y = ts2xy(load_results(reward_log), "timesteps")
if not _RESULTS_PLOTTER_AVAILABLE:
continue
if len(x) > 0:
# Retrieve training timesteps from config
exp_id = get_latest_run_id(Path(args.log_dir) / algo, env_id)
log_path = Path(args.log_dir) / algo / f"{env_id}_{exp_id}" / env_id
hyperparams, _ = get_saved_hyperparams(log_path)
# Model not downloaded
if len(hyperparams) == 0 and _HUB_DOWNLOAD_AVAILABLE:
print(
"Pretrained model not found, trying to download it from sb3 Huggingface hub: https://huggingface.co/sb3"
)
exp_id = 1
log_path = Path(args.log_dir) / algo / f"{env_id}_{exp_id}" / env_id
# Auto-download
try:
download_from_hub(
algo=algo,
env_name=env_id, # type: ignore[arg-type]
exp_id=exp_id,
folder=args.log_dir,
organization="sb3",
repo_name=None,
force=False,
)
# Try again
hyperparams, _ = get_saved_hyperparams(log_path)
except Exception as e:
print(f"Failed to download: {e}")
# Hack to format it properly
if hyperparams.get("n_timesteps", 0) < 1e6:
n_training_timesteps = f"{int(hyperparams.get('n_timesteps', 0) / 1e3)}k"
else:
n_training_timesteps = f"{int(hyperparams.get('n_timesteps', 0) / 1e6)}M"
mean_reward = np.mean(y)
std_reward = np.std(y)
results["algo"].append(algo)
results["env_id"].append(env_id)
results["mean_reward"].append(mean_reward)
results["std_reward"].append(std_reward)
results["n_timesteps"].append(n_training_timesteps)
results["eval_timesteps"].append(x[-1])
results["eval_episodes"].append(len(y))
if args.verbose >= 1:
print(x[-1], "timesteps")
print(len(y), "Episodes")
print(f"Mean reward: {mean_reward:.2f} +- {std_reward:.2f}")
print()
else:
print("Not enough timesteps")
if args.test_mode:
break
if not _TABLE_DEPS_AVAILABLE:
print("Warning: pandas and pytablewriter are required to generate the benchmark table.")
print("Install them with: pip install pandas pytablewriter")
return
# Create DataFrame
results_df = pd.DataFrame(results)
# Sort results
results_df = results_df.sort_values(by=["algo", "env_id"])
header = """
## Performance of trained agents
Final performance of the trained agents can be found in the table below.
This was computed by running `python -m lit_sb3 benchmark`:
it runs the trained agent (trained on `n_timesteps`)
for `eval_timesteps` and then reports the mean episode reward
during this evaluation.
It uses the deterministic policy except for Atari games.
*NOTE: this is not a quantitative benchmark as it corresponds to only one run
(cf [issue #38](https://github.com/araffin/rl-baselines-zoo/issues/38)).
This benchmark is meant to check algorithm (maximal) performance, find potential bugs
and also allow users to have access to pretrained agents.*
"M" stands for Million (1e6)
"""
writer = pytablewriter.MarkdownTableWriter(max_precision=3)
writer.from_dataframe(results_df)
Path(args.benchmark_dir).mkdir(parents=True, exist_ok=True)
tmp_path = Path(args.benchmark_dir) / "benchmark.md"
with Path(tmp_path).open("w") as f:
f.write(header)
writer.stream = f
writer.write_table()
print(f"Results written to: {tmp_path}")
# Update root benchmark file
if not args.test_mode:
shutil.copy(tmp_path, "benchmark.md")
print("Results copied to: benchmark.md")
|