def info(
log_dirs: List[Path] = ["~/logs/compiler_gym/llvm_autotuning"],
all_runs: bool = False,
group_by_working_directory: bool = False,
only_nonzero_reward: bool = False,
):
experiments = experiments_from_paths(log_dirs)
results = []
for experiment in experiments:
df = experiment.dataframe
# Exclude runs where reward was zero, used for pruning false results if
# the environment is flaky or can fail.
if only_nonzero_reward:
df = df[df.reward != 0]
if not len(df):
continue
df.to_csv(experiment.working_directory / "results.csv", index=False)
walltimes = df[["benchmark", "walltime"]].groupby("benchmark").mean()
rewards = df[["benchmark", "reward"]].groupby("benchmark").agg(geometric_mean)
num_results = len(df)
num_benchmarks = len(set(df["benchmark"]))
df = pd.concat((walltimes, rewards), axis=1)
avg_walltime = df["walltime"].mean()
avg_reward = geometric_mean(df["reward"])
df = pd.concat(
(
df,
pd.DataFrame(
[{"walltime": avg_walltime, "reward": avg_reward}],
index=["Average"],
),
)
)
df = df.reset_index()
df.insert(0, "config", experiment.configuration_number)
df.insert(0, "timestamp", experiment.timestamp)
df.insert(0, "experiment", experiment.experiment)
if all_runs:
print(experiment.working_directory)
print(tabulate(df, showindex=False, headers="keys", tablefmt="grid"))
print()
results.append(
{
"working_directory": experiment.working_directory,
"experiment": experiment.experiment,
"timestamp": experiment.timestamp,
"config": experiment.configuration_number,
"num_benchmarks": num_benchmarks,
"num_results": num_results,
"walltime": avg_walltime,
"reward": avg_reward,
}
)
df = pd.DataFrame(results)
if not len(df):
print("No results")
return
print("---------------------------------------")
print("Aggregate over experiments:")
if group_by_working_directory:
df = df.groupby(["working_directory"]).mean()
else:
df = df.groupby(["experiment", "timestamp", "config"]).mean()
# Cast float back to int.
df["num_benchmarks"] = [int(x) for x in df["num_benchmarks"]]
df["num_results"] = [int(x) for x in df["num_results"]]
# Better column names.
df = df.rename(columns={"reward": "geomean_reward", "walltime": "walltime (s)"})
pd.set_option("display.max_rows", None)
print(df)
def train(log_dirs: List[Path] = ["~/logs/compiler_gym/llvm_rl"]):
init_logging()
models = models_from_paths(log_dirs)
dfs = []
for model in models:
df = model.dataframe
if not len(df):
continue
# Select only the rows with a checkpoint.
df = df[df["checkpoint"].values]
df = df[[
"trial_name",
"experiment_timestamp",
"episodes_total",
"episode_reward_geomean",
"episode_reward_mean",
"evaluation/episode_reward_mean",
"evaluation/episode_reward_geomean",
"time_total_s",
"complete",
"cpus",
"gpus",
]]
sdf = df.groupby(
["experiment", "config", "replica", "experiment_timestamp"]).max()
test_results = model.test_dataframes
sdf["test_results"] = [
test_results.get(d, pd.DataFrame()) for d in sdf["trial_name"]
]
sdf["test_ic_mean"] = [
sum(d["instruction_count_reduction"]) /
len(d) if not d.empty else float("nan")
for d in sdf["test_results"]
]
sdf["test_ic_geomean"] = [
geometric_mean(d["instruction_count_reduction"])
if not d.empty else float("nan") for d in sdf["test_results"]
]
sdf["test_os_mean"] = [
sum(d["object_size_reduction"]) /
len(d) if not d.empty else float("nan")
for d in sdf["test_results"]
]
sdf["test_os_geomean"] = [
geometric_mean(d["object_size_reduction"])
if not d.empty else float("nan") for d in sdf["test_results"]
]
sdf["test_checkpoint"] = [
int(d["test_checkpoint"].values[0].split("-")[-1])
if not d.empty else "" for d in sdf["test_results"]
]
dfs.append(sdf.reset_index())
df = pd.concat(dfs)
# Print everything.
pd.set_option("display.max_columns", None)
pd.set_option("display.max_rows", None)
pd.set_option("display.width", None)
df = df.rename(
columns={
"experiment_timestamp": "timestamp",
"episodes_total": "episodes",
"evaluation/episode_reward_geomean": "val_geomean",
"evaluation/episode_reward_mean": "val_mean",
"episode_reward_mean": "train_mean",
"episode_reward_geomean": "train_geomean",
"time_total_s": "training_time",
"test_reward_mean": "test_mean",
"test_reward_geomean": "test_geomean",
})
# Format for printing.
df["complete"] = [f"{x:.1%}" for x in df["complete"]]
df["episodes"] = [f"{int(x):,d}" for x in df["episodes"]]
df["training_time"] = [
humanize.naturaldelta(x) for x in df["training_time"]
]
for reward in [
"train_mean",
"train_geomean",
"val_mean",
"val_geomean",
"test_ic_geomean",
"test_os_geomean",
"test_ic_mean",
"test_os_mean",
]:
df[reward] = [f"{x:.4f}" for x in df[reward].values]
df = df[[
"trial_name",
"timestamp",
"complete",
"episodes",
"training_time",
"test_checkpoint",
"train_geomean",
"val_geomean",
]]
print(tabulate(df, headers="keys", showindex=False, tablefmt="psql"))
def reward_aggregation(a):
return geometric_mean(np.clip(a, 0, None))
def main(argv):
assert len(argv) == 1, f"Unknown args: {argv[:1]}"
assert FLAGS.n > 0, "n must be > 0"
with gym.make("llvm-ic-v0") as env:
# Stream verbose CompilerGym logs to file.
logger = logging.getLogger("compiler_gym")
logger.setLevel(logging.DEBUG)
log_handler = logging.FileHandler(FLAGS.leaderboard_logfile)
logger.addHandler(log_handler)
logger.propagate = False
print(f"Writing results to {FLAGS.leaderboard_results}")
print(f"Writing logs to {FLAGS.leaderboard_logfile}")
# Build the list of benchmarks to evaluate.
benchmarks = env.datasets[FLAGS.test_dataset].benchmark_uris()
if FLAGS.max_benchmarks:
benchmarks = islice(benchmarks, FLAGS.max_benchmarks)
benchmarks = list(benchmarks)
# Repeat the searches for the requested number of iterations.
benchmarks *= FLAGS.n
total_count = len(benchmarks)
# If we are resuming from a previous job, read the states that have
# already been proccessed and remove those benchmarks from the list
# of benchmarks to evaluate.
init_states = []
if FLAGS.resume and Path(FLAGS.leaderboard_results).is_file():
with CompilerEnvStateReader(open(
FLAGS.leaderboard_results)) as reader:
for state in reader:
init_states.append(state)
if state.benchmark in benchmarks:
benchmarks.remove(state.benchmark)
# Run the benchmark loop in background so that we can asynchronously
# log progress.
worker = _EvalPolicyWorker(env, benchmarks, policy, init_states)
worker.start()
timer = Timer().reset()
try:
print(f"=== Evaluating policy on "
f"{humanize.intcomma(total_count)} "
f"{FLAGS.test_dataset} benchmarks ==="
"\n\n" # Blank lines will be filled below
)
while worker.is_alive():
done_count = len(worker.states)
remaining_count = total_count - done_count
time = timer.time
gmean_reward = geometric_mean(
[s.reward for s in worker.states])
mean_walltime = (arithmetic_mean(
[s.walltime for s in worker.states]) or time)
print(
"\r\033[2A"
"\033[K"
f"Runtime: {humanize_duration_hms(time)}. "
f"Estimated completion: {humanize_duration_hms(mean_walltime * remaining_count)}. "
f"Completed: {humanize.intcomma(done_count)} / {humanize.intcomma(total_count)} "
f"({done_count / total_count:.1%})."
"\n\033[K"
f"Current mean walltime: {mean_walltime:.3f}s / benchmark."
"\n\033[K"
f"Current geomean reward: {gmean_reward:.4f}.",
flush=True,
end="",
)
sleep(1)
except KeyboardInterrupt:
print("\nkeyboard interrupt", flush=True)
worker.alive = False
# User interrupt, don't validate.
FLAGS.validate = False
if FLAGS.validate:
FLAGS.env = "llvm-ic-v0"
validate(["argv0", FLAGS.leaderboard_results])
def test_geometric_mean_123():
assert geometric_mean([1, 2, 3]) == approx(1.8171205928321)
def test_geometric_mean_negative():
assert geometric_mean([-1, 1, 2]) == 0
def test_geometric_mean_zero_value():
assert geometric_mean([0, 1, 2]) == 0
def test_geometric_mean_empty_list():
assert geometric_mean([]) == 0
def _trial_to_dataframe(self, directory: Path) -> Optional[pd.DataFrame]:
components = directory.name.split("-")
if len(components) < 3:
logger.warning(
"Directory name does not match expected "
"{experiment}-{config}-{replica} format: %s",
directory,
)
return
replica = components[-1]
config = components[-2]
experiment = "-".join(components[:-2])
if not (directory / "progress.csv").is_file():
logger.warning("File not found: %s", directory / "progress.csv")
return
try:
df = pd.read_csv(directory / "progress.csv")
except pd.errors.EmptyDataError:
return None
df.insert(0, "logsdir", str(directory))
df.insert(
0,
"experiment_timestamp",
" ".join([
self.working_directory.parent.parent.name,
self.working_directory.parent.name,
]),
)
df.insert(0, "trial_name", directory.name)
df.insert(0, "replica", replica)
df.insert(0, "config", config)
df.insert(0, "experiment", experiment)
df["checkpoint"] = [(directory / f"checkpoint_{i:06d}").is_dir()
for i in df["training_iteration"]]
df["checkpoint_path"] = [
str(directory / f"checkpoint_{i:06d}" / f"checkpoint-{i}") if
(directory / f"checkpoint_{i:06d}").is_dir() else None
for i in df["training_iteration"]
]
df["evaluation/episode_reward_geomean"] = [
geometric_mean(eval(x))
for x in df["evaluation/hist_stats/episode_reward"]
]
df["episode_reward_geomean"] = [
geometric_mean(eval(x)) for x in df["hist_stats/episode_reward"]
]
df["complete"] = [
min(d / self.training.episodes, 1) for d in df["episodes_total"]
]
df["cpus"] = self.executor.cpus
df["gpus"] = self.executor.gpus
df = df.set_index(
["experiment", "config", "replica", "training_iteration"])
return df
