def experiment_main(opt_class, args=None): # pragma: main
"""This is in effect the `main` routine for this experiment. However, it is called from the optimizer wrapper file
so the class can be passed in. The optimizers are assumed to be outside the package, so the optimizer class can't
be named from inside the main function without using hacky stuff like `eval`.
"""
if args is None:
description = "Run a study with one benchmark function and an optimizer"
args = cmd.parse_args(cmd.experiment_parser(description))
args[CmdArgs.opt_rev] = opt_class.get_version()
run_uuid = uuid.UUID(args[CmdArgs.uuid])
logging.captureWarnings(True)
# Setup logging to both a file and stdout (if verbose is set to True)
logger.setLevel(logging.INFO) # Note this is the module-wide logger
logfile = XRSerializer.logging_path(args[CmdArgs.db_root],
args[CmdArgs.db], run_uuid)
logger_file_handler = logging.FileHandler(logfile, mode="w")
logger.addHandler(logger_file_handler)
if args[CmdArgs.verbose]:
logger.addHandler(logging.StreamHandler())
warnings_logger = logging.getLogger("py.warnings")
warnings_logger.addHandler(logger_file_handler)
if args[CmdArgs.verbose]:
warnings_logger.addHandler(logging.StreamHandler())
logger.info("running: %s" % str(cmd.serializable_dict(args)))
logger.info("cmd: %s" % cmd.cmd_str())
assert (args[CmdArgs.metric]
in METRICS_LOOKUP[get_problem_type(args[CmdArgs.data])]
), "reg/clf metrics can only be used on compatible dataset"
# Setup random streams for computing the signature, must use same seed
# across all runs to ensure signature is consistent. This seed is random:
_setup_seeds(
"7e9f2cabb0dd4f44bc10cf18e440b427") # pragma: allowlist secret
signature = get_objective_signature(args[CmdArgs.classifier],
args[CmdArgs.data],
args[CmdArgs.metric],
data_root=args[CmdArgs.data_root])
logger.info("computed signature: %s" % str(signature))
opt_kwargs = load_optimizer_kwargs(args[CmdArgs.optimizer],
args[CmdArgs.optimizer_root])
# Setup the call back for intermediate logging
if cc.BASELINE not in XRSerializer.get_derived_keys(args[CmdArgs.db_root],
db=args[CmdArgs.db]):
warnings.warn("Baselines not found. Will not log intermediate scores.")
callback = None
else:
test_case_str = SklearnModel.test_case_str(args[CmdArgs.classifier],
args[CmdArgs.data],
args[CmdArgs.metric])
optimizer_str = str_join_safe(
ARG_DELIM, (args[CmdArgs.optimizer], args[CmdArgs.opt_rev],
args[CmdArgs.rev]))
baseline_ds, baselines_meta = XRSerializer.load_derived(
args[CmdArgs.db_root], db=args[CmdArgs.db], key=cc.BASELINE)
# Check the objective function signatures match in the baseline file
sig_errs, _ = analyze_signature_pair({test_case_str: signature[1]},
baselines_meta["signature"])
logger.info("Signature errors:\n%s" % sig_errs.to_string())
print(json.dumps({"exp sig errors": sig_errs.T.to_dict()}))
def log_mean_score_json(evals, iters):
assert evals.shape == (len(OBJECTIVE_NAMES), )
assert not np.any(np.isnan(evals))
log_msg = {
cc.TEST_CASE: test_case_str,
cc.METHOD: optimizer_str,
cc.TRIAL: args[CmdArgs.uuid],
cc.ITER: iters,
}
for idx, obj in enumerate(OBJECTIVE_NAMES):
assert OBJECTIVE_NAMES[idx] == obj
# Extract relevant rescaling info
slice_ = {cc.TEST_CASE: test_case_str, OBJECTIVE: obj}
best_opt = baseline_ds[cc.PERF_BEST].sel(
slice_, drop=True).values.item()
base_clip_val = baseline_ds[cc.PERF_CLIP].sel(
slice_, drop=True).values.item()
# Perform the same rescaling as found in experiment_analysis.compute_aggregates()
score = linear_rescale(evals[idx],
best_opt,
base_clip_val,
0.0,
1.0,
enforce_bounds=False)
# Also, clip the score from below at -1 to limit max influence of single run on final average
score = np.clip(score, -1.0, 1.0)
score = score.item() # Make easiest for logging in JSON
assert isinstance(score, float)
# Note: This is not the raw score but the rescaled one!
log_msg[obj] = score
log_msg = json.dumps(log_msg)
print(log_msg, flush=True)
# One second safety delay to protect against subprocess stdout getting lost
sleep(1)
callback = log_mean_score_json
# Now set the seeds for the actual experiment
_setup_seeds(args[CmdArgs.uuid])
# Now do the experiment
logger.info("starting sklearn study %s %s %s %s %d %d" % (
args[CmdArgs.optimizer],
args[CmdArgs.classifier],
args[CmdArgs.data],
args[CmdArgs.metric],
args[CmdArgs.n_calls],
args[CmdArgs.n_suggest],
))
logger.info("with data root: %s" % args[CmdArgs.data_root])
function_evals, timing, suggest_log = run_sklearn_study(
opt_class,
opt_kwargs,
args[CmdArgs.classifier],
args[CmdArgs.data],
args[CmdArgs.metric],
args[CmdArgs.n_calls],
args[CmdArgs.n_suggest],
data_root=args[CmdArgs.data_root],
callback=callback,
)
# Curate results into clean dataframes
eval_ds = build_eval_ds(function_evals, OBJECTIVE_NAMES)
time_ds = build_timing_ds(*timing)
suggest_ds = build_suggest_ds(suggest_log)
# setup meta:
meta = {"args": cmd.serializable_dict(args), "signature": signature}
logger.info("saving meta data: %s" % str(meta))
# Now the final IO to export the results
logger.info("saving results")
XRSerializer.save(eval_ds,
meta,
args[CmdArgs.db_root],
db=args[CmdArgs.db],
key=cc.EVAL,
uuid_=run_uuid)
logger.info("saving timing")
XRSerializer.save(time_ds,
meta,
args[CmdArgs.db_root],
db=args[CmdArgs.db],
key=cc.TIME,
uuid_=run_uuid)
logger.info("saving suggest log")
XRSerializer.save(suggest_ds,
meta,
args[CmdArgs.db_root],
db=args[CmdArgs.db],
key=cc.SUGGEST_LOG,
uuid_=run_uuid)
logger.info("done")
def experiment_main(opt_class, args=None): # pragma: main
"""This is in effect the `main` routine for this experiment. However, it is called from the optimizer wrapper file
so the class can be passed in. The optimizers are assumed to be outside the package, so the optimizer classs can't
be named from inside the main function without using hacky stuff like `eval`.
"""
if args is None:
description = "Run a study with one benchmark function and an optimizer"
args = cmd.parse_args(cmd.experiment_parser(description))
args[CmdArgs.opt_rev] = opt_class.get_version()
run_uuid = uuid.UUID(args[CmdArgs.uuid])
logging.captureWarnings(True)
# Setup logging to both a file and stdout (if verbose is set to True)
logger.setLevel(logging.INFO) # Note this is the module-wide logger
logfile = XRSerializer.logging_path(args[CmdArgs.db_root],
args[CmdArgs.db], run_uuid)
logger_file_handler = logging.FileHandler(logfile, mode="w")
logger.addHandler(logger_file_handler)
if args[CmdArgs.verbose]:
logger.addHandler(logging.StreamHandler())
warnings_logger = logging.getLogger("py.warnings")
warnings_logger.addHandler(logger_file_handler)
if args[CmdArgs.verbose]:
warnings_logger.addHandler(logging.StreamHandler())
logger.info("running: %s" % str(cmd.serializable_dict(args)))
logger.info("cmd: %s" % cmd.cmd_str())
assert (args[CmdArgs.metric]
in METRICS_LOOKUP[get_problem_type(args[CmdArgs.data])]
), "reg/clf metrics can only be used on compatible dataset"
# Setup random streams for computing the signature, must use same seed
# across all runs to ensure signature is consistent. This seed is random:
_setup_seeds(
"7e9f2cabb0dd4f44bc10cf18e440b427") # pragma: allowlist secret
signature = get_objective_signature(args[CmdArgs.classifier],
args[CmdArgs.data],
args[CmdArgs.metric],
data_root=args[CmdArgs.data_root])
logger.info("computed signature: %s" % str(signature))
opt_kwargs = load_optimizer_kwargs(args[CmdArgs.optimizer],
args[CmdArgs.optimizer_root])
# Now set the seeds for the actual experiment
_setup_seeds(args[CmdArgs.uuid])
# Now do the experiment
logger.info("starting sklearn study %s %s %s %s %d %d" % (
args[CmdArgs.optimizer],
args[CmdArgs.classifier],
args[CmdArgs.data],
args[CmdArgs.metric],
args[CmdArgs.n_calls],
args[CmdArgs.n_suggest],
))
logger.info("with data root: %s" % args[CmdArgs.data_root])
function_evals, timing = run_sklearn_study(
opt_class,
opt_kwargs,
args[CmdArgs.classifier],
args[CmdArgs.data],
args[CmdArgs.metric],
args[CmdArgs.n_calls],
args[CmdArgs.n_suggest],
data_root=args[CmdArgs.data_root],
)
# Curate results into clean dataframes
eval_ds = build_eval_ds(function_evals)
time_ds = build_timing_ds(*timing)
# setup meta:
meta = {"args": cmd.serializable_dict(args), "signature": signature}
logger.info("saving meta data: %s" % str(meta))
# Now the final IO to export the results
logger.info("saving results")
XRSerializer.save(eval_ds,
meta,
args[CmdArgs.db_root],
db=args[CmdArgs.db],
key=cc.EVAL,
uuid_=run_uuid)
logger.info("saving timing")
XRSerializer.save(time_ds,
meta,
args[CmdArgs.db_root],
db=args[CmdArgs.db],
key=cc.TIME,
uuid_=run_uuid)
logger.info("done")