def _setup_seeds(hex_str): # pragma: main
"""This function should only be called from main. Be careful with this function as it manipulates the global random
streams.
This is part of the general experiment setup before a study.
If torch becomes used in any of our optimizers then this will need to come back, could also do TF seed init.
```
torch.manual_seed(random_seed(master_stream))
if torch.cuda.is_available():
torch.cuda.manual_seed(random_seed(master_stream))
```
"""
# Set all random seeds: avoid correlated streams ==> must use diff seeds.
# Could use UUID class, but more direct to just convert the hex to py int.
# pyrandom is better for master because it is not limited to 32-bit seeds.
master_stream = pyrandom.Random(int(hex_str, 16))
pyrandom.seed(random_seed(master_stream))
np.random.seed(random_seed(master_stream))
def _hyperopt_suggest(self):
new_ids = self.trials.new_trial_ids(1)
assert len(new_ids) == 1
self.trials.refresh()
seed = random_seed(self.random)
new_trials = tpe.suggest(new_ids, self.domain, self.trials, seed)
assert len(new_trials) == 1
self.trials.insert_trial_docs(new_trials)
self.trials.refresh()
new_trial, = new_trials # extract singleton
return new_trial
def main():
"""See README for instructions on calling launcher.
"""
description = "Launch series of studies across functions and optimizers"
args = cmd.parse_args(cmd.launcher_parser(description))
logger.setLevel(logging.INFO) # Note this is the module-wide logger
if args[CmdArgs.verbose]:
logger.addHandler(logging.StreamHandler())
# Get optimizer settings, says which file to call for each optimizer
settings = cmd.load_optimizer_settings(args[CmdArgs.optimizer_root])
opt_file_lookup = {
optimizer: wrapper_file
for optimizer, (wrapper_file, _) in settings.items()
}
# Setup uuid
if args[CmdArgs.uuid] is None:
args[CmdArgs.uuid] = pyuuid.uuid4(
).hex # debatable if uuid1 or uuid4 is better here
else:
warnings.warn(
"User UUID supplied. This is only desired for debugging. Careless use could lead to study id conflicts.",
UserWarning,
)
run_uuid = pyuuid.UUID(hex=args[CmdArgs.uuid])
assert run_uuid.hex == args[CmdArgs.uuid]
logger.info("Supply --uuid %s to reproduce this run." % run_uuid.hex)
# Log all the options
print("Launcher options (JSON):\n")
print(json.dumps({"bayesmark-launch-args": cmd.serializable_dict(args)}))
print("\n")
# Set the master seed (derive from the uuid we just setup)
pyrandom.seed(run_uuid.int)
np.random.seed(random_seed(pyrandom))
# Now run it, either to dry run file or executes sub-processes
if args[CmdArgs.dry_run]:
with absopen(args[CmdArgs.jobs_file], "w") as fp:
dry_run(args, opt_file_lookup, run_uuid, fp)
else:
timeout = args[CmdArgs.timeout] if args[CmdArgs.timeout] > 0 else None
real_run(args, opt_file_lookup, run_uuid, timeout)
logger.info("done")
def _suggest(self):
"""Helper function to `suggest` that does the work of calling
`hyperopt` via its dumb API.
"""
new_ids = self.trials.new_trial_ids(1)
assert len(new_ids) == 1
self.trials.refresh()
seed = random_seed(self.random)
new_trials = tpe.suggest(new_ids, self.domain, self.trials, seed)
assert len(new_trials) == 1
self.trials.insert_trial_docs(new_trials)
self.trials.refresh()
new_trial, = new_trials # extract singleton
return new_trial
def test_random_seed(seed):
random = np.random.RandomState(seed)
seed = np_util.random_seed(random)
