def run(ctx, args):
"""Runs an experiment
"""
print(args)
# Strip potential operation from experiment name
experiment, op_name = _strip_op_name_from_experiment(args)
# Safe load of experiment file path
try:
experiment_config_file = \
config.get_project_config()["experiments"].get(experiment)
except KeyError:
cli.error("No experiments found. "
"Are you sure you're in a Tracker project?")
# Load configuration file
experiment_config = config.load(experiment_config_file)
# Create operation object
# - Here we scan through the sourcecode
# and extract the (hyper-)parameters
op = oplib.Operation(
op_name, _op_run_dir(args),
_get_experiment_dict_by_name(experiment, experiment_config),
_op_gpus(args), args.yes)
# Prompt user to confirm run parameters
if args.yes or _confirm_run(args, experiment, op):
for n in range(args.trials):
cli.out("Trial {}/{}".format(n + 1, args.trials))
# Run the trial
_run(args, op)
def run(ctx, args):
"""Runs an experiment
"""
# Strip potential operation from experiment name
exp_name, op_name = _strip_op_name_from_experiment(args)
if op_name is None:
op_name = DEFAULT_OP
log.debug("Running experiment: '{}' with default operation: '{}' "
"as no op was provided by the user!".format(
exp_name, op_name))
# Safe load of experiment file path
try:
exp_conf_file = \
config.get_project_config()["experiments"].get(exp_name)
except KeyError:
cli.error("No experiments found. "
"Are you sure you're in a Tracker project?")
# Load configuration file
exp_conf = config.load(exp_conf_file)
# Create operation object
op = oplib.Operation(op_name, _op_run_dir(args),
_op_experiment(exp_name, exp_conf), _op_remote(args),
_op_gpus(args))
# Prompt user to confirm run parameters
if args.yes or _confirm_run(args, exp_name, op):
for n in range(args.trials):
cli.out("Trial {}/{}".format(n + 1, args.trials))
# Run the trial
_run(args, op)
def run_dir_for_id(run_id):
try:
return _path_for_id(run_id)
except NoSuchRun:
cli.out("The trial with id: '{}' was not found\n"
"Show trials by running 'tracker experiment NAME "
"--list_trials'.".format(run_id))
cli.error("No such directory", errno.ENOENT)
def _op_run_dir(args):
if args.run_dir:
run_dir = os.path.abspath(args.run_dir)
if os.getenv("NO_WARN_RUNDIR") != "1":
cli.out("Run directory is '{}' (results will not be "
"visible to Tracker)".format(run_dir))
return run_dir
else:
return None
def remote_op(op, prompt, default_resp, args):
if not args.yes:
cli.out(prompt)
if args.yes or cli.confirm("Continue?", default_resp):
try:
op()
except OperationNotSupported as e:
cli.error(e)
except OperationError as e:
cli.error(e)
def _maybe_apply_default_runs(args):
n_runs = len(args.runs)
if n_runs == 0:
raise NotImplementedError
# args.run = ("2", "1")
elif n_runs == 1:
cli.out(
"The `diff` command requires two runs.\n"
"Try specifying a second run or 'tracker diff --help' "
"for more information.")
cli.error()
elif n_runs > 2:
cli.out(
"The `diff` command cannot compare more than two runs.\n"
"Try specifying just two runs or 'tracker diff --help' "
"for more information.")
cli.error()
else:
assert n_runs == 2, args
def _run_remote(op, args):
remote = remotelib.remote_for_args(args)
try:
run_id = remote.run_op(**_run_kw(args))
except remotelib.RunFailed as e:
_handle_remote_run_failed(e, remote)
except remotelib.RemoteProcessError as e:
_handle_remote_process_error(e)
except remotelib.RemoteProcessDetached as e:
_handle_remote_process_detached(e, args.remote)
except remotelib.OperationError as e:
_handle_remote_op_error(e, remote)
except remotelib.OperationNotSupported:
cli.error("{} does not support this operation".format(remote.name))
else:
if args.background:
cli.out("{run_id} is running remotely on {remote}\n"
"To watch use 'tracker watch {run_id} -r {remote}'".format(
run_id=run_id[:8], remote=args.remote))
def _tail(run):
if os.getenv("NO_WATCHING_MSG") != "1":
cli.out("Watching run %s (pid: %s)" % (run.id, run.pid), err=True)
if run.pid is None:
_print_output(run)
return
proc = psutil.Process(run.pid)
output_path = run.tracker_path("output")
f = None
while proc.is_running():
f = f or _try_open(output_path)
if not f:
time.sleep(1.0)
continue
line = f.readline()
if not line:
time.sleep(0.1)
continue
sys.stdout.write(line)
sys.stdout.flush()
def _handle_remote_process_detached(e, remote):
run_id = e.args[0]
cli.out("\nDetached from remote run {run_id} (still running)\n"
"To re-attach use 'tracker watch {run_id} -r {remote}'".format(
run_id=run_id[:8], remote=remote))
def _handle_remote_run_failed(e, remote):
run_id = os.path.basename(e.remote_run_dir)
cli.out("Try 'tracker runs info %s -O -r %s' to view its output." %
(run_id[:8], remote.name),
err=True)
cli.error()
def _print_run_status(run):
cli.out("Run %s stopped with a status of '%s'" %
(run.short_id, run.status),
err=True)
def _stopped_msg(run):
msg = "\nStopped watching %s" % run.short_id
if run.pid and psutil.Process(run.pid).is_running():
msg += " (%s still running)" % run.pid
cli.out(msg)