def describe_trial(args: Namespace) -> None:
if args.metrics:
r = api.get(args.master, "trials/{}/metrics".format(args.trial_id))
else:
r = api.get(args.master, "trials/{}".format(args.trial_id))
trial = r.json()
if args.json:
print(json.dumps(trial, indent=4))
return
# Print information about the trial itself.
headers = [
"Experiment ID",
"State",
"H-Params",
"Start Time",
"End Time",
]
values = [[
trial["experiment_id"],
trial["state"],
json.dumps(trial["hparams"], indent=4),
render.format_time(trial["start_time"]),
render.format_time(trial["end_time"]),
]]
render.tabulate_or_csv(headers, values, args.csv)
# Print information about individual steps.
headers = [
"# of Batches",
"State",
"Start Time",
"End Time",
"Checkpoint",
"Checkpoint UUID",
"Checkpoint Metadata",
"Validation",
"Validation Metrics",
]
if args.metrics:
headers.append("Workload Metrics")
values = [[
s["prior_batches_processed"] + s["num_batches"],
s["state"],
render.format_time(s["start_time"]),
render.format_time(s["end_time"]),
*format_checkpoint(s["checkpoint"]),
*format_validation(s["validation"]),
*([json.dumps(s["metrics"], indent=4)] if args.metrics else []),
] for s in trial["steps"]]
print()
print("Workloads:")
render.tabulate_or_csv(headers, values, args.csv)
def format_experiment(e: Any) -> List[Any]:
result = [
e["id"],
e["owner"]["username"],
e["config"]["description"],
e["state"],
render.format_percent(e["progress"]),
render.format_time(e["start_time"]),
render.format_time(e["end_time"]),
e["config"]["resources"].get("resource_pool"),
]
if args.all:
result.append(e["archived"])
return result
def format_experiment(e: Any) -> List[Any]:
result = [
e.id,
e.username,
e.name,
e.forkedFrom,
e.state.value.replace("STATE_", ""),
render.format_percent(e.progress),
render.format_time(e.startTime),
render.format_time(e.endTime),
e.resourcePool,
]
if args.all:
result.append(e.archived)
return result
def list_agents(args: argparse.Namespace) -> None:
r = api.get(args.master, "agents")
agents = r.json()
agents = [
OrderedDict([
("id", local_id(agent_id)),
("registered_time", render.format_time(agent["registered_time"])),
("num_slots", len(agent["slots"])),
("num_containers", agent["num_containers"]),
("resource_pool", agent["resource_pool"]),
("label", agent["label"]),
("addresses", ", ".join(agent["addresses"])),
]) for agent_id, agent in sorted(agents.items())
]
if args.json:
print(json.dumps(agents, indent=4))
return
headers = [
"Agent ID",
"Registered Time",
"Slots",
"Containers",
"Resource Pool",
"Label",
"Addresses",
]
values = [a.values() for a in agents]
render.tabulate_or_csv(headers, values, args.csv)
def list_trials(args: Namespace) -> None:
r = api.get(args.master, "experiments/{}/summary".format(args.experiment_id))
experiment = r.json()
headers = ["Trial ID", "State", "H-Params", "Start Time", "End Time", "# of Batches"]
values = [
[
t["id"],
t["state"],
json.dumps(t["hparams"], indent=4),
render.format_time(t["start_time"]),
render.format_time(t["end_time"]),
t["total_batches_processed"],
]
for t in experiment["trials"]
]
render.tabulate_or_csv(headers, values, args.csv)
def render_tasks(args: Namespace, tasks: Dict[str, Dict[str, Any]]) -> None:
def agent_info(t: Dict[str, Any]) -> Union[str, List[str]]:
resources = t.get("resources", [])
if not resources:
return "unassigned"
agents = [a for r in resources for a in r["agent_devices"]]
if len(agents) == 1:
agent = agents[0] # type: str
return agent
return agents
if args.json:
print(json.dumps(tasks, indent=4))
return
headers = [
"Task ID",
"Allocation ID",
"Name",
"Slots Needed",
"Registered Time",
"Agent",
"Priority",
"Resource Pool",
]
values = [
[
task["task_id"],
task["allocation_id"],
task["name"],
task["slots_needed"],
render.format_time(task["registered_time"]),
agent_info(task),
task["priority"] if task["scheduler_type"] == "priority" else "N/A",
task["resource_pool"],
]
for task_id, task in sorted(
tasks.items(),
key=lambda tup: (render.format_time(tup[1]["registered_time"]),),
)
]
render.tabulate_or_csv(headers, values, args.csv)
def list_tasks(args: Namespace) -> None:
r = api.get(args.master, "tasks")
def agent_info(t: Dict[str, Any]) -> Union[str, List[str]]:
containers = t.get("containers", [])
if not containers:
return "unassigned"
if len(containers) == 1:
agent = containers[0]["agent"] # type: str
return agent
return [c["agent"] for c in containers]
tasks = r.json()
headers = [
"ID",
"Name",
"Slots Needed",
"Registered Time",
"Agent",
"Priority",
"Resource Pool",
]
values = [[
task["id"],
task["name"],
task["slots_needed"],
render.format_time(task["registered_time"]),
agent_info(task),
task["priority"] if task["scheduler_type"] == "priority" else "N/A",
task["resource_pool"],
] for task_id, task in sorted(
tasks.items(),
key=lambda tup: (render.format_time(tup[1]["registered_time"]), ),
)]
render.tabulate_or_csv(headers, values, args.csv)
def list_trials(args: Namespace) -> None:
all_trials: List[bindings.trialv1Trial] = limit_offset_paginator(
bindings.get_GetExperimentTrials,
"trials",
setup_session(args),
experimentId=args.experiment_id,
limit=args.limit,
offset=args.offset,
)
headers = [
"Trial ID", "State", "H-Params", "Start Time", "End Time",
"# of Batches"
]
values = [[
t.id,
t.state.value.replace("STATE_", ""),
json.dumps(t.hparams, indent=4),
render.format_time(t.startTime),
render.format_time(t.endTime),
t.totalBatchesProcessed,
] for t in all_trials]
render.tabulate_or_csv(headers, values, args.csv)
def list_agents(args: argparse.Namespace) -> None:
resp = bindings.get_GetAgents(setup_session(args))
agents = [
OrderedDict([
("id", local_id(a.id)),
("version", a.version),
("registered_time", render.format_time(a.registeredTime)),
("num_slots", len(a.slots) if a.slots is not None else ""),
("num_containers",
len(a.containers) if a.containers is not None else ""),
(
"resource_pools",
", ".join(a.resourcePools)
if a.resourcePools is not None else "",
),
("enabled", a.enabled),
("draining", a.draining),
("label", a.label),
("addresses",
", ".join(a.addresses) if a.addresses is not None else ""),
]) for a in sorted(resp.agents or [], key=attrgetter("id"))
]
if args.json:
print(json.dumps(agents, indent=4))
return
headers = [
"Agent ID",
"Version",
"Registered Time",
"Slots",
"Containers",
"Resource Pool",
"Enabled",
"Draining",
"Label",
"Addresses",
]
values = [a.values() for a in agents]
render.tabulate_or_csv(headers, values, args.csv)
def describe(args: Namespace) -> None:
docs = []
for experiment_id in args.experiment_ids.split(","):
if args.metrics:
r = api.get(args.master, "experiments/{}/metrics/summary".format(experiment_id))
else:
r = api.get(args.master, "experiments/{}".format(experiment_id))
docs.append(r.json())
if args.json:
print(json.dumps(docs, indent=4))
return
# Display overall experiment information.
headers = [
"Experiment ID",
"State",
"Progress",
"Start Time",
"End Time",
"Description",
"Archived",
"Resource Pool",
"Labels",
]
values = [
[
doc["id"],
doc["state"],
render.format_percent(doc["progress"]),
render.format_time(doc.get("start_time")),
render.format_time(doc.get("end_time")),
doc["config"].get("description"),
doc["archived"],
doc["config"]["resources"].get("resource_pool"),
", ".join(sorted(doc["config"].get("labels") or [])),
]
for doc in docs
]
if not args.outdir:
outfile = None
print("Experiment:")
else:
outfile = args.outdir.joinpath("experiments.csv")
render.tabulate_or_csv(headers, values, args.csv, outfile)
# Display trial-related information.
headers = ["Trial ID", "Experiment ID", "State", "Start Time", "End Time", "H-Params"]
values = [
[
trial["id"],
doc["id"],
trial["state"],
render.format_time(trial.get("start_time")),
render.format_time(trial.get("end_time")),
json.dumps(trial["hparams"], indent=4),
]
for doc in docs
for trial in doc["trials"]
]
if not args.outdir:
outfile = None
print("\nTrials:")
else:
outfile = args.outdir.joinpath("trials.csv")
render.tabulate_or_csv(headers, values, args.csv, outfile)
# Display step-related information.
if args.metrics:
# Accumulate the scalar training and validation metric names from all provided experiments.
t_metrics_names = sorted({n for doc in docs for n in scalar_training_metrics_names(doc)})
t_metrics_headers = ["Training Metric: {}".format(name) for name in t_metrics_names]
v_metrics_names = sorted({n for doc in docs for n in scalar_validation_metrics_names(doc)})
v_metrics_headers = ["Validation Metric: {}".format(name) for name in v_metrics_names]
else:
t_metrics_headers = []
v_metrics_headers = []
headers = (
["Trial ID", "# of Batches", "State", "Start Time", "End Time"]
+ t_metrics_headers
+ [
"Checkpoint State",
"Checkpoint Start Time",
"Checkpoint End Time",
"Validation State",
"Validation Start Time",
"Validation End Time",
]
+ v_metrics_headers
)
values = []
for doc in docs:
for trial in doc["trials"]:
for step in trial["steps"]:
t_metrics_fields = []
if step.get("metrics"):
avg_metrics = step["metrics"]["avg_metrics"]
for name in t_metrics_names:
if name in avg_metrics:
t_metrics_fields.append(avg_metrics[name])
else:
t_metrics_fields.append(None)
checkpoint = step.get("checkpoint")
if checkpoint:
checkpoint_state = checkpoint["state"]
checkpoint_start_time = checkpoint.get("start_time")
checkpoint_end_time = checkpoint.get("end_time")
else:
checkpoint_state = None
checkpoint_start_time = None
checkpoint_end_time = None
validation = step.get("validation")
if validation:
validation_state = validation["state"]
validation_start_time = validation.get("start_time")
validation_end_time = validation.get("end_time")
else:
validation_state = None
validation_start_time = None
validation_end_time = None
if args.metrics:
v_metrics_fields = [
api.metric.get_validation_metric(name, validation)
for name in v_metrics_names
]
else:
v_metrics_fields = []
row = (
[
step["trial_id"],
step["total_batches"],
step["state"],
render.format_time(step.get("start_time")),
render.format_time(step.get("end_time")),
]
+ t_metrics_fields
+ [
checkpoint_state,
render.format_time(checkpoint_start_time),
render.format_time(checkpoint_end_time),
validation_state,
render.format_time(validation_start_time),
render.format_time(validation_end_time),
]
+ v_metrics_fields
)
values.append(row)
if not args.outdir:
outfile = None
print("\nWorkloads:")
else:
outfile = args.outdir.joinpath("workloads.csv")
render.tabulate_or_csv(headers, values, args.csv, outfile)
def describe(args: Namespace) -> None:
session = setup_session(args)
exps = []
for experiment_id in args.experiment_ids.split(","):
r = bindings.get_GetExperiment(session, experimentId=experiment_id)
if args.json:
exps.append(r.to_json())
else:
exps.append(r.experiment)
if args.json:
print(json.dumps(exps, indent=4))
return
# Display overall experiment information.
headers = [
"Experiment ID",
"State",
"Progress",
"Start Time",
"End Time",
"Name",
"Description",
"Archived",
"Resource Pool",
"Labels",
]
values = [[
exp.id,
exp.state.value.replace("STATE_", ""),
render.format_percent(exp.progress),
render.format_time(exp.startTime),
render.format_time(exp.endTime),
exp.name,
exp.description,
exp.archived,
exp.resourcePool,
", ".join(sorted(exp.labels or [])),
] for exp in exps]
if not args.outdir:
outfile = None
print("Experiment:")
else:
outfile = args.outdir.joinpath("experiments.csv")
render.tabulate_or_csv(headers, values, args.csv, outfile)
# Display trial-related information.
trials_for_experiment: Dict[str, Sequence[bindings.trialv1Trial]] = {}
for exp in exps:
trials_for_experiment[exp.id] = bindings.get_GetExperimentTrials(
session, experimentId=exp.id).trials
headers = [
"Trial ID", "Experiment ID", "State", "Start Time", "End Time",
"H-Params"
]
values = [[
trial.id,
exp.id,
trial.state.value.replace("STATE_", ""),
render.format_time(trial.startTime),
render.format_time(trial.endTime),
json.dumps(trial.hparams, indent=4),
] for exp in exps for trial in trials_for_experiment[exp.id]]
if not args.outdir:
outfile = None
print("\nTrials:")
else:
outfile = args.outdir.joinpath("trials.csv")
render.tabulate_or_csv(headers, values, args.csv, outfile)
# Display step-related information.
t_metrics_headers: List[str] = []
t_metrics_names: List[str] = []
v_metrics_headers: List[str] = []
v_metrics_names: List[str] = []
if args.metrics:
# Accumulate the scalar training and validation metric names from all provided experiments.
for exp in exps:
sample_trial = trials_for_experiment[exp.id][0]
sample_workloads = bindings.get_GetTrial(
session, trialId=sample_trial.id).workloads
t_metrics_names += scalar_training_metrics_names(sample_workloads)
v_metrics_names += scalar_validation_metrics_names(
sample_workloads)
t_metrics_names = sorted(set(t_metrics_names))
t_metrics_headers = [
"Training Metric: {}".format(name) for name in t_metrics_names
]
v_metrics_names = sorted(set(v_metrics_names))
v_metrics_headers = [
"Validation Metric: {}".format(name) for name in v_metrics_names
]
headers = (["Trial ID", "# of Batches", "State", "Report Time"] +
t_metrics_headers + [
"Checkpoint State",
"Checkpoint Report Time",
"Validation State",
"Validation Report Time",
] + v_metrics_headers)
wl_output: Dict[int, List[Any]] = {}
for exp in exps:
for trial in trials_for_experiment[exp.id]:
workloads = bindings.get_GetTrial(session,
trialId=trial.id).workloads
for workload in workloads:
t_metrics_fields = []
wl_detail: Optional[
Union[bindings.v1MetricsWorkload,
bindings.v1CheckpointWorkload]] = None
if workload.training:
wl_detail = workload.training
for name in t_metrics_names:
if wl_detail.metrics and (name in wl_detail.metrics):
t_metrics_fields.append(wl_detail.metrics[name])
else:
t_metrics_fields.append(None)
else:
t_metrics_fields = [None for name in t_metrics_names]
if workload.checkpoint:
wl_detail = workload.checkpoint
if workload.checkpoint and wl_detail:
checkpoint_state = wl_detail.state.value
checkpoint_end_time = wl_detail.endTime
else:
checkpoint_state = ""
checkpoint_end_time = None
v_metrics_fields = []
if workload.validation:
wl_detail = workload.validation
validation_state = wl_detail.state.value
validation_end_time = wl_detail.endTime
for name in v_metrics_names:
if wl_detail.metrics and (name in wl_detail.metrics):
v_metrics_fields.append(wl_detail.metrics[name])
else:
v_metrics_fields.append(None)
else:
validation_state = ""
validation_end_time = None
v_metrics_fields = [None for name in v_metrics_names]
if wl_detail:
if wl_detail.totalBatches in wl_output:
# condense training, checkpoints, validation workloads into one step-like
# row for compatibility with previous versions of describe
merge_row = wl_output[wl_detail.totalBatches]
merge_row[3] = max(
merge_row[3],
render.format_time(wl_detail.endTime))
for idx, tfield in enumerate(t_metrics_fields):
if tfield and merge_row[4 + idx] is None:
merge_row[4 + idx] = tfield
start_checkpoint = 4 + len(t_metrics_fields)
if checkpoint_state:
merge_row[
start_checkpoint] = checkpoint_state.replace(
"STATE_", "")
merge_row[start_checkpoint +
1] = render.format_time(
checkpoint_end_time)
if validation_end_time:
merge_row[start_checkpoint +
3] = render.format_time(
validation_end_time)
if validation_state:
merge_row[start_checkpoint +
2] = validation_state.replace(
"STATE_", "")
for idx, vfield in enumerate(v_metrics_fields):
if vfield and merge_row[start_checkpoint + idx +
4] is None:
merge_row[start_checkpoint + idx + 4] = vfield
else:
row = ([
trial.id,
wl_detail.totalBatches,
wl_detail.state.value.replace("STATE_", ""),
render.format_time(wl_detail.endTime),
] + t_metrics_fields + [
checkpoint_state.replace("STATE_", ""),
render.format_time(checkpoint_end_time),
validation_state.replace("STATE_", ""),
render.format_time(validation_end_time),
] + v_metrics_fields)
wl_output[wl_detail.totalBatches] = row
if not args.outdir:
outfile = None
print("\nWorkloads:")
else:
outfile = args.outdir.joinpath("workloads.csv")
values = sorted(wl_output.values(), key=lambda a: int(a[1]))
render.tabulate_or_csv(headers, values, args.csv, outfile)
