ClearML/clearml-server
1
0
Fork 0
mirror of https://github.com/clearml/clearml-server synced 2025-05-08 05:54:58 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix task scalars comparison aggregation

Browse Source
This commit is contained in:
allegroai 2020-02-04 18:16:27 +02:00
parent a392a0e6ff
commit 0722b20c1c
3 changed files with 140 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

68
server/bll/event/event_metrics.py
View File

@ -4,8 +4,9 @@ from concurrent.futures import ThreadPoolExecutor
from functools import partial
from operator import itemgetter
from boltons.iterutils import bucketize
from elasticsearch import Elasticsearch
from typing import Sequence, Tuple, Callable
from typing import Sequence, Tuple, Callable, Iterable
from mongoengine import Q
@ -21,9 +22,10 @@ log = config.logger(__file__)
class EventMetrics:
MAX_TASKS_COUNT = 100
MAX_TASKS_COUNT = 50
MAX_METRICS_COUNT = 200
MAX_VARIANTS_COUNT = 500
MAX_AGGS_ELEMENTS_COUNT = 50
def __init__(self, es: Elasticsearch):
self.es = es
@ -62,6 +64,11 @@ class EventMetrics:
Compare scalar metrics for different tasks per metric and variant
The amount of points in each histogram should not exceed the requested samples
"""
if len(task_ids) > self.MAX_TASKS_COUNT:
raise errors.BadRequest(
f"Up to {self.MAX_TASKS_COUNT} tasks supported for comparison", len(task_ids)
)
task_name_by_id = {}
with translate_errors_context():
task_objs = Task.get_many(
@ -97,6 +104,31 @@ class EventMetrics:
MetricInterval = Tuple[int, Sequence[TaskMetric]]
MetricData = Tuple[str, dict]
def _split_metrics_by_max_aggs_count(
self, task_metrics: Sequence[TaskMetric]
) -> Iterable[Sequence[TaskMetric]]:
"""
Return task metrics in groups where amount of task metrics in each group
is roughly limited by MAX_AGGS_ELEMENTS_COUNT. The split is done on metrics and
variants while always preserving all their tasks in the same group
"""
if len(task_metrics) < self.MAX_AGGS_ELEMENTS_COUNT:
yield task_metrics
return
tm_grouped = bucketize(task_metrics, key=itemgetter(1, 2))
groups = []
for group in tm_grouped.values():
groups.append(group)
if sum(map(len, groups)) >= self.MAX_AGGS_ELEMENTS_COUNT:
yield list(itertools.chain(*groups))
groups = []
if groups:
yield list(itertools.chain(*groups))
return
def _run_get_scalar_metrics_as_parallel(
self,
company_id: str,
@ -126,21 +158,27 @@ class EventMetrics:
if not intervals:
return {}
with ThreadPoolExecutor(len(intervals)) as pool:
metrics = list(
itertools.chain.from_iterable(
pool.map(
partial(
get_func, task_ids=task_ids, es_index=es_index, key=key
),
intervals,
)
intervals = list(
itertools.chain.from_iterable(
zip(itertools.repeat(i), self._split_metrics_by_max_aggs_count(tms))
for i, tms in intervals
)
)
max_concurrency = config.get("services.events.max_metrics_concurrency", 4)
with ThreadPoolExecutor(max_workers=max_concurrency) as pool:
metrics = itertools.chain.from_iterable(
pool.map(
partial(
get_func, task_ids=task_ids, es_index=es_index, key=key
),
intervals,
)
)
ret = defaultdict(dict)
for metric_key, metric_values in metrics:
ret[metric_key].update(metric_values)
return ret
def _get_metric_intervals(
@ -310,7 +348,13 @@ class EventMetrics:
"variants": {
"terms": {"field": "variant", "size": self.MAX_VARIANTS_COUNT},
"aggs": {
"tasks": {"terms": {"field": "task"}, "aggs": aggregation}
"tasks": {
"terms": {
"field": "task",
"size": self.MAX_TASKS_COUNT,
},
"aggs": aggregation,
}
},
}
},

6
server/config/default/services/events.conf
View File

@ -2,4 +2,8 @@ es_index_prefix: "events"
ignore_iteration {
metrics: [":monitor:machine", ":monitor:gpu"]
}
}
# max number of concurrent queries to ES when calculating events metrics
# should not exceed the amount of concurrent connections set in the ES driver
max_metrics_concurrency: 4

121
server/tests/automated/test_task_events.py
View File

@ -2,9 +2,12 @@
Comprehensive test of all(?) use cases of datasets and frames
"""
import json
import time
import unittest
from statistics import mean
from typing import Sequence
import es_factory
from config import config
from tests.automated import TestService
@ -16,69 +19,56 @@ class TestTaskEvents(TestService):
def setUp(self, version="1.7"):
super().setUp(version=version)
self.created_tasks = []
self.task = dict(
name="test task events",
type="training",
input=dict(mapping={}, view=dict(entries=[])),
def _temp_task(self, name="test task events"):
task_input = dict(
name=name, type="training", input=dict(mapping={}, view=dict(entries=[])),
)
res, self.task_id = self.api.send("tasks.create", self.task, extract="id")
assert res.meta.result_code == 200
self.created_tasks.append(self.task_id)
return self.create_temp("tasks", **task_input)
def tearDown(self):
log.info("Cleanup...")
for task_id in self.created_tasks:
try:
self.api.send("tasks.delete", dict(task=task_id, force=True))
except Exception as ex:
log.exception(ex)
def create_task_event(self, type, iteration):
def _create_task_event(self, type_, task, iteration):
return {
"worker": "test",
"type": type,
"task": self.task_id,
"type": type_,
"task": task,
"iter": iteration,
"timestamp": es_factory.get_timestamp_millis()
"timestamp": es_factory.get_timestamp_millis(),
}
def copy_and_update(self, src_obj, new_data):
def _copy_and_update(self, src_obj, new_data):
obj = src_obj.copy()
obj.update(new_data)
return obj
def test_task_logs(self):
events = []
for iter in range(10):
log_event = self.create_task_event("log", iteration=iter)
task = self._temp_task()
for iter_ in range(10):
log_event = self._create_task_event("log", task, iteration=iter_)
events.append(
self.copy_and_update(
self._copy_and_update(
log_event,
{"msg": "This is a log message from test task iter " + str(iter)},
{"msg": "This is a log message from test task iter " + str(iter_)},
)
)
# sleep so timestamp is not the same
import time
time.sleep(0.01)
self.send_batch(events)
data = self.api.events.get_task_log(task=self.task_id)
data = self.api.events.get_task_log(task=task)
assert len(data["events"]) == 10
self.api.tasks.reset(task=self.task_id)
data = self.api.events.get_task_log(task=self.task_id)
self.api.tasks.reset(task=task)
data = self.api.events.get_task_log(task=task)
assert len(data["events"]) == 0
def test_task_metric_value_intervals_keys(self):
metric = "Metric1"
variant = "Variant1"
iter_count = 100
task = self._temp_task()
events = [
{
**self.create_task_event("training_stats_scalar", iteration),
**self._create_task_event("training_stats_scalar", task, iteration),
"metric": metric,
"variant": variant,
"value": iteration,
@ -88,19 +78,65 @@ class TestTaskEvents(TestService):
self.send_batch(events)
for key in None, "iter", "timestamp", "iso_time":
with self.subTest(key=key):
data = self.api.events.scalar_metrics_iter_histogram(task=self.task_id, key=key)
data = self.api.events.scalar_metrics_iter_histogram(task=task, key=key)
self.assertIn(metric, data)
self.assertIn(variant, data[metric])
self.assertIn("x", data[metric][variant])
self.assertIn("y", data[metric][variant])
def test_multitask_events_many_metrics(self):
tasks = [
self._temp_task(name="test events1"),
self._temp_task(name="test events2"),
]
iter_count = 10
metrics_count = 10
variants_count = 10
events = [
{
**self._create_task_event("training_stats_scalar", task, iteration),
"metric": f"Metric{metric_idx}",
"variant": f"Variant{variant_idx}",
"value": iteration,
}
for iteration in range(iter_count)
for task in tasks
for metric_idx in range(metrics_count)
for variant_idx in range(variants_count)
]
self.send_batch(events)
data = self.api.events.multi_task_scalar_metrics_iter_histogram(tasks=tasks)
self._assert_metrics_and_variants(
data.metrics,
metrics=metrics_count,
variants=variants_count,
tasks=tasks,
iterations=iter_count,
)
def _assert_metrics_and_variants(
self, data: dict, metrics: int, variants: int, tasks: Sequence, iterations: int
):
self.assertEqual(len(data), metrics)
for m in range(metrics):
metric_data = data[f"Metric{m}"]
self.assertEqual(len(metric_data), variants)
for v in range(variants):
variant_data = metric_data[f"Variant{v}"]
self.assertEqual(len(variant_data), len(tasks))
for t in tasks:
task_data = variant_data[t]
self.assertEqual(len(task_data["x"]), iterations)
self.assertEqual(len(task_data["y"]), iterations)
def test_task_metric_value_intervals(self):
metric = "Metric1"
variant = "Variant1"
iter_count = 100
task = self._temp_task()
events = [
{
**self.create_task_event("training_stats_scalar", iteration),
**self._create_task_event("training_stats_scalar", task, iteration),
"metric": metric,
"variant": variant,
"value": iteration,
@ -109,13 +145,13 @@ class TestTaskEvents(TestService):
]
self.send_batch(events)
data = self.api.events.scalar_metrics_iter_histogram(task=self.task_id)
data = self.api.events.scalar_metrics_iter_histogram(task=task)
self._assert_metrics_histogram(data[metric][variant], iter_count, 100)
data = self.api.events.scalar_metrics_iter_histogram(task=self.task_id, samples=100)
data = self.api.events.scalar_metrics_iter_histogram(task=task, samples=100)
self._assert_metrics_histogram(data[metric][variant], iter_count, 100)
data = self.api.events.scalar_metrics_iter_histogram(task=self.task_id, samples=10)
data = self.api.events.scalar_metrics_iter_histogram(task=task, samples=10)
self._assert_metrics_histogram(data[metric][variant], iter_count, 10)
def _assert_metrics_histogram(self, data, iters, samples):
@ -130,7 +166,8 @@ class TestTaskEvents(TestService):
)
def test_task_plots(self):
event = self.create_task_event("plot", 0)
task = self._temp_task()
event = self._create_task_event("plot", task, 0)
event["metric"] = "roc"
event.update(
{
@ -179,7 +216,7 @@ class TestTaskEvents(TestService):
)
self.send(event)
event = self.create_task_event("plot", 100)
event = self._create_task_event("plot", task, 100)
event["metric"] = "confusion"
event.update(
{
@ -222,11 +259,11 @@ class TestTaskEvents(TestService):
)
self.send(event)
data = self.api.events.get_task_plots(task=self.task_id)
data = self.api.events.get_task_plots(task=task)
assert len(data["plots"]) == 2
self.api.tasks.reset(task=self.task_id)
data = self.api.events.get_task_plots(task=self.task_id)
self.api.tasks.reset(task=task)
data = self.api.events.get_task_plots(task=task)
assert len(data["plots"]) == 0
def send_batch(self, events):