tahamajs's picture
download
raw
4.63 kB
import uuid
from django.db import models
from experiment.models import Experiment
class Sweep(models.Model):
class Status(models.TextChoices):
ACTIVE = "active", "Active"
PAUSED = "paused", "Paused"
COMPLETED = "completed", "Completed"
FAILED = "failed", "Failed"
class Direction(models.TextChoices):
MAXIMIZE = "maximize", "Maximize"
MINIMIZE = "minimize", "Minimize"
id = models.UUIDField(primary_key=True, default=uuid.uuid4, editable=False)
project = models.ForeignKey(
Experiment, on_delete=models.CASCADE, related_name="sweep_set"
)
name = models.CharField(max_length=255)
description = models.TextField(blank=True)
config = models.JSONField(
default=dict,
blank=True,
help_text="Full sweep configuration: search space, base_config, etc.",
)
status = models.CharField(
max_length=20, choices=Status.choices, default=Status.ACTIVE
)
direction = models.CharField(
max_length=20, choices=Direction.choices, default=Direction.MAXIMIZE
)
directions = models.JSONField(default=list, blank=True)
metric_name = models.CharField(max_length=100, default="eval_accuracy")
metric_names = models.JSONField(default=list, blank=True)
algorithm = models.CharField(max_length=50, default="tpe")
pruner = models.CharField(max_length=50, default="hyperband")
scheduler_type = models.CharField(max_length=50, default="fifo")
max_trials = models.IntegerField(default=20)
max_failures = models.IntegerField(default=3)
parallel_workers = models.IntegerField(default=1)
warm_start_from = models.ForeignKey(
"self",
on_delete=models.SET_NULL,
null=True,
blank=True,
related_name="warm_started_sweeps",
)
# Preferential optimization
preference_enabled = models.BooleanField(default=False)
preference_queries = models.JSONField(default=list, blank=True)
# Artifact store
artifact_dir = models.CharField(max_length=500, blank=True)
# Tags
tags = models.JSONField(default=list, blank=True)
created_at = models.DateTimeField(auto_now_add=True)
updated_at = models.DateTimeField(auto_now=True)
class Meta:
ordering = ["-created_at"]
indexes = [
models.Index(fields=["project", "status"]),
models.Index(fields=["status"]),
]
def __str__(self):
return self.name
@property
def best_trial(self):
if self.metric_names:
return None
ordering = (
"-objective_value"
if self.direction == self.Direction.MAXIMIZE
else "objective_value"
)
return (
self.trials.exclude(objective_value__isnull=True).order_by(ordering).first()
)
@property
def pareto_trials(self):
if not self.metric_names:
return []
return self.trials.filter(is_pareto=True)
class SweepTrial(models.Model):
class Status(models.TextChoices):
QUEUED = "queued", "Queued"
RUNNING = "running", "Running"
COMPLETED = "completed", "Completed"
FAILED = "failed", "Failed"
PRUNED = "pruned", "Pruned"
id = models.UUIDField(primary_key=True, default=uuid.uuid4, editable=False)
sweep = models.ForeignKey(Sweep, on_delete=models.CASCADE, related_name="trials")
trial_number = models.IntegerField()
status = models.CharField(
max_length=20, choices=Status.choices, default=Status.QUEUED
)
objective_value = models.FloatField(null=True, blank=True)
objectives = models.JSONField(null=True, blank=True)
params = models.JSONField(default=dict, blank=True)
user_attrs = models.JSONField(default=dict, blank=True)
notes = models.TextField(blank=True)
is_pareto = models.BooleanField(default=False)
run = models.OneToOneField(
"experiment.TrainingRun",
on_delete=models.SET_NULL,
null=True,
blank=True,
related_name="sweep_trial",
)
# Artifact store
artifacts = models.JSONField(default=list, blank=True)
# Timing
duration_seconds = models.FloatField(null=True, blank=True)
created_at = models.DateTimeField(auto_now_add=True)
class Meta:
ordering = ["sweep", "trial_number"]
unique_together = [["sweep", "trial_number"]]
indexes = [
models.Index(fields=["sweep", "status"]),
models.Index(fields=["objective_value"]),
]
def __str__(self):
return f"{self.sweep.name} trial {self.trial_number}"

Xet Storage Details

Size:
4.63 kB
·
Xet hash:
7e9aca6be6394b0dfda3d2760ee6962391d6db48138653a3dae270f0d066cc26

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.