Integración de MLflow para Ultralytics YOLO
Introducción#
El seguimiento de experimentos es una parte esencial de los flujos de trabajo de aprendizaje automático: registra métricas, parámetros y artefactos para que puedas reproducir resultados, depurar problemas y mejorar el rendimiento del modelo. Ultralytics YOLO, conocido por sus capacidades de detección de objetos en tiempo real, se integra con MLflow, una plataforma de código abierto para gestionar todo el ciclo de vida del aprendizaje automático.
En esta guía se explica cómo configurar y usar el registro de MLflow en tu proyecto Ultralytics YOLO.
¿Qué es MLflow?#
MLflow es una plataforma de código abierto desarrollada por Databricks para gestionar el ciclo de vida completo del aprendizaje automático. Incluye herramientas para hacer un seguimiento de experimentos, empaquetar código en ejecuciones reproducibles y compartir y desplegar modelos. MLflow está diseñado para funcionar con cualquier biblioteca de aprendizaje automático y lenguaje de programación.
Funciones#
- Registro de métricas: Registra las tasas de aprendizaje, las pérdidas de entrenamiento y las métricas de validación al final de cada época.
- Registro de parámetros: Registra todos los parámetros utilizados durante el entrenamiento.
- Registro de artefactos: Registra los artefactos del modelo, incluidos los pesos y los archivos de configuración, al final del entrenamiento.
Configuración y requisitos previos#
Asegúrate de que MLflow esté instalado. Si no lo está, instálalo con pip:
pip install mlflowEl registro de MLflow está activado de forma predeterminada mediante la clave mlflow en los ajustes de Ultralytics. Si lo has desactivado, vuelve a activarlo:
En el entorno Python, llama al método update del objeto settings para cambiar la configuración:
from ultralytics import settings
# Actualizar un ajuste
settings.update({"mlflow": True})
# Restablecer los ajustes a sus valores predeterminados
settings.reset()Cómo usarlo#
Comandos#
-
Establecer el nombre de un proyecto: Puedes establecer el nombre del proyecto mediante una variable de entorno:
export MLFLOW_EXPERIMENT_NAME=YOUR_EXPERIMENT_NAMETambién puedes usar el argumento
project=<project>al entrenar un modelo YOLO, por ejemplo,yolo train project=my_project. -
Establecer el nombre de una ejecución: Al igual que con el nombre del proyecto, puedes establecer el nombre de la ejecución mediante una variable de entorno:
export MLFLOW_RUN=YOUR_RUN_NAMETambién puedes usar el argumento
name=<name>al entrenar un modelo YOLO, por ejemplo,yolo train project=my_project name=my_name. -
Iniciar un servidor local de MLflow: Sin
MLFLOW_TRACKING_URI, las ejecuciones se registran en el almacén local predeterminado de MLflow, en el directorio en el que entrenes (mlflow.dben MLflow 3.7 y versiones posteriores, salvo que el directorio ya contenga un historial demlruns/;mlruns/en versiones anteriores). Para verlas, ejecuta esto desde el mismo directorio:mlflow uiEsto iniciará de forma predeterminada un servidor local en
http://127.0.0.1:5000. Para dirigir tus ejecuciones de entrenamiento a otro servidor de seguimiento, exportaMLFLOW_TRACKING_URIantes de entrenar:export MLFLOW_TRACKING_URI=http://127.0.0.1:5000 -
Mantener activa la ejecución de MLflow después del entrenamiento: de forma predeterminada, Ultralytics finaliza la ejecución de MLflow que inició al terminar el entrenamiento, con el estado
FAILEDsi se produjo un error durante el entrenamiento antes de cerrar la ejecución. Se reutiliza y se deja abierta una ejecución que hayas iniciado tú antes del entrenamiento. Para mantener abierta la ejecución (por ejemplo, para registrar métricas o artefactos adicionales desde el mismo cuaderno), estableceMLFLOW_KEEP_RUN_ACTIVEenTrueantes del entrenamiento:export MLFLOW_KEEP_RUN_ACTIVE=TrueLos valores
1,true,yes,on,yytactivan este comportamiento (sin distinguir entre mayúsculas y minúsculas). Cualquier otro valor, incluida una variable sin definir, mantiene el comportamiento predeterminado de cerrar la ejecución. Recuerda cerrarla manualmente después conmlflow.end_run().
Registro#
El registro se gestiona mediante on_pretrain_routine_end, on_train_epoch_end, on_fit_epoch_end y on_train_end funciones de devolución de llamada. Estas funciones se llaman automáticamente durante las etapas correspondientes del proceso de entrenamiento y se encargan de registrar parámetros, métricas y artefactos. La función de devolución de llamada teardown finaliza una ejecución que un entrenamiento fallido dejó abierta.
Ejemplos#
-
Registro de métricas personalizadas: Puedes añadir métricas personalizadas para que se registren modificando el diccionario
trainer.metricsantes de que se llame aon_fit_epoch_end. -
Ver experimento: Para consultar tus registros, ve a tu servidor MLflow (normalmente
http://127.0.0.1:5000) y selecciona tu experimento y la ejecución.
-
Ver ejecución: Las ejecuciones son modelos individuales dentro de un experimento. Haz clic en una ejecución para ver sus detalles, incluidos los artefactos subidos y los pesos del modelo.

Desactivar MLflow#
Para desactivar el registro de MLflow:
yolo settings mlflow=FalseConclusión#
La integración del registro de MLflow con Ultralytics YOLO ofrece una forma sencilla de hacer un seguimiento de tus experimentos de aprendizaje automático. Te permite supervisar las métricas de rendimiento y gestionar los artefactos durante el desarrollo y la implementación del modelo.
Preguntas frecuentes#
Para configurar el registro de MLflow con Ultralytics YOLO, primero debes asegurarte de que MLflow esté instalado. Puedes instalarlo con pip:
pip install mlflowEl registro de MLflow está activado de forma predeterminada mediante la clave
mlflowen los ajustes de Ultralytics. Si lo has desactivado, vuelve a activarlo:Actualizar la configuración de MLflow de Ultralyticsfrom ultralytics import settings # Actualizar un ajuste settings.update({"mlflow": True}) # Restablecer los ajustes a sus valores predeterminados settings.reset()Por último, después de entrenar, puedes ver tus ejecuciones iniciando un servidor local de MLflow desde el directorio en el que hayas entrenado:
mlflow uiUltralytics YOLO con MLflow admite el registro de diversas métricas, parámetros y artefactos durante todo el proceso de entrenamiento:
- Registro de métricas: Registra las métricas al final de cada época y al finalizar el entrenamiento.
- Registro de parámetros: Registra todos los parámetros utilizados en el proceso de entrenamiento.
- Registro de artefactos: Guarda artefactos del modelo, como los pesos y los archivos de configuración, después del entrenamiento.
Sí, puedes desactivar el registro de MLflow para Ultralytics YOLO actualizando la configuración. Así puedes hacerlo con la CLI:
yolo settings mlflow=FalsePara personalizar más la configuración y restablecerla, consulta la guía de configuración.
Para ver tus experimentos de Ultralytics YOLO en un servidor local de MLflow, ejecuta el siguiente comando desde el directorio en el que hayas entrenado:
mlflow uiEste comando inicia un servidor local en
http://127.0.0.1:5000de forma predeterminada. PulsaCtrl+Cen ese terminal para detenerlo.Consulta la sección de comandos para ver más opciones.
Antes del entrenamiento, establece la variable de entorno
MLFLOW_KEEP_RUN_ACTIVEenTrue:export MLFLOW_KEEP_RUN_ACTIVE=TrueDe forma predeterminada, este valor es
False, por lo que Ultralytics llama amlflow.end_run()en la ejecución que inició cuando termina el entrenamiento. ConMLFLOW_KEEP_RUN_ACTIVE=True, la ejecución permanece abierta para que puedas registrar métricas, parámetros o artefactos adicionales desde la misma sesión de Python; ciérrala tú conmlflow.end_run()cuando hayas terminado. Los valores1,true,yes,on,yytactivan este comportamiento (sin distinguir entre mayúsculas y minúsculas).Integrar MLflow con Ultralytics YOLO ofrece varias ventajas para gestionar tus experimentos de aprendizaje automático:
- Seguimiento mejorado de experimentos: Haz un seguimiento de distintas ejecuciones y sus resultados, y compáralos.
- Mayor reproducibilidad de los modelos: Asegúrate de que tus experimentos sean reproducibles registrando todos los parámetros y artefactos.
- Supervisión del rendimiento: Visualiza las métricas de rendimiento a lo largo del tiempo para tomar decisiones sobre cómo mejorar los modelos basadas en datos.
- Flujo de trabajo optimizado: Automatiza el proceso de registro para dedicar más tiempo al desarrollo de modelos y menos al seguimiento manual.
- Desarrollo colaborativo: Comparte los resultados de los experimentos con tu equipo para colaborar mejor e intercambiar conocimientos.
Comentarios30
Haz una pregunta, comparte tus resultados o ayuda a otra persona.
UltralyticsAssistantUltralytics
👋 Bonjour, merci pour votre intérêt pour Ultralytics 🚀! Nous vous recommandons de consulter notre documentation complète sur les Intégrations MLflow pour découvrir comment configurer et utiliser MLflow avec Ultralytics YOLO afin d'améliorer le suivi des expériences et la reproductibilité des modèles.
Si ceci est un 🐛 Rapport de bug, veuillez fournir un exemple minimal reproductible pour nous aider à résoudre le problème.
Si vous avez une ❓ Question ou besoin d’aide avec la configuration MLflow, n'hésitez pas à partager plus de détails, tels que votre environnement d'exécution, les étapes effectuées et les éventuels messages d'erreur rencontrés.
Mise à jour
Assurez-vous que vous utilisez la version la plus récente d’
ultralyticsavec toutes les dépendances dans un environnement Python>=3.8 et PyTorch>=1.8. Vous pouvez mettre à jour avec la commande suivante :pip install -U ultralyticsEnvironnements
Vous pouvez exécuter YOLO dans les environnements suivants avec MLflow :
- Notebooks avec GPU gratuit :
- Cloud : Consultez nos guides rapides pour GCP et AWS.
- Docker Image : Consultez notre guide Docker
.
Communauté
Rejoignez la communauté Ultralytics où cela vous convient le mieux ! Pour des discussions en temps réel, rendez-vous sur notre Discord 🎧. Pour des discussions approfondies, visitez notre Discourse, ou échangez avec la communauté sur notre Subreddit.
Statut
Si ce badge est vert, tous les tests Ultralytics CI sont actuellement réussis, garantissant le bon fonctionnement de toutes les Modes et Tâches sur macOS, Windows et Ubuntu.
Ceci est une réponse automatisée 🤖. Un ingénieur Ultralytics examinera également votre discussion prochainement. 😊
Glenn JocherUltralyticsBonjour! Cette discussion est dédiée à l'intégration de MLflow avec Ultralytics YOLO. MLflow est un excellent outil pour suivre vos expériences, mesurer les performances et gérer les artefacts de vos modèles YOLO.
N'hésitez pas à poser vos questions sur cette intégration dans les commentaires ci-dessous. Notre équipe Ultralytics sera ravie de vous aider avec la configuration, le suivi des expériences ou tout problème que vous pourriez rencontrer.
Si vous utilisez déjà MLflow avec Ultralytics, nous serions également intéressés par vos retours d'expérience et vos cas d'utilisation.
- Notebooks avec GPU gratuit :
BlitzBricksterYY-db
For multiple GPU training a YOLO model, with autolog enabled and also I manually logged a few items by using mlflow.log_xxx() and I also logged the customized PyFunc model, it turns out to be multiple runs are generated per training.
I want to only have 1 run and it better have all the logged artifacts.
- Can you give a best practice advice?
- which GPU rank in a multiple GPU training is used for autologging purpose? I suppose is it the highest rank GPU?
Thanks Best, Yang
BlitzBricksterYY-db
my code example:
from pyspark.ml.torch.distributor import TorchDistributor settings.update({"mlflow":True}) # if you do want to autolog. mlflow.autolog(disable = False) def train_fn(): from ultralytics import YOLO import torch import mlflow import torch.distributed as dist from ultralytics import settings from mlflow.types.schema import Schema, ColSpec from mlflow.models.signature import ModelSignature model = YOLO(f"{project_location}/raw_model/yolov8n.pt") model.train( batch=8, device=[0, 1], data="./coco8.yaml", # ref: https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/datasets/coco8.yaml epochs=5, project=f'{tmp_project_location}', exist_ok=True, fliplr=1, flipud=1, perspective=0.001, degrees=.45 ) # after training is done. if not dist.is_initialized(): # import torch.distributed as dist dist.init_process_group("nccl") rank = dist.get_rank() # we divert all other manual logging to the last GPU with max rank. if rank == torch.cuda.device_count() - 1: mlflow.log_params({"rank":rank}) yolo_wrapper = YOLOC(model.trainer.best) mlflow.pyfunc.log_model( artifact_path="model", artifacts={'model_path': str(model.trainer.save_dir), "best_point": str(model.trainer.best)}, python_model=yolo_wrapper, signature=signature ) mlflow.end_run() experiment_name = "/Users/yang.yang@databricks.com/ConvaTec Workshop Computer Vision/Experiments_YOLO_CoCo" mlflow.set_experiment(experiment_name) experiment_id = mlflow.get_experiment_by_name(experiment_name).experiment_id # os.environ['MLFLOW_EXPERIMENT_NAME'] = experiment_name os.environ['MLFLOW_EXPERIMENT_ID'] = experiment_id os.environ['MLFLOW_ENABLE_SYSTEM_METRICS_LOGGING'] = "true" with mlflow.start_run(experiment_id = experiment_id) as run: print("Active run_id: {}".format(run.info.run_id)) os.environ['MLFLOW_RUN_ID'] = run.info.run_id # try to register under the same run distributor = TorchDistributor(num_processes=2, local_mode=True, use_gpu=True ) distributor.run(train_fn)BlitzBricksterYY-db
Also, is it possible to enable multi-GPU training to use the project folder on Databricks Unity Catalog Volume directory?
e.g.,
project=f'{tmp_project_location}',from above code in themodel.train()args.
Glenn JocherUltralyticsTo achieve a single MLflow run during multi-GPU training, ensure manual logging is executed only once, typically on the main GPU (rank 0). For project directories, you can specify the Unity Catalog Volume directory by setting the
projectparameter to your desired path. This should work as expected on Databricks. If you encounter issues, verify the permissions and path configurations for the Unity Catalog Volume.BlitzBricksterYY-db(editado)
I have tested, in multi-GPU settings,
- dbfs path wont work in
project= - Unity Catalog path wont work in
project= - VM path such as
project='/tmp/project/'works.
I have included in the
train_fn()the os.environ() for both databricks_host and databricks_pat (token). This doesn't help.In a single GPU setting, these all work.
I suspect the credential env vars are not passed correctly to worker GPUs or threads. Could you verify?
- dbfs path wont work in
BlitzBricksterYY-db
2nd Issue: for autologging system metrics, there are confusions of logging each node, each GPU's system metrics into the MLflow run.
What is the best practice if I have two nodes and each node has two GPUs? for each node, it has GPU 0 and GPU 1 as local rank.
-
Plan A I tried: Logging all into the same MLflow run will yield 2 GPUs. Looks like overwritten issues happened between two nodes.
-
Logging into 4 different mlflow runs (each GPU created a run), then each run has its own GPU 0 and GPU 1, in total, I have 2*4 = 8 GPU system metrics, there must be duplications there.
What is the right way to do that?
-
Glenn JocherUltralytics@BlitzBricksterYY-db to avoid overwriting issues during multi-GPU and multi-node training with MLflow autologging, the best practice is usually to log system metrics from only the primary GPU (rank 0) on each node. This ensures unique entries and avoids duplication or conflicts. Configure logging in your script such that only rank 0 GPUs per node execute the logging logic. Let us know if further clarification is needed!
BlitzBricksterYY-db
For system metrics like memory and network bandwidth usage, it is automatically done in the .train() function. I have the need to monitor each GPU on each node, i.e., [[0,1],[0,1]], what is the best way to achieve that? Automatically logging system metrics will overwritten each other cause only record GPU_0 and GPU_1, no matter it is node_0 or node_1.
I know for now single node with multiGPU is a clear solution as it is [0,1,2,3], logged perfectly by your auto logging within .train()
Glenn JocherUltralytics@BlitzBricksterYY-db to monitor system metrics for each GPU on each node in a multi-node, multi-GPU setup, the best approach is to disable automatic MLflow autologging for system metrics and implement custom logging. You can use GPU-specific identifiers (e.g., node rank and GPU local rank) to avoid overwriting and ensure distinct logs for each GPU. For assistance with custom implementations, refer to the MLflow documentation. Let us know if further clarification is needed!
1
vini-c-costa
Hey
I'm facing a strange situation between yolov8 and mlflow. I created a local container to run a postgres environment, and then I tried the following script:
import pandas as pd from ultralytics import YOLO, settings import os import json import mlflow settings.reset() settings.update({ 'mlflow': True, 'clearml': False, 'comet': False, 'dvc': False, 'hub': False, 'neptune': False, 'raytune': False, 'tensorboard': False, 'wandb': False }) mlflow.set_experiment("Teste 01") mlflow.set_tracking_uri("postgresql://user:senha@localhost:5432/postgres") model_name = "yolov8n-cls.pt" model = YOLO(model_name) epochs = 3 image_size = 128 with mlflow.start_run(run_name=f'{model_name}-{epochs}-{image_size}'): model.train(name=f'{model_name}-{epochs}-{image_size}', data='dataset', epochs=epochs, imgsz=image_size, hsv_h=1, hsv_s=1, hsv_v=1)The reason I maintained some mlflow configs out of with statement is because I would like to preserve all parameters, metrics and artifacts generated automatically by yolo-mlflow settings.
And the error I found was during the db creation: the experiment_id generated by yolo is bigger than serial4 type, which is the type used as postgres-mlflow default (a bigint type would solve I suppose). So, due to that, I'm facing the following error:
Traceback (most recent call last): File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/base.py", line 1969, in _exec_single_context self.dialect.do_execute( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/default.py", line 922, in do_execute cursor.execute(statement, parameters) psycopg2.errors.NumericValueOutOfRange: value "653626019723855191" is out of range for type integer LINE 3: WHERE experiments.experiment_id = '653626019723855191' AND e... ^ The above exception was the direct cause of the following exception: Traceback (most recent call last): File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/store/db/utils.py", line 141, in make_managed_session yield session File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/store/tracking/sqlalchemy_store.py", line 386, in get_experiment return self._get_experiment( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/store/tracking/sqlalchemy_store.py", line 360, in _get_experiment .one_or_none() File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/orm/query.py", line 2774, in one_or_none return self._iter().one_or_none() # type: ignore File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/orm/query.py", line 2847, in _iter result: Union[ScalarResult[_T], Result[_T]] = self.session.execute( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/orm/session.py", line 2308, in execute return self._execute_internal( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/orm/session.py", line 2190, in _execute_internal result: Result[Any] = compile_state_cls.orm_execute_statement( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/orm/context.py", line 293, in orm_execute_statement result = conn.execute( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/base.py", line 1416, in execute return meth( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/sql/elements.py", line 517, in _execute_on_connection return connection._execute_clauseelement( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/base.py", line 1639, in _execute_clauseelement ret = self._execute_context( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/base.py", line 1848, in _execute_context return self._exec_single_context( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/base.py", line 1988, in _exec_single_context self._handle_dbapi_exception( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/base.py", line 2344, in _handle_dbapi_exception raise sqlalchemy_exception.with_traceback(exc_info[2]) from e File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/base.py", line 1969, in _exec_single_context self.dialect.do_execute( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/sqlalchemy/engine/default.py", line 922, in do_execute cursor.execute(statement, parameters) sqlalchemy.exc.DataError: (psycopg2.errors.NumericValueOutOfRange) value "653626019723855191" is out of range for type integer LINE 3: WHERE experiments.experiment_id = '653626019723855191' AND e... ^ [SQL: SELECT experiments.experiment_id AS experiments_experiment_id, experiments.name AS experiments_name, experiments.artifact_location AS experiments_artifact_location, experiments.lifecycle_stage AS experiments_lifecycle_stage, experiments.creation_time AS experiments_creation_time, experiments.last_update_time AS experiments_last_update_time FROM experiments WHERE experiments.experiment_id = %(experiment_id_1)s AND experiments.lifecycle_stage IN (%(lifecycle_stage_1_1)s, %(lifecycle_stage_1_2)s)] [parameters: {'experiment_id_1': '653626019723855191', 'lifecycle_stage_1_1': 'active', 'lifecycle_stage_1_2': 'deleted'}] (Background on this error at: https://sqlalche.me/e/20/9h9h) During handling of the above exception, another exception occurred: Traceback (most recent call last): File "/home/viniciusconti/carbigdata/repos/color-ml/yolo-test/yolo_impl.py", line 32, in <module> with mlflow.start_run(run_name=f'{model_name}-3-128'): File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/tracking/fluent.py", line 377, in start_run active_run_obj = client.create_run( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/tracking/client.py", line 339, in create_run return self._tracking_client.create_run(experiment_id, start_time, tags, run_name) File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/tracking/_tracking_service/client.py", line 132, in create_run return self.store.create_run( File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/store/tracking/sqlalchemy_store.py", line 462, in create_run experiment = self.get_experiment(experiment_id) File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/store/tracking/sqlalchemy_store.py", line 385, in get_experiment with self.ManagedSessionMaker() as session: File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/contextlib.py", line 153, in __exit__ self.gen.throw(typ, value, traceback) File "/home/viniciusconti/miniconda3/envs/carbigdata_py310/lib/python3.10/site-packages/mlflow/store/db/utils.py", line 155, in make_managed_session raise MlflowException(message=e, error_code=BAD_REQUEST) mlflow.exceptions.MlflowException: (psycopg2.errors.NumericValueOutOfRange) value "653626019723855191" is out of range for type integer LINE 3: WHERE experiments.experiment_id = '653626019723855191' AND e... ^ [SQL: SELECT experiments.experiment_id AS experiments_experiment_id, experiments.name AS experiments_name, experiments.artifact_location AS experiments_artifact_location, experiments.lifecycle_stage AS experiments_lifecycle_stage, experiments.creation_time AS experiments_creation_time, experiments.last_update_time AS experiments_last_update_time FROM experiments WHERE experiments.experiment_id = %(experiment_id_1)s AND experiments.lifecycle_stage IN (%(lifecycle_stage_1_1)s, %(lifecycle_stage_1_2)s)] [parameters: {'experiment_id_1': '653626019723855191', 'lifecycle_stage_1_1': 'active', 'lifecycle_stage_1_2': 'deleted'}] (Background on this error at: https://sqlalche.me/e/20/9h9h)I was wondering if I could just define the experiment_id as a serial4, maintaining it in MLFlow patterns (numerical order e.g 1, 2, 3, 4...). I thought about fix it manually by changing the type directly on my database, but I believe that there's a smarter way to fix it.
Does anyone know how to solve it?
Glenn JocherUltralytics@vini-c-costa it seems the issue is related to the experiment_id exceeding the integer range in PostgreSQL. You might consider altering the database schema to use a
bigintfor the experiment_id. This should resolve the out-of-range error. For further details on MLflow integration, please visit the Ultralytics documentation.
MoAbbasid
@pderrenger Ive set the tracking uri correctly to point to the sqlite database but the db itself created after the Yolo training run is empty and has experiments or run, this is automatically generated by Ultralytics right? so this is an issue on Ultralytics side?
Paula DerrengerUltralyticsHi @MoAbbasid,
Thanks for reaching out! The creation and management of the SQLite database for MLflow is generally handled by MLflow itself rather than directly by Ultralytics. If the database is empty after a training run, it's possible that the MLflow tracking URI might not be set correctly, or the logging calls are not being executed as expected.
Here are a few things you can check:
- Ensure that MLflow is correctly installed and that the MLflow server is running during your training sessions.
- Verify that the
MLFLOW_TRACKING_URIis correctly pointing to your SQLite database. - Confirm that MLflow logging is enabled in your Ultralytics settings and that the logging functions (
on_pretrain_routine_end,on_fit_epoch_end, andon_train_end) are being triggered.
If these steps don't resolve the issue, it might be helpful to check the MLflow logs for any errors or warnings that could provide more insight.
Let us know how it goes or if you need further assistance!
MoAbbasid
how to set mlflow to true using CLI?
MoAbbasid
Im trying to log YOLO experiments using mlflow on a remote server, any idea how to do that?
MoAbbasid
do I have to use python and with mlflow.start run()? or can I do something similar from the command line?
Paula DerrengerUltralyticsYou can enable MLflow logging directly from the command line without needing to use
mlflow.start_run(). Simply update the settings using the CLI:yolo settings mlflow=TrueThen, you can set the project and run names when training your model:
yolo train project=my_project name=my_runThis will handle the logging of metrics, parameters, and artifacts automatically. For more detailed guidance, please refer to the documentation. Happy logging! 🚀
MoAbbasid
@pderranger, when I run this and check my mlflow server running locally I dont see the experiment results, only when I specify it in python mlflow.start run() do I see the run
MoAbbasid
and mlflow.autolog doesnt work, could you provide a working example of all segmentation results logged to mlflow?
Paula DerrengerUltralytics(editado)Thank you for your feedback! To ensure all segmentation results are logged to MLflow, you can follow these steps:
-
Enable MLflow Logging:
yolo settings mlflow=True -
Set Project and Run Names:
export MLFLOW_EXPERIMENT_NAME=my_experiment export MLFLOW_RUN=my_run -
Train Your Model:
yolo segment train project=my_project name=my_run
If
mlflow.autologisn't working as expected, you can manually log metrics and artifacts using themlflowlibrary within your training script. Here’s a basic example:import mlflow from ultralytics import YOLO # Start an MLflow run with mlflow.start_run(): # Load and train your model model = YOLO('yolov8n-seg.pt') results = model.train(data='coco128-seg.yaml', epochs=100, imgsz=640) # Log metrics and artifacts mlflow.log_metrics({'mAP50_95': results.box.map}) mlflow.log_artifacts(str(model.trainer.save_dir))This approach ensures all relevant data is logged to your MLflow server. For more detailed guidance, please refer to the MLflow Integration Documentation. If you have any further questions, feel free to ask! 😊
-
MoAbbasid
@pderranger, hi thanks I managed to get it to run for training with your code, but my val runs dont show up, any idea how I can log val mode in mlflow?
MoAbbasid
@pderranger, also it is reading from "runs/mlflow" directory fine but when I try to set up a sqlite db it fails to read the logged files
MoAbbasid
@pderranger, hi when setting mlflow=True, Yolov8 logs it under runs/mlflow, which includes the experiments, runs, and artifacts and there is an mlflow.db file, if I want these logs to be an sqlite format and have mlflow read them, what can I do? USING CLI
MoAbbasid
even tho --backend-store-uri runs the logged experiments correctly, the "runs/mlflow.db" database folder has empty experiments and runs, this means it an Ultralytics logging issue not an mlflow issue, please advise
MoAbbasid
@pderranger, how do I get yolov8 to automatically log to sqlite database?
Paula DerrengerUltralytics(editado)Hello @MoAbbasid,
For logging validation runs in MLflow, you can manually log the metrics after validation. Here's a basic example:
import mlflow from ultralytics import YOLO model = YOLO('yolov8n-seg.pt') results = model.val(data='coco128-seg.yaml') # Log the validation metrics you need, for example mAP50-95 mlflow.log_metrics({'mAP50_95': results.box.map})Regarding your issue with setting up a SQLite database for MLflow, ensure that your MLflow server is correctly configured to use the SQLite backend. You might need to specify the SQLite URI when starting the MLflow server:
mlflow server --backend-store-uri sqlite:///mlflow.db --default-artifact-root ./artifactsMake sure the path to the SQLite database (
mlflow.db) is correct and accessible. If you continue to face issues, checking the MLflow server logs for any error messages could provide more insights.If you need further assistance, feel free to ask!
MoAbbasid
@pderranger, I initiated the server using your code: "mlflow server --backend-store-uri sqlite:///mlflow.db --default-artifact-root ./artifacts", and I have set mlflow=True in yolo settings but after I train the results are still logged to directory runs/mlflow and not the sqlite database, I also tried setting "export SET_TRACKING_URI=/runs/mlflow.db" but that threw an error. any advise about how to have the logs stored in an sql database in the CLI automatically?
Paula DerrengerUltralyticsHello @MoAbbasid,
It seems like there might be a misunderstanding in setting the MLflow tracking URI. The
SET_TRACKING_URIenvironment variable you tried setting is not recognized by MLflow. Instead, you should set theMLFLOW_TRACKING_URIenvironment variable to point to your SQLite database. Here’s how you can do it:-
Set the MLflow Tracking URI:
export MLFLOW_TRACKING_URI=sqlite:///mlflow.db -
Start the MLflow Server: Ensure you start the MLflow server with the correct backend store URI if you are using a server:
mlflow server --backend-store-uri sqlite:///mlflow.db --default-artifact-root ./artifacts -
Run Your Training: With the environment variable set, start your training as usual. MLflow should now log directly to the SQLite database specified.
If you continue to face issues, please ensure that the path to the SQLite database is correct and accessible, and that there are no permission issues. Also, verify that the MLflow version you are using supports the features and configurations you are trying to use.
Let me know if this helps or if you need further assistance!
-