monailabel.tasks.train.basic_train module

class monailabel.tasks.train.basic_train.BasicTrainTask(model_dir, description=None, config=None, amp=True, load_path=None, load_dict=None, publish_path=None, stats_path=None, train_save_interval=20, val_interval=1, n_saved=5, final_filename='checkpoint_final.pt', key_metric_filename='model.pt', model_dict_key='model', find_unused_parameters=False, load_strict=False, labels=None, disable_meta_tracking=False, tracking='mlflow', tracking_uri='', tracking_experiment_name=None)

Bases: monailabel.interfaces.tasks.train.TrainTask

This provides Basic Train Task to train a model using SupervisedTrainer and SupervisedEvaluator from MONAI

Parameters

• model_dir – Base Model Dir to save the model checkpoints, events etc…

• description – Description for this task

• config – K,V pairs to be part of user config

• amp – Enable AMP for training

• load_path – Initialize model from existing checkpoint (pre-trained)

• load_dict – Provide dictionary to load from checkpoint. If None, then net will be loaded

• publish_path – Publish path for best trained model (based on best key metric)

• stats_path – Path to save the train stats

• train_save_interval – checkpoint save interval for training

• val_interval – validation interval (run every x epochs)

• n_saved – max checkpoints to save

• final_filename – name of final checkpoint that will be saved

• key_metric_filename – best key metric model file name

• model_dict_key – key to save network weights into checkpoint

• find_unused_parameters – Applicable for DDP/Multi GPU training

• load_strict – Load pre-trained model in strict mode

• labels – Labels to be used as part of training context (some transform might need)

• disable_meta_tracking – Disable tracking for faster training rate (unless you are using MetaTensor/batched transforms)

• tracking – Tracking Manager for Experiment Management (only ‘mlflow’ is supported)

• tracking_uri – Tracking URI for Experiment Management

• tracking_experiment_name – Name for tracking experiment

TRAIN_METRIC_ACCURACY = 'train_acc'

TRAIN_METRIC_MEAN_DICE = 'train_mean_dice'

VAL_METRIC_ACCURACY = 'val_acc'

VAL_METRIC_MEAN_DICE = 'val_mean_dice'

__init__(model_dir, description=None, config=None, amp=True, load_path=None, load_dict=None, publish_path=None, stats_path=None, train_save_interval=20, val_interval=1, n_saved=5, final_filename='checkpoint_final.pt', key_metric_filename='model.pt', model_dict_key='model', find_unused_parameters=False, load_strict=False, labels=None, disable_meta_tracking=False, tracking='mlflow', tracking_uri='', tracking_experiment_name=None)

Parameters

• model_dir – Base Model Dir to save the model checkpoints, events etc…

• description – Description for this task

• config – K,V pairs to be part of user config

• amp – Enable AMP for training

• load_path – Initialize model from existing checkpoint (pre-trained)

• load_dict – Provide dictionary to load from checkpoint. If None, then net will be loaded

• publish_path – Publish path for best trained model (based on best key metric)

• stats_path – Path to save the train stats

• train_save_interval – checkpoint save interval for training

• val_interval – validation interval (run every x epochs)

• n_saved – max checkpoints to save

• final_filename – name of final checkpoint that will be saved

• key_metric_filename – best key metric model file name

• model_dict_key – key to save network weights into checkpoint

• find_unused_parameters – Applicable for DDP/Multi GPU training

• load_strict – Load pre-trained model in strict mode

• labels – Labels to be used as part of training context (some transform might need)

• disable_meta_tracking – Disable tracking for faster training rate (unless you are using MetaTensor/batched transforms)

• tracking – Tracking Manager for Experiment Management (only ‘mlflow’ is supported)

• tracking_uri – Tracking URI for Experiment Management

• tracking_experiment_name – Name for tracking experiment

cleanup(request)

config()

event_names(context)

finalize(context)

get_cache_dir(request)

info()

load_path(output_dir, pretrained=True)

abstract loss_function(context)

lr_scheduler_handler(context)

abstract network(context)

abstract optimizer(context)

partition_datalist(context, shuffle=False)

pre_process(request, datastore)

stats()

train(rank, world_size, request, datalist)

train_additional_metrics(context)

train_data_loader(context, num_workers=0, shuffle=True)

train_handlers(context)

train_inferer(context)

train_iteration_update(context)

train_key_metric(context)

abstract train_post_transforms(context)

abstract train_pre_transforms(context)

val_additional_metrics(context)

val_data_loader(context, num_workers=0)

val_handlers(context)

abstract val_inferer(context)

val_iteration_update(context)

val_key_metric(context)

val_post_transforms(context)

val_pre_transforms(context)

class monailabel.tasks.train.basic_train.Context

Bases: object

monailabel.tasks.train.basic_train.main_worker(rank, world_size, request, datalist, task)

previous

monailabel.tasks.train package

next

monailabel.tasks.train.bundle module