Plotting TRAINING DICE and VALIDATION LOSS in tensorboard

[vlgonz]

Hi, I am using monai event handlers, specifically tensorboard, to plot training loss and validation dice score. Besides, I have seen a couple of examples in monai notebooks provided in this github. However, I haven't found any information about how to plot training dice and validation loss using tensorboard handler, can anyone provide an example on how to do this? Thank you very much.

[KumoLiu]

Hi @vlgonz, by default, 'TensorBoardStatsHandler' plots loss and metric; therefore, you must ensure that the metric and loss are calculated in the correct engine.
https://github.com/Project-MONAI/MONAI/blob/e9e273858b6904659b0cd581d357662dc5c782f6/monai/handlers/tensorboard_handlers.py#L79-L83
You can set output_transform=lambda x: x to plot loss.

tutorials/3d_classification/ignite/densenet_training_dict.py

Line 149 in 570b19e

output_transform=lambda x: None, # no need to plot loss value, so disable per iteration output

Hope it help, thanks.

[vlgonz]

Okay, thank you for the answer. I have already tried that (commented lines below) but it does not work. This is the code part related to validation and train ignite handlers, any different idea on how to log validation loss and training dice?

val_handlers = [
# TensorBoardStatsHandler(log_dir="../runs/Mar11_03-16_00/val/graphs/FOLD_{}".format(index),
# tag_name="val_loss", output_transform=lambda x:x),

    TensorBoardStatsHandler(log_dir="../runs/Mar11_03-16_00/val/graphs/FOLD_{}".format(index),
        output_transform=lambda x:None, global_epoch_transform=lambda x: evaluator.state.epoch),
    CheckpointSaver(save_dir="../runs/Mar11_03-16_00/val/model_checkpoints/FOLD_{}".format(index), save_dict={"net": net}, save_key_metric=True)]

evaluator = SupervisedEvaluator(
    device=device,
    val_data_loader=val_loaders[index],
    network=net,
    inferer=SlidingWindowInferer(roi_size=(64,64,64), sw_batch_size=4, overlap=0.5, progress=True),
    postprocessing=val_post_transforms,
    key_val_metric={
        "val_mean_dice": MeanDice(include_background=False,output_transform=from_engine(["pred", "label"]))},
    val_handlers=val_handlers)

train_handlers = [

    ValidationHandler(validator=evaluator, interval=1, epoch_level=True),
    TensorBoardStatsHandler(log_dir="../runs/Mar11_03-16_00/train/graphs/FOLD_{}".format(index), tag_name="train_loss", output_transform=from_engine(["loss"], first=True))]

trainer = SupervisedTrainer(
    device=device,
    max_epochs=100,
    train_data_loader=train_loaders[index],
    network=net,
    optimizer=opt,
    loss_function=loss,
    inferer=SimpleInferer(),
    train_handlers=train_handlers,
)