Log both loss and metrics on both train and eval dataset into a csv file

[anais2390]

Hello,
I am using monai librairy for segmentation and I've been looking for the cleanest way to log at each epoch:

• mean loss computation on train dataset
• mean metrics computations on train dataset
• mean loss computation on eval dataset
• mean metrics computations on eval dataset
• learning rate

I am using engine SupervisedTrainer and SupervisedEvaluator and the trickest part is that the output_transforms are different for loss computation (no postprocessing) and metrics computation (activation + asdiscrete).
Also I coudn't manage to get the loss values into the csv.

A typical csv file sould contain the following columns (if we use the MeanDice metric for exemple)

Epoch | MeanDice_train | MeanDice_eval | Loss_train | Loss_eval |

Thank you so much for your help.

Anaïs

[ericspod]

Hi @anais2390 you can use the MetricLogger class to capture all the metric and loss data during training and evaluation, probably with two separate instances. These will capture the values which can then be used to fill in a csv file afterward. You can also use event handlers to trigger at the end of an epoch to record the learning rate to a data structure somewhere, something like this:

opt = SomeOptimizer()
epoch_lrs=[]
...

@trainer.on(Events.EPOCH_COMPLETED)
def get_lr(engine):
    epoch_lrs.append(opt.lr)