log the output of process_function

[ydcjeff]

Hi,

I have setup to log at ITERATION_COMPLETED(every=100), how can I log the output of process_function at specified interval (100 here) with WandBLogger?
I have found engine.state.metrics not to be None in order to log. So, there is RunningAverage I found out. But, I don't think it fits my needs.
The output of the process_function is a dictionary with many keys with loss values.
So how can I do that? Should I use Loss or RunningAverage or other else?

Thank you!

[vfdev-5]

@ydcjeff how about this example :

ignite/examples/contrib/mnist/mnist_with_wandb_logger.py

Lines 105 to 110 in 915d69e

	wandb_logger.attach_output_handler(
	trainer,
	event_name=Events.ITERATION_COMPLETED(every=100),
	tag="training",
	output_transform=lambda loss: {"batchloss": loss},
	)

See also OutputHandler docs about output_transform.

In this case you will log loss value on the iteration and not averaged one. If you like to compute an average value, then use RunningAverage for that and then metric_names from OutputHandler.

[ydcjeff]

Thank you @vfdev-5 , this works!

[ydcjeff]

@vfdev-5 Thank you for your answer. Let me try it.
I have another question about RunningAverage and Loss.

1. If I want to calculate and log the loss at each iteration, I should go with the answer you provided?
2. If I want to calculate and log the loss at each epoch, RunninAverage and Loss can do that, I think. While in RunningAverage, I need to calculate the loss beforehand and attach to RunningAverage to get epoch-wise loss, in Loss, i just need to pass loss_fn to calculate the loss (epoch-wise, I think). What's the difference?
3. By default, RunningAverage do epoch-wise, what would it do in batch-wise?

[sdesrozis]

@ydcjeff Thanks for these questions

1. That is correct using Events.ITERATION_COMPLETED without every filter
2. For both, computations are accumulated during iterations. Using RunningAverage, the loss is computed once by process function, while the Loss object will call the loss function.
3. I mean you take about MetricUsage. I think it will log loss without accumulation. You can see in the following this specific usage
   

   ignite/ignite/metrics/metric.py

   Line 76 in 915d69e

   class BatchWise(MetricUsage):

HTH :)