Does eval_metric affect model training?

[Gazillionth]

Hi,

The following reflects my understanding of training a network with MXNet. I'd appreciate it if someone could let me know if it's not correct:

• The final output layer of a network defines the loss function with respect to which the network will be optimized, e.g. using mxnet.symbol.LogisticRegressionOutput means that the cross-entropy is being minimized (at least, I suppose it's cross entropy, but one way or another, this symbol defines the loss function).
• The eval_metric argument supplied to the model.fit function serves only to report to the user the performance of the model with respect to some prediction task (e.g. like top-5 correct), but it does not affect the training process at all.

Is this correct? Or does eval_metric somehow define the loss function that is optimized during training?

Thanks,

Gaz.

[piiswrong]

It's just for reporting

[Gazillionth]

Thanks, good to have that confirmed. As a follow-up question, that must mean that when training with a single set of hyperparameters, the validation set is only for reporting too, right?

[pluskid]

Yes, validation set is only for reporting. Except you use some customized learning rate scheduler that adjust the learning rate according to the performance on the validation set, etc.