[R] Is the mxnet always minimize the custom loss function?

[yiyusheng]

Package used (Python/R/Scala/Julia): R
Before I write my own custom loss function with mx.metric.custom, I checked the https://github.com/dmlc/mxnet/blob/master/R-package/R/metric.R.
I find two functions: mx.metric.accuracy is used as a metric for classification and mx.metric.rmse is used as a metric for regression.
So far as I know, the model should minimize the loss function like MSE. But minimize accuracy indicates a bad model. When I test the mx.metric.accuracy, the increasing accuracy of each epoch indicates the mxnet know how to treat the loss function.
This confuses me when I write my own loss function. I do not know when the mxnet will maximum the loss function and when the mxnet will minimize the loss function. Could you explain this stuff? Thank you.

[back2yes]

I guess adding a minus '-' in front of the loss function would turn the minimization into maxmization.

[yiyusheng]

@back2yes , yes.
But I am not sure that the mxnet will always minimize the loss function. Because I find one mx.metric.accuracy which should be maximized.

[marugari]

mx.metric are not loss functions to optimize, but for evaluation.

In your case, MXNet minimized log_loss (If you choose SoftmaxOutput).
It follows that the accuracy score increases.

[back2yes]

@yiyusheng I understand you. @marugari was right.
The 'metric' is just a score, it is evaluated after the weights and biases are updated.
Whereas the loss function is calculated before the update, the backward process is based on the loss function value.

[yiyusheng]

@marugari, thank you for correcting my fault. I'd like to use my own loss function. Do you know how to do that?

[yiyusheng]

@back2yes, yes, I recognize the mistake, thanks. Is there any idea to implement my own loss function?

[marugari]

1. add a new Output layer (C++)
   https://github.com/dmlc/mxnet/blob/e7514fe1b3265aaf15870b124bb6ed0edd82fa76/src/operator/softmax_output-inl.h

2. use mx.sym.MakeLoss (Python/R? I don't know details)
   custom loss symbol in R/Python #3368

[thirdwing]

In R, you can use mx.symbol.MakeLoss

[thirdwing]

@yiyusheng You should be able to find some simple examples in https://github.com/dmlc/mxnet/blob/master/docs/tutorials/r/CustomLossFunction.md

[yiyusheng]

@thirdwing thank you for your advice. I have sent an e-mail to you to detail my question. Could you have a look on it? Thank you very much.