How to interpret "perplexity" data

[Digital Mine]

Hello,

I'm new with Magenta, and I just started to train my model yesterday.

So far the results are very bad but I guess the model needs more time to train, at least I hope so.

I'm getting output similar to this one now:

```

Accuracy = 0.99274343, Global Step = 7841, Loss = 0.026616095, Perplexity = 1.0269735

```

Could you kindly help me to understand how to interpret them please.

Thanks in advance

[Ian Simon]

Not sure which model you're training, but the meaning of those terms is:

Accuracy: fraction of steps at which the model assigns highest probability to the actual event that occurred

Global Step: number of batches processed so far

Loss: the value the model is attempting to minimize; in this case it is likely a cross-entropy between the actual event and the distribution predicted by the model, averaged across all time steps in all sequences

Perplexity: a little tricky to define; in this case the perplexity is exp(loss), in general it is a measure of how well the model can predict each event: https://en.wikipedia.org/wiki/Perplexity

Your accuracy seems quite high and loss/perplexity quite low.  If these are being computed on the training data, my guess is you are overfitting.

-Ian

--
Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discu...@tensorflow.org
---
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discu...@tensorflow.org.

[Digital Mine]

thx Ian for your reply,

At this point I guess I'm doing something wrong, because when I try to generate something it is really unlistenable, but if the model data are so hight there must be a mistake somewhere, and I guess that it is in the training dataset.

I'm trying with some chopin midi, maybe they are noth enough in total are like less than 200 traks