Why not choose the breakpoint with lowest test perplexity?

[MrSeven77]

Recently, I've been reproducing the paper's result using the oringinal data and this code.
The picture is a visulization from tensorboard, with record of testing perplexity from every 500 step. And from this picture, I noticed that the last step (ppl:37.76) doesn't have the lowest test perplexity (ppl:25.11). However, the value of the last step consists with paper's result(ppl:36.9) .
So, why not choose the breakpoint with lowest test perplexity? Or, what is the criterion of the convergence of the model?

[claude-zhou]

Criterion of convergence is that the perplexity has stabilized not that it reaches a lowest point.

[MrSeven77]

Criterion of convergence is that the perplexity has stabilized not that it reaches a lowest point.

Thanks for reply. However, when testing perplexity is stablized, is there a chance that this model is overfitted?

More over, when i changed the code for NLPCC 2017 dataset and trained the CVAE model, I got extremely large testing perplexty and after training for a while, the losses became NaN. What problem did I encounter? Are there any tutorials of how to train a CVAE model?

Note that the CVAE model is initialized with a pretrained seq2seq model, as the paper said.

[MrSeven77]

Thanks a lot for your reply.

[hqlin2018]

Criterion of convergence is that the perplexity has stabilized not that it reaches a lowest point.

Thanks for reply. However, when testing perplexity is stablized, is there a chance that this model is overfitted?

More over, when i changed the code for NLPCC 2017 dataset and trained the CVAE model, I got extremely large testing perplexty and after training for a while, the losses became NaN. What problem did I encounter? Are there any tutorials of how to train a CVAE model?

Note that the CVAE model is initialized with a pretrained seq2seq model, as the paper said.

hello , do you solve the problem that using the NLPCC 2017 dataset to trained the CVAE model?
and also that is i want to do now.