Skip to content

Cannot reproduce perplexities in the word language model #415

@lezcano

Description

@lezcano

I've trained a few of the example models present in the folder word_language_model, and I have not been able to reproduce the results given in the README.
python main.py --cuda --epochs 6 # I get test ppl 142.07.
python main.py --cuda --epochs 6 --tied # I get a test perplexity of 130.32

I do get similar training perplexities to the ones given in the README, but I believe that this is not necessarily a good measure of how well the model is performing.

For the model
python main.py --cuda --emsize 650 --nhid 650 --dropout 0.5 --epochs 40
I get a training perplexity of about 52, and a test perplexity of 94.30. None of these numbers are even close to the supposed test perplexity of 80.97.

Can anyone reproduce these results or the results given in README.md?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions