Frequently Asked Questions¶
This document contains a number of questions that are regularly asked on GitHub Issues.
Why is my model not generating a response?¶
For a generative model, check that
--skip-generation is set to
Why can’t I reproduce the results of an evaluation on a task with a pretrained model?¶
One common culprit for this is that the flags for that task may not be
correctly set. When loading a pretrained checkpoint, all of the parameters for
the model itself will be loaded from the model’s
.opt file, but all
task-specific parameters will need to be re-specified.
If results differ by a few small decimal places, this can often be attributed to differences in hardware or software environment.
I want to generate a lot of responses to fixed utterances¶
The easiest way to do this is to create a
teacher in ParlAI Dialog Format. Then, use
eval_model with world logging to store all the responses:
parlai eval_model -t fromfile:parlaiformat --fromfile-datapath yourtextfile.txt \ -mf some_model_file --world-logs outputfile
Why is my generative model’s perplexity so high (>1000) when evaluating?¶
One first thing to check is whether there is a problem with your dictionary or token embeddings, because this high perplexity implies that the model is very bad at predicting the next token in a string of text.
I changed my teacher and now its tests won’t pass.¶
Take a careful look at the diff outputs that those tests produce. If the results look expected, then you can update the regression fixtures (stored, expected results) with:
pytest --force-regen parlai/tasks/TASK_NAME_HERE/test.py