Ensure additional fitting works after loading models - Githubissues

allenai / deep_qa

A deep NLP library, based on Keras / tf, focused on question answering (but useful for other NLP too)

Apache License 2.0

404 stars 132 forks source link

Ensure additional fitting works after loading models #233

Closed nelson-liu closed 7 years ago

nelson-liu commented 7 years ago

This PR adds some tests to verify that we're able to fit loaded, serialized models on additional data for fine-tuning.

[X] Add a tests for fitting after loading for MultipleTrueFalseMemoryNetwork, which should already be implemented correctly
[x] Fix _set_max_lengths_from_model for TrueFalseModel
[x] Add and pass tests for fitting after loading for TrueFalseModel
[x] Fix _set_max_lengths_from_model for BiDAF
[x] Add and pass tests for fitting after loading for BiDAF

nelson-liu commented 7 years ago

@matt-gardner --- theano tests seem to fail on my machine for even MultipleTrueFalseMemoryNetwork, which is interesting considering TF is fine / this model should already be correct. I'm looking into it, but wanted to let you know just in case you had any idea off the top of your head why this could be the case.

nelson-liu commented 7 years ago

interesting, seems like TF is failing on the CPU as well when trying to fit MultipleTrueFalseMemoryNetwork after loading.

nelson-liu commented 7 years ago

@matt-gardner this is (finally) ready to be looked at (except maybe for the problem I messaged you on slack about). a more real-world test would be to load an actual weights file from one of the training epochs of BiDAF on SQuAD (still running...) and see if I can get fitting after loading to work on that.

nelson-liu commented 7 years ago

merging this, thanks for looking @matt-gardner