Model not saved after training

[iamyihwa]

I have trained a NER however there was no model saved after the training was completed.

From what I have seen in sequence_tagger_trainer.py, save_model = True by default, and I didn't give any change to that.

SequenceTagger and SequenceTaggerTrainer were used.

This is the part of codes used to train.

trainer.train('resources/taggers/poa-ner', learning_rate=0.1, mini_batch_size=2, max_epochs=20)
When I look at the content of the directory where I saved the training data, there is no model which ends in .pt.

[ec2-user@ip-172-31-28-32 poa-ner]$ ls
loss.txt  test.tsv

For this case, I had to reduce the mini_batch_size = 2, because the sentence was very very long. ( About 3000 words long each. In fact it was a whole document, because in this document, there are dependencies between sentences, so our teammate wanted to put it together.. I don't know if using whole document as a sentence is a good idea or not.. any ideas on that also? ). I wonder if this could have been the cause of the problem?

The training seems to have completed okay ..

[stefan-it]

Have you installed the latest version of flair? The save_model = True default was recently introduced with #86, so with a version prior to #86 the model was not saved :(

[iamyihwa]

@stefan-it You are right!! Actually I installed with pip, but I was looking at the codes on the git repository (the most recent version!), and in fact my version has that save_model = False!
I will do that !

So I guess instead of installing through pip, I should git clone from master branch.
Uhm ... I realize now that I made a couple of changes to the flair locally (e.g. accepting different type of embedding, etc.) , so I just decided to set save_model = True in the file sequence_tagger_trainer.py.
Are there lots of differences btwn pip version and master branch? other than this one?

[stefan-it]

Exactly, you can clone the master branch and then just use pip install --user -e . - then the cloned version of flair will be installed :)

[iamyihwa]

thanks @stefan-it ! Will try that! :-)

[alanakbik]

@iamyihwa It's probably better to split the document into sentences and use a larger mini-batch-size. We find that a mini-batch-size of 16 or 32 generally works best if you have enough data.

Also, I note that test_mode is set to True in your code. If you want the best results, it's better to set test_mode to False because then the data gets shuffled at each epoch. You should use test_mode=True only for debugging.

So, better use:

trainer = SequenceTaggerTrainer(tagger, corpus, test_mode=False)

Or leave out the test_mode parameter since it is False by default.

[iamyihwa]

@alanakbik thanks!!!
will try with those!
just changed test_mode = False!