Retraining models

[chaitjo]

Is it possible to retrain models (for example, the one's from https://github.com/smartschat/cort/blob/master/COREFERENCE.md#model-downloads) with new data?

I tried training using-

cort-train -in new_retraining_data.conll \
           -out pretrained_model.obj \
           -extractor cort.coreference.approaches.mention_ranking.extract_substructures \
           -perceptron cort.coreference.approaches.mention_ranking.RankingPerceptron \
           -cost_function cort.coreference.cost_functions.cost_based_on_consistency \
           -n_iter 5 \ 
           -cost_scaling 100 \
           -random_seed 23

but I think it overwrites the model.

[smartschat]

If I understand you correctly, you want to take a model as input, and then initialize training on new data with that model. Is that what you want to do?

[chaitjo]

Yes. For example, take http://smartschat.de/downloads/model-pair-train.obj and continue training on it using some new data.

[smartschat]

That's not implemented, but the code can be adapted. Unfortunately, I will not be able to have a closer look at this during this week. If you want to do it by yourself, I can give you some pointers.

[chaitjo]

Please do!

I'll create a pull request if I am able to successfully implement this.

[smartschat]

The constructor of perceptrons.pyx has priors and weights parameters. However, for training, these are overwritten in the fit method.

You need to make this overwriting optional, for example by adding a boolean parameter which controls whether weights/priors should be initialized or not. Then you also need to adapt the experiments.py-API training/predicting scripts with the new parameters.

I hope this helps. if you have any questions, I'm happy to answer them!

[rakesh-malviya]

Hi Sebastian,

Can you give approximate information on the amount of time it took you to train CORT on CONLL data? What was the hardware you used ?

Thanks and regards, Rakesh Malviya

[smartschat]

Hi Rakesh,

training the ranking model takes around two minutes per epoch. Preprocessing takes ~ 20 minutes if I remember correctly. Due to high memory requirements I train the models on a server with > 100GB RAM, using ~20 2.3GHz CPUs. However, only preprocessing is parallelized.