search

DTUComputeCognitiveSystems / deep_detektor

Automated Factual-Claim Detection in Danish Broadcasting
Other
2 stars 0 forks source link

Move all model and training-definitions into the constructors of models #18

Open NorthGuard opened 7 years ago

NorthGuard commented 7 years ago

No parameters should be passed on to the fit() method. This allows for predefining multiple versions of the same model, with different parameters, for training.

sfvnielsen commented 7 years ago

First attempt has been made to do this. See latest commit. One problem persists - if models are constructed before train/test split the placeholder dimensions can be incosistent due to the following call (in single_training.py)

# Make and set BoW-vocabulary
bow_vocabulary = the_tensor_provider.extract_programs_vocabulary(train_idx)
the_tensor_provider.set_bow_vocabulary(bow_vocabulary)

Current solution is to create train/test split before running all fit() methods (see new run file model_comparison.py)