#!/usr/bin/env python
from dictlearn.extractive_qa_training import evaluate_extractive_qa
from dictlearn.extractive_qa_configs import qa_config_registry
from dictlearn.main import main_evaluate
if __name__ == "__main__":
main_evaluate(qa_config_registry, evaluate_extractive_qa)
#!/usr/bin/env python
from dictlearn.nli_training import evaluate
from dictlearn.main import main_evaluate
from dictlearn.nli_esim_config_registry import nli_esim_config_registry
from functools import partial
if __name__ == "__main__":
main_evaluate(nli_esim_config_registry, partial(evaluate, model="esim"))
print "aggregated stats:", aggregated
print "# of parameters {}".format(n_params)
#TODO: check that different batch_size yields same validation error than
# end of training validation error.
# TODO: I think blocks aggreg is simply mean which should break
# when we use masks??? investigate
if not os.path.exists(dest_path):
os.makedirs(dest_path)
if part == 'test_unseen':
np.savez(
os.path.join(dest_path, "predictions"),
words=input_data['words'],
words_mask=input_data['words_mask'],
#unk_ratio = to_save['unk_ratio'],
#def_unk_ratio = to_save['def_unk_ratio'],
proba_out=to_save['languagemodel_apply_proba_out'],
vocab_in=lm._vocab.words[:c['num_input_words']],
vocab_out=lm._vocab.words[:c['num_output_words']])
json.dump(aggregated,
open(os.path.join(dest_path, "aggregates.json"), "w"),
sort_keys=True,
indent=2)
if __name__ == "__main__":
main_evaluate(lm_config_registry, evaluate_lm)