model_sentences.py 文件源码

def test(self, vocab_size, use_onto_lstm, S_ind_test=None, C_ind_test=None, hierarchical=False, base=2, oov_list=None):
    X_test = C_ind_test[:,:-1] if use_onto_lstm else S_ind_test[:,:-1] # remove the last words' hyps in all sentences
    Y_inds_test = S_ind_test[:,1:]
    if hierarchical:
      test_targets = self._factor_target_indices(Y_inds_test, vocab_size, base=base)
    else:
      test_targets = [self._make_one_hot(Y_inds_test, vocab_size)]
    print >>sys.stderr, "Evaluating model on test data"
    test_loss = self.model.evaluate(X_test, test_targets)
    print >>sys.stderr, "Test loss: %.4f"%test_loss
    if oov_list is not None:
      oov_inds = [self.dp.word_index[w] for w in oov_list]
      non_oov_Y_inds = numpy.copy(Y_inds_test)
      for ind in oov_inds:
    non_oov_Y_inds[non_oov_Y_inds == ind] = 0
      non_oov_test_targets = self._factor_target_indices(non_oov_Y_inds, vocab_size, base=base)
      non_oov_test_loss = self.model.evaluate(X_test, non_oov_test_targets)
      print >>sys.stderr, "Non-oov test loss: %.4f"%non_oov_test_loss
    factored_test_preds = [-((numpy.log(pred) * target).sum(axis=-1)) for pred, target in zip(self.model.predict(X_test), test_targets)]
    test_preds = sum(factored_test_preds)
    #non_null_probs = []
    #for test_pred, inds in zip(test_preds, Y_inds_test):
    #  wanted_probs = []
    #  for tp, ind in zip(test_pred, inds):
    #    if ind != 0:
    #      wanted_probs.append(tp)
    #  non_null_probs.append(wanted_probs)
    #return non_null_probs
    return test_preds