Method loss

examples/word-embedding/train_tok-embed.cc:136–162

    p_start = model.add_parameters({EMBED_DIM});
  }
  Expression loss(ComputationGraph& cg, const Expression& v, const string& code) {
    decoder.new_graph(cg);
    Expression h = tanh(v);
    vector<Expression> init = {v, h};
    decoder.start_new_sequence(init);
    Expression start = parameter(cg, p_start);
    PrefixNode* cur = &pfc->root;
    decoder.add_input(start);
    size_t i = 0;
    vector<Expression> errs(code.size());
    while(i < code.size()) {
      assert(cur);
      Expression pred = decoder.back();
      Expression rp = parameter(cg, cur->pred);
      Expression bias = parameter(cg, cur->bias);
      Expression p = logistic(dot_product(pred, rp) + bias);
      // maybe squared error instead of xentropy?
      if (code[i] == '0') p = 1.f - p;
      errs[i] = log(p);
      Expression cond = parameter(cg, code[i] == '0' ? cur->zero_cond : cur->one_cond);
      decoder.add_input(cond);
      cur = code[i] == '0' ? cur->zero_child : cur->one_child;
      ++i;
    }
    assert(cur->terminal);
    return -sum(errs);
  }
};

template <class Builder>