def evaluate()

in dialogue_personalization/model/common_layer.py [0:0]

• 39 lines of code
• 16 McCabe index (conditional complexity)

def evaluate(model, data, model_name='trs', ty='valid', writer=None, n_iter=0, ty_eval="before", verbose=False ):

    dial,ref, hyp_b= [],[],[]
    t = Translator(model, model.vocab)

    l = []
    p = []
    ent_b = []
    
    pbar = tqdm(enumerate(data),total=len(data))
    for j, batch in pbar:
        loss, ppl, _ = model.train_one_batch(batch, train=False)
        l.append(loss)
        p.append(ppl)
        if((j<3 and ty != "test") or ty =="test"): 

            sent_b, _ = t.translate_batch(batch)

            for i in range(len(batch["target_txt"])):
                new_words = []
                for w in sent_b[i][0]:
                    if w==config.EOS_idx:
                        break
                    new_words.append(w)
                    if len(new_words)>2 and (new_words[-2]==w):
                        new_words.pop()
                
                sent_beam_search = ' '.join([model.vocab.index2word[idx] for idx in new_words])
                hyp_b.append(sent_beam_search)
                ref.append(batch["target_txt"][i])
                dial.append(batch['input_txt'][i])
                ent_b.append(bert.predict_label([sent_beam_search for _ in range(len(batch['persona_txt'][i]))], batch['persona_txt'][i]))

        pbar.set_description("loss:{:.4f} ppl:{:.1f}".format(np.mean(l),np.mean(p)))
        if(j>4 and ty=="train"): break
    loss = np.mean(l)
    ppl = np.mean(p)
    ent_b = np.mean(ent_b)
    bleu_score_b = moses_multi_bleu(np.array(hyp_b), np.array(ref), lowercase=True)
 
    if(verbose):
        print("----------------------------------------------------------------------")
        print("----------------------------------------------------------------------")
        print_all(dial,ref,hyp_b,max_print=3 if ty != "test" else 100000000 )
        print("EVAL\tLoss\tPeplexity\tEntl_b\tBleu_b")
        print("{}\t{:.4f}\t{:.4f}\t{:.2f}\t{:.2f}".format(ty,loss,ppl,ent_b,bleu_score_b))
    return loss,ppl,ent_b,bleu_score_b