def featureEvaluate(encoder, decoder, char2i, pairs, use_cuda):
    correct = 0
    total = 0
    i2char = {i: c for c, i in char2i.items()}
    i=0
    print_at = random.sample(range(1, 1000), 5)

    for inp, out in pairs:
        preds = featurePredict(encoder, decoder, inp, use_cuda)
        # Squeeze off the batch_size = 1 dim
        targets = out
        total+=1
        print("Evaluating dev example %i" % i)
        if i in print_at:
            print(''.join([i2char[int(c)] for c in preds]))
            print(''.join([i2char[int(c)] for c in targets]))

        i+=1
        if targets.equal(preds):
            correct += 1

    return "Accuracy: %.2f %% \n" % (correct / total * 100)

def featurePredict(encoder, decoder, inp, use_cuda):
    enc_out, enc_hidden = encoder(inp)

    decoder_input = Variable(torch.LongTensor([EOS_index]))

    dec_hidden = decoder.init_hidden()

    eos = 1
    preds = []
    while (eos < 2):
        dec_out, dec_hidden = decoder(decoder_input,\
                        dec_hidden, enc_out, use_cuda)

        topv, topi = dec_out.data.topk(1)
        # Get the index from the tensor as an integer
        topi_int = topi[0][0][0]
        if topi_int == EOS_index:
            eos += 1

        preds.append(topi_int)
        decoder_input = Variable(topi.view(1))

    return Variable(torch.LongTensor(preds))