def cal_ROUGE(generated, reference, is_corpus=False):
# ref and sample are both dict
# scorers = [
# (Bleu(4), ["Bleu_1", "Bleu_2", "Bleu_3", "Bleu_4"]),
# (Meteor(),"METEOR"),
# (Rouge(), "ROUGE_L"),
# (Cider(), "CIDEr")
# ]
# output rouge 1-4 and rouge L and rouge L from pycocoevaluate
ROUGEscore = [0.0] * 6
for idx, g in enumerate(generated):
score = [0.0] * 6
if is_corpus:
for order in range(4):
score[order] = rouge_n(g.split(),
[x.split() for x in reference[0]],
order + 1, 0.5)
score[4] = rouge_l(g.split(), [x.split() for x in reference[0]],
0.5)
score[5], _ = Rouge().compute_score(reference, {0: [g]})
else:
for order in range(4):
score[order] = rouge_n(g.split(), [reference[0][idx].split()],
order + 1, 0.5)
score[4] = rouge_l(g.split(), [reference[0][idx].split()], 0.5)
score[5], _ = Rouge().compute_score({0: [reference[0][idx]]},
{0: [g]})
#pdb.set_trace()
#print g, score
ROUGEscore = [r + score[idx] for idx, r in enumerate(ROUGEscore)]
#BLEUscore += nltk.translate.bleu_score.sentence_bleu(reference, g, weight)
ROUGEscore = [r / len(generated) for r in ROUGEscore]
return ROUGEscore
def run_eval(sess, inputs, model):
sess.run(inputs.initializer)
total_losses = []
total_sizes = []
references = []
candidates = []
while True:
try:
loss, true_word, sample_word = sess.run(
[model.loss, model.true_word, model.sample_word])
total_losses.append(loss)
total_sizes.append(len(sample_word))
candidates.append(decode_for_human(sample_word, EOS, join=False))
references.append(decode_for_human(true_word, EOS, join=False))
except tf.errors.OutOfRangeError:
avg_loss = get_avg_loss(total_losses, total_sizes)
print("Evaluation done. Avg eval loss: %f" % avg_loss)
candidates = list(chain(*candidates))
references = list(chain(*references))
#TODO add bleu and rouge
print(rouge_1(candidates[1], [references[1]], [0, 0.5]))
rouge_1_recall, rouge_1_F1 = list(
np.mean([
rouge_1(c, [r], [0, 0.5])
for c, r in zip(candidates, references)
],
axis=0))
rouge_2_recall, rouge_2_F1 = list(
np.mean([
rouge_2(c, [r], [0, 0.5])
for c, r in zip(candidates, references)
],
axis=0))
rouge_L_recall, rouge_L_F1 = list(
np.mean([
rouge_l(c, [r], [0, 0.5])
for c, r in zip(candidates, references)
],
axis=0))
print("rouge-1 recall: %.5f \t F1: %.5f" %
(rouge_1_recall, rouge_1_F1))
print("rouge-2 recall: %.5f \t F1: %.5f" %
(rouge_2_recall, rouge_2_F1))
print("rouge-L recall: %.5f \t F1: %.5f" %
(rouge_L_recall, rouge_L_F1))
for c, r in random.sample(list(zip(candidates, references)), 20):
print('\n'.join([
"candidate:" + ''.join(c), "reference:" + ''.join(r),
'------'
]))
break
return avg_loss
def _rouge_evaluation(self, predicts, answers):
rouge_1s = []
rouge_2s = []
rouge_ls = []
for predict, answer in zip(predicts, answers):
answer = [w.replace('_UNK', '_UNKNOWN') for w in answer]
rouge_1 = rougescore.rouge_1(predict, [answer], 0.5)
rouge_2 = rougescore.rouge_2(predict, [answer], 0.5)
rouge_l = rougescore.rouge_l(predict, [answer], 0.5)
rouge_1s.append(rouge_1)
rouge_2s.append(rouge_2)
rouge_ls.append(rouge_l)
return {
"rouge_1": np.mean(rouge_1s),
"rouge_2": np.mean(rouge_2s),
"rouge_l": np.mean(rouge_ls)
}
def test_rouge_l(self):
data = self.load_test_data()
rouge = RougeCalculator(stopwords=True)
for eval_id in data:
summaries = data[eval_id]["summaries"]
references = data[eval_id]["references"]
for s in summaries:
baseline = Pythonrouge(summary_file_exist=False,
summary=[[s]],
reference=[[[r] for r in references]],
n_gram=1,
recall_only=False,
ROUGE_L=True,
length_limit=True,
length=50,
stemming=False,
stopwords=True)
b1_v = baseline.calc_score()
b2_v = rouge_l(rouge.tokenize(s),
[rouge.tokenize(r) for r in references], 0.5)
v = rouge.rouge_l(s, references)
self.assertLess(abs(b2_v - v), 1e-5)
self.assertLess(abs(b1_v["ROUGE-L-F"] - v), 1e-5)
