def compute_gleu(source, references, prediction_path):
"""get sentence-level gleu scores"""
sys.stderr.write('Running GLEU...\n')
gleu_calculator = GLEU(4)
gleu_calculator.load_sources(source)
num_iterations = 200
gleu_calculator.load_references(references)
return np.array(
[float(g[0]) for g in gleu_calculator.run_iterations(num_iterations=num_iterations,
#num_references=len(references),
source=source,
hypothesis=prediction_path,
per_sent=True)])
def gleu_scores(source,
references,
systems,
ngrams_len=4,
num_iterations=500,
debug=False):
# if there is only one reference, just do one iteration
if len(references) == 1:
num_iterations = 1
gleu_calculator = GLEU(ngrams_len)
if isinstance(source, six.string_types):
gleu_calculator.load_sources(source)
else:
gleu_calculator.set_sources(source)
if isinstance(references[0], six.string_types):
gleu_calculator.load_references(references)
else:
gleu_calculator.set_references(references)
total = []
per_sentence = []
for hpath in systems:
if isinstance(hpath, six.string_types):
with open(hpath) as instream:
hyp = [line.split() for line in instream]
if not debug:
print(os.path.basename(hpath), )
else:
instream = hpath
hyp = [line.split() for line in instream]
# first generate a random list of indices, using a different seed
# for each iteration
indices = []
for j in range(num_iterations):
random.seed(j * 101)
indices.append([
random.randint(0,
len(references) - 1) for i in range(len(hyp))
])
if debug:
print()
print('===== Sentence-level scores =====')
print('SID Mean Stdev 95%CI GLEU')
iter_stats = [[0 for i in range(2 * ngrams_len + 2)]
for j in range(num_iterations)]
for i, h in enumerate(hyp):
gleu_calculator.load_hypothesis_sentence(h)
# we are going to store the score of this sentence for each ref
# so we don't have to recalculate them 500 times
stats_by_ref = [None for r in range(len(references))]
for j in range(num_iterations):
ref = indices[j][i]
this_stats = stats_by_ref[ref]
if this_stats is None:
this_stats = [
s for s in gleu_calculator.gleu_stats(i, r_ind=ref)
]
stats_by_ref[ref] = this_stats
iter_stats[j] = [
sum(scores) for scores in zip(iter_stats[j], this_stats)
]
per_sentence.append(
get_gleu_stats([
gleu_calculator.gleu(stats, smooth=True)
for stats in stats_by_ref
]))
if debug:
# sentence-level GLEU is the mean GLEU of the hypothesis
# compared to each reference
for r in range(len(references)):
if stats_by_ref[r] is None:
stats_by_ref[r] = [
s for s in gleu_calculator.gleu_stats(i, r_ind=r)
]
print(i, )
print(' '.join(per_sentence[-1]))
total.append(
get_gleu_stats(
[gleu_calculator.gleu(stats) for stats in iter_stats]))
if debug:
print('\n==== Overall score =====')
print('Mean Stdev 95%CI GLEU')
print(' '.join(total[-1]))
else:
print("total", total[-1][0])
return total, per_sentence
help='path to src sentences')
parser.add_argument('-r',
'--ref',
nargs='*',
required=True,
help='references to use')
parser.add_argument('-d',
'--debug',
default=False,
action='store_true',
help='print debugging messages')
parser.add_argument('-c',
'--cand',
nargs='*',
required=True,
help='candidate(s) to score')
args = parser.parse_args()
gleu_calculator = GLEU(4)
gleu_calculator.load_sources(args.src)
num_iterations = 200
gleu_calculator.load_references(args.ref)
for cand in args.cand:
print cand, [
float(g[0]) for g in gleu_calculator.run_iterations(
num_iterations=num_iterations,
source=args.src,
hypothesis=cand,
per_sent=False)
][0]