def score(args):
if args.dataset == 'coco':
import vg.vendrov_provider as dp
elif args.dataset == 'places':
import vg.places_provider as dp
elif args.dataset == 'flickr8k':
import vg.flickr8k_provider as dp
logging.info('Loading data')
prov = dp.getDataProvider(args.dataset, root=args.root, audio_kind='mfcc')
tokenize = characters if args.text else lambda x: x['audio']
config = dict(split=args.split, tokenize=tokenize, batch_size=args.batch_size)
if args.text:
config['encode_sentences'] = encode_texts
scorer = Scorer(prov, config)
output = []
for path in args.model:
task = load(path)
task.eval().cuda()
rsa = scorer.rsa_image(task)
para = scorer.retrieval_para(task)
result = dict(path=path, rsa=rsa, para=para)
if not args.text:
result['speaker_id'] = scorer.speaker_id(task)
output.append(result)
json.dump(output, open(args.output, 'w'), indent=2)
import numpy
import vg.simple_data as sd
import vg.experiment as E
import vg.vendrov_provider as dp
import vg.defn.baseline_mfcc as D
dataset = 'coco'
prov = dp.getDataProvider(dataset, root='../..', audio_kind='mfcc')
def audio(sent):
return sent['audio']
eval_config = dict(tokenize=audio, split='val', encode_sentences=D.aggregate_mean, para=True)
numpy.save("scores.1.npy", D.eval_para(prov, eval_config))
def __init__(self, root='.', truncate=None, load_images=True):
self.places = places.getDataProvider('places', root=root, truncate=truncate, load_images=load_images)
self.coco = vendrov.getDataProvider('coco', root=root, truncate=truncate, load_images=load_images)
self.speakers = self.places.speakers.union(self.coco.speakers)