def language_eval(dataset, preds, preds_n, eval_kwargs, split):
model_id = eval_kwargs['id']
eval_oracle = eval_kwargs.get('eval_oracle', 0)
# create output dictionary
out = {}
if len(preds_n) > 0:
# vocab size and novel sentences
if 'coco' in dataset:
dataset_file = 'data/dataset_coco.json'
elif 'flickr30k' in dataset or 'f30k' in dataset:
dataset_file = 'data/dataset_flickr30k.json'
training_sentences = set([' '.join(__['tokens']) for _ in json.load(open(dataset_file))['images'] if not _['split'] in ['val', 'test'] for __ in _['sentences']])
generated_sentences = set([_['caption'] for _ in preds_n])
novels = generated_sentences - training_sentences
out['novel_sentences'] = float(len(novels)) / len(preds_n)
tmp = [_.split() for _ in generated_sentences]
words = []
for _ in tmp:
words += _
out['vocab_size'] = len(set(words))
# encoder.FLOAT_REPR = lambda o: format(o, '.3f')
cache_path = os.path.join('eval_results/', '.cache_'+ model_id + '_' + split + '.json')
coco = getCOCO(dataset)
valids = coco.getImgIds()
# filter results to only those in MSCOCO validation set
preds_filt = [p for p in preds if p['image_id'] in valids]
mean_perplexity = sum([_['perplexity'] for _ in preds_filt]) / len(preds_filt)
mean_entropy = sum([_['entropy'] for _ in preds_filt]) / len(preds_filt)
print('using %d/%d predictions' % (len(preds_filt), len(preds)))
json.dump(preds_filt, open(cache_path, 'w')) # serialize to temporary json file. Sigh, COCO API...
cocoRes = coco.loadRes(cache_path)
cocoEval = COCOEvalCap(coco, cocoRes)
cocoEval.params['image_id'] = cocoRes.getImgIds()
cocoEval.evaluate()
for metric, score in cocoEval.eval.items():
out[metric] = score
# Add mean perplexity
out['perplexity'] = mean_perplexity
out['entropy'] = mean_entropy
imgToEval = cocoEval.imgToEval
for k in list(imgToEval.values())[0]['SPICE'].keys():
if k != 'All':
out['SPICE_'+k] = np.array([v['SPICE'][k]['f'] for v in imgToEval.values()])
out['SPICE_'+k] = (out['SPICE_'+k][out['SPICE_'+k]==out['SPICE_'+k]]).mean()
for p in preds_filt:
image_id, caption = p['image_id'], p['caption']
imgToEval[image_id]['caption'] = caption
if len(preds_n) > 0:
import eval_multi
cache_path_n = os.path.join('eval_results/', '.cache_'+ model_id + '_' + split + '_n.json')
spice_n = eval_multi.eval_spice_n(dataset, preds_n, model_id, split)
out.update(spice_n['overall'])
div_stats = eval_multi.eval_div_stats(dataset, preds_n, model_id, split)
out.update(div_stats['overall'])
if eval_oracle:
oracle = eval_multi.eval_oracle(dataset, preds_n, model_id, split)
out.update(oracle['overall'])
else:
oracle = None
self_cider = eval_multi.eval_self_cider(dataset, preds_n, model_id, split)
out.update(self_cider['overall'])
with open(cache_path_n, 'w') as outfile:
json.dump({'spice_n': spice_n, 'div_stats': div_stats, 'oracle': oracle, 'self_cider': self_cider}, outfile)
out['bad_count_rate'] = sum([count_bad(_['caption']) for _ in preds_filt]) / float(len(preds_filt))
outfile_path = os.path.join('eval_results/', model_id + '_' + split + '.json')
with open(outfile_path, 'w') as outfile:
json.dump({'overall': out, 'imgToEval': imgToEval}, outfile)
return out
def language_eval(dataset, preds, preds_n, eval_kwargs, split):
model_id = eval_kwargs['id']
eval_oracle = eval_kwargs.get('eval_oracle', 0)
import sys
sys.path.append("coco-caption")
annFile = 'coco-caption/annotations/captions_val2014.json'
from pycocotools.coco import COCO
from pycocoevalcap.eval import COCOEvalCap
# encoder.FLOAT_REPR = lambda o: format(o, '.3f')
if not os.path.isdir('eval_results'):
os.mkdir('eval_results')
cache_path = os.path.join('eval_results/',
'.cache_' + model_id + '_' + split + '.json')
coco = COCO(annFile)
valids = coco.getImgIds()
# filter results to only those in MSCOCO validation set (will be about a third)
preds_filt = [p for p in preds if p['image_id'] in valids]
mean_perplexity = sum([_['perplexity']
for _ in preds_filt]) / len(preds_filt)
mean_entropy = sum([_['entropy'] for _ in preds_filt]) / len(preds_filt)
print('using %d/%d predictions' % (len(preds_filt), len(preds)))
json.dump(preds_filt,
open(cache_path,
'w')) # serialize to temporary json file. Sigh, COCO API...
cocoRes = coco.loadRes(cache_path)
cocoEval = COCOEvalCap(coco, cocoRes)
cocoEval.params['image_id'] = cocoRes.getImgIds()
cocoEval.evaluate()
# create output dictionary
out = {}
for metric, score in cocoEval.eval.items():
out[metric] = score
# Add mean perplexity
out['perplexity'] = mean_perplexity
out['entropy'] = mean_entropy
imgToEval = cocoEval.imgToEval
for k in list(imgToEval.values())[0]['SPICE'].keys():
if k != 'All':
out['SPICE_' + k] = np.array(
[v['SPICE'][k]['f'] for v in imgToEval.values()])
out['SPICE_' + k] = (out['SPICE_' + k][out['SPICE_' +
k] == out['SPICE_' +
k]]).mean()
for p in preds_filt:
image_id, caption = p['image_id'], p['caption']
imgToEval[image_id]['caption'] = caption
if len(preds_n) > 0:
cache_path_n = os.path.join(
'eval_results/', '.cache_' + model_id + '_' + split + '_n.json')
spice_n = eval_multi.eval_spice_n(preds_n, model_id, split)
out.update(spice_n['overall'])
div_stats = eval_multi.eval_div_stats(preds_n, model_id, split)
out.update(div_stats['overall'])
if eval_oracle:
oracle = eval_multi.eval_oracle(preds_n, model_id, split)
out.update(oracle['overall'])
with open(cache_path_n, 'w') as outfile:
json.dump(
{
'spice_n': spice_n,
'div_stats': div_stats,
'oracle': oracle
}, outfile)
out['bad_count_rate'] = sum([count_bad(_['caption'])
for _ in preds_filt]) / float(len(preds_filt))
outfile_path = os.path.join('eval_results/',
model_id + '_' + split + '.json')
with open(outfile_path, 'w') as outfile:
json.dump({'overall': out, 'imgToEval': imgToEval}, outfile)
return out
