def benchmark_ap(vocab, gt_label, num_references, mil_prob, eval_file=None):
n_words = len(vocab['words'])
P = np.zeros(mil_prob.shape, dtype=np.float)
R = np.zeros(mil_prob.shape, dtype=np.float)
score = np.zeros(mil_prob.shape, dtype=np.float)
ap = np.zeros((1, n_words), dtype=np.float)
human_prec = np.zeros((1, n_words), dtype=np.float)
human_rec = np.zeros((1, n_words), dtype=np.float)
prec_at_human_rec = np.zeros((1, n_words), dtype=np.float)
rec_at_human_prec = np.zeros((1, n_words), dtype=np.float)
rec_at_half_prec = np.zeros((1, n_words), dtype=np.float)
prec_at_human_rec[...] = np.nan
for i in range(len(vocab['words'])):
utils.tic_toc_print(1,
'benchmarking : {:4d} / {:4d}'.format(i, n_words))
P[:, i], R[:, i], score[:, i], ap[0, i] = cap_eval_utils.calc_pr_ovr(
gt_label[:, i], mil_prob[:, i], num_references)
human_prec[0, i], human_rec[0, i] = cap_eval_utils.human_agreement(
gt_label[:, i], num_references)
ind = np.where(R[:, i] >= human_rec[0, i])[0]
if len(ind) > 0:
ind = np.sort(ind)
prec_at_human_rec[0, i] = P[ind[0], i]
ind = np.where(P[:, i] >= human_prec[0, i])[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_human_prec[0, i] = R[ind[-1], i]
ind = np.where(P[:, i] >= 0.5)[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_half_prec[0, i] = R[ind[-1], i]
details = {'precision': P, 'recall': R, 'ap': ap, 'score': score, \
'prec_at_human_rec': prec_at_human_rec, 'rec_at_human_prec': rec_at_human_prec, \
'human_prec': human_prec, 'human_rec': human_rec, 'rec_at_half_prec': rec_at_half_prec}
if eval_file is not None:
eval_file_details_keys = details.keys()
eval_file_details_vals = [details[x] for x in eval_file_details_keys]
utils.save_variables(eval_file,
eval_file_details_vals,
eval_file_details_keys,
overwrite=True)
return details
def benchmark_ap(vocab, gt_label, num_references, mil_prob, eval_file = None):
n_words = len(vocab['words'])
P = np.zeros(mil_prob.shape, dtype = np.float)
R = np.zeros(mil_prob.shape, dtype = np.float)
score = np.zeros(mil_prob.shape, dtype = np.float)
ap = np.zeros((1, n_words), dtype = np.float)
human_prec = np.zeros((1, n_words), dtype = np.float)
human_rec = np.zeros((1, n_words), dtype = np.float)
prec_at_human_rec = np.zeros((1, n_words), dtype = np.float)
rec_at_human_prec = np.zeros((1, n_words), dtype = np.float)
rec_at_half_prec = np.zeros((1, n_words), dtype = np.float)
prec_at_human_rec[...] = np.nan
for i in range(len(vocab['words'])):
utils.tic_toc_print(1, 'benchmarking : {:4d} / {:4d}'.format(i, n_words))
P[:,i], R[:,i], score[:,i], ap[0,i] = cap_eval_utils.calc_pr_ovr(gt_label[:,i], mil_prob[:,i], num_references)
human_prec[0,i], human_rec[0,i] = cap_eval_utils.human_agreement(gt_label[:,i], num_references)
ind = np.where(R[:,i] >= human_rec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
prec_at_human_rec[0,i] = P[ind[0], i];
ind = np.where(P[:,i] >= human_prec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_human_prec[0,i] = R[ind[-1], i];
ind = np.where(P[:,i] >= 0.5)[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_half_prec[0,i] = R[ind[-1], i];
details = {'precision': P, 'recall': R, 'ap': ap, 'score': score, \
'prec_at_human_rec': prec_at_human_rec, 'rec_at_human_prec': rec_at_human_prec, \
'human_prec': human_prec, 'human_rec': human_rec, 'rec_at_half_prec': rec_at_half_prec};
if eval_file is not None:
eval_file_details_keys = details.keys();
eval_file_details_vals = [details[x] for x in eval_file_details_keys];
utils.save_variables(eval_file, eval_file_details_vals, eval_file_details_keys, overwrite=True);
return details;
def benchmark_one_word(i, P, R, score, ap, human_rec, human_prec, \
prec_at_human_rec, rec_at_human_prec, rec_at_half_prec,\
gt_label, mil_prob, num_references):
P[:,i], R[:,i], score[:,i], ap[0,i] = cap_eval_utils.calc_pr_ovr(gt_label[:,i], mil_prob[:,i], num_references)
human_prec[0,i], human_rec[0,i] = cap_eval_utils.human_agreement(gt_label[:,i], num_references)
ind = np.where(R[:,i] >= human_rec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
prec_at_human_rec[0,i] = P[ind[0], i];
ind = np.where(P[:,i] >= human_prec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_human_prec[0,i] = R[ind[-1], i];
ind = np.where(P[:,i] >= 0.5)[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_half_prec[0,i] = R[ind[-1], i];
def benchmark_one_word(i, P, R, score, ap, human_rec, human_prec, \
prec_at_human_rec, rec_at_human_prec, rec_at_half_prec,\
gt_label, mil_prob, num_references):
P[:, i], R[:, i], score[:, i], ap[0, i] = cap_eval_utils.calc_pr_ovr(
gt_label[:, i], mil_prob[:, i], num_references)
human_prec[0, i], human_rec[0, i] = cap_eval_utils.human_agreement(
gt_label[:, i], num_references)
ind = np.where(R[:, i] >= human_rec[0, i])[0]
if len(ind) > 0:
ind = np.sort(ind)
prec_at_human_rec[0, i] = P[ind[0], i]
ind = np.where(P[:, i] >= human_prec[0, i])[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_human_prec[0, i] = R[ind[-1], i]
ind = np.where(P[:, i] >= 0.5)[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_half_prec[0, i] = R[ind[-1], i]
def benchmark(imdb, vocab, gt_label, num_references, detection_file, eval_file = None):
# Get ground truth
# dt = utils.scio.loadmat(detection_file)
dt = utils.load_variables(detection_file)
mil_prob = dt['mil_prob'];
# Benchmark the output, and return a result struct
n_words = len(vocab['words'])
P = np.zeros(mil_prob.shape, dtype = np.float)
R = np.zeros(mil_prob.shape, dtype = np.float)
score = np.zeros(mil_prob.shape, dtype = np.float)
ap = np.zeros((1, n_words), dtype = np.float)
human_prec = np.zeros((1, n_words), dtype = np.float)
human_rec = np.zeros((1, n_words), dtype = np.float)
prec_at_human_rec = np.zeros((1, n_words), dtype = np.float)
rec_at_human_prec = np.zeros((1, n_words), dtype = np.float)
rec_at_half_prec = np.zeros((1, n_words), dtype = np.float)
prec_at_human_rec[...] = np.nan
for i in range(len(vocab['words'])):
utils.tic_toc_print(1, 'benchmarking : {:4d} / {:4d}'.format(i, n_words))
P[:,i], R[:,i], score[:,i], ap[0,i] = cap_eval_utils.calc_pr_ovr(gt_label[:,i], mil_prob[:,i], num_references)
human_prec[0,i], human_rec[0,i] = cap_eval_utils.human_agreement(gt_label[:,i], num_references)
ind = np.where(R[:,i] >= human_rec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
prec_at_human_rec[0,i] = P[ind[0], i];
ind = np.where(P[:,i] >= human_prec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_human_prec[0,i] = R[ind[-1], i];
ind = np.where(P[:,i] >= 0.5)[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_half_prec[0,i] = R[ind[-1], i];
# # print '{:20s}: {:.3f}'.format(vocab['words'][i], ap[0,i]*100)
details = {'precision': P, 'recall': R, 'ap': ap, 'score': score, \
'prec_at_human_rec': prec_at_human_rec, 'rec_at_human_prec': rec_at_human_prec, \
'human_prec': human_prec, 'human_rec': human_rec, 'rec_at_half_prec': rec_at_half_prec};
# Collect statistics over the POS
agg = [];
for pos in list(set(vocab['poss'])):
ind = [i for i,x in enumerate(vocab['poss']) if pos == x]
print " {:5s} [{:4d}] : {:5.2f} {:5.2f} {:5.2f}". \
format(pos, len(ind), 100*np.mean(ap[0, ind]), 100*np.mean(prec_at_human_rec[0, ind]), \
100*np.mean(human_prec[0, ind]))
agg.append({'pos': pos, 'ap': 100*np.mean(ap[0, ind]), \
'prec_at_human_rec': 100*np.mean(prec_at_human_rec[0, ind]), \
'human_prec': 100*np.mean(human_prec[0, ind])})
ind = range(len(vocab['words'])); pos = 'all';
print " {:5s} [{:4d}] : {:5.2f} {:5.2f} {:5.2f}". \
format(pos, len(ind), 100*np.mean(ap[0, ind]), 100*np.mean(prec_at_human_rec[0, ind]), \
100*np.mean(human_prec[0, ind]))
agg.append({'pos': pos, 'ap': 100*np.mean(ap[0, ind]), \
'prec_at_human_rec': 100*np.mean(prec_at_human_rec[0, ind]), \
'human_prec': 100*np.mean(human_prec[0, ind])})
if eval_file is not None:
try:
utils.save_variables(eval_file, [details, agg],
['details', 'agg'], overwrite = True)
except:
print 'Error trying to save to pickle, will try hdf5 hack';
namespace = globals().copy()
namespace.update(locals())
code.interact(local=namespace)
eval_file_details = eval_file.replace('.pkl','_details.h5');
eval_file_details_keys = details.keys();
eval_file_details_vals = [details[x] for x in eval_file_details_keys];
utils.save_variables(eval_file_details, eval_file_details_vals, eval_file_details_keys, overwrite=True);
eval_file_agg = eval_file.replace('.pkl','_agg.h5');
eval_file_agg_keys = agg.keys();
eval_file_agg_vals = [agg[x] for x in eval_file_agg_keys];
utils.save_variables(eval_file_agg, eval_file_agg_vals, eval_file_agg_keys, overwrite=True);
return details
def benchmark(imdb, vocab, gt_label, num_references, detection_file, eval_file = None):
# Get ground truth
# dt = utils.scio.loadmat(detection_file)
dt = utils.load_variables(detection_file)
mil_prob = dt['mil_prob'];
# Benchmark the output, and return a result struct
n_words = len(vocab['words'])
P = np.zeros(mil_prob.shape, dtype = np.float)
R = np.zeros(mil_prob.shape, dtype = np.float)
score = np.zeros(mil_prob.shape, dtype = np.float)
ap = np.zeros((1, n_words), dtype = np.float)
human_prec = np.zeros((1, n_words), dtype = np.float)
human_rec = np.zeros((1, n_words), dtype = np.float)
prec_at_human_rec = np.zeros((1, n_words), dtype = np.float)
rec_at_human_prec = np.zeros((1, n_words), dtype = np.float)
rec_at_half_prec = np.zeros((1, n_words), dtype = np.float)
prec_at_human_rec[...] = np.nan
for i in range(len(vocab['words'])):
utils.tic_toc_print(1, 'benchmarking : {:4d} / {:4d}'.format(i, n_words))
P[:,i], R[:,i], score[:,i], ap[0,i] = cap_eval_utils.calc_pr_ovr(gt_label[:,i], mil_prob[:,i], num_references)
human_prec[0,i], human_rec[0,i] = cap_eval_utils.human_agreement(gt_label[:,i], num_references)
ind = np.where(R[:,i] >= human_rec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
prec_at_human_rec[0,i] = P[ind[0], i];
ind = np.where(P[:,i] >= human_prec[0,i])[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_human_prec[0,i] = R[ind[-1], i];
ind = np.where(P[:,i] >= 0.5)[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_half_prec[0,i] = R[ind[-1], i];
# # print '{:20s}: {:.3f}'.format(vocab['words'][i], ap[0,i]*100)
details = {'precision': P, 'recall': R, 'ap': ap, 'score': score, \
'prec_at_human_rec': prec_at_human_rec, 'rec_at_human_prec': rec_at_human_prec, \
'human_prec': human_prec, 'human_rec': human_rec, 'rec_at_half_prec': rec_at_half_prec};
# Collect statistics over the POS
agg = [];
for pos in list(set(vocab['poss'])):
ind = [i for i,x in enumerate(vocab['poss']) if pos == x]
print " {:5s} [{:4d}] : {:5.2f} {:5.2f} {:5.2f}". \
format(pos, len(ind), 100*np.mean(ap[0, ind]), 100*np.mean(prec_at_human_rec[0, ind]), \
100*np.mean(human_prec[0, ind]))
agg.append({'pos': pos, 'ap': 100*np.mean(ap[0, ind]), \
'prec_at_human_rec': 100*np.mean(prec_at_human_rec[0, ind]), \
'human_prec': 100*np.mean(human_prec[0, ind])})
ind = range(len(vocab['words'])); pos = 'all';
print " {:5s} [{:4d}] : {:5.2f} {:5.2f} {:5.2f}". \
format(pos, len(ind), 100*np.mean(ap[0, ind]), 100*np.mean(prec_at_human_rec[0, ind]), \
100*np.mean(human_prec[0, ind]))
agg.append({'pos': pos, 'ap': 100*np.mean(ap[0, ind]), \
'prec_at_human_rec': 100*np.mean(prec_at_human_rec[0, ind]), \
'human_prec': 100*np.mean(human_prec[0, ind])})
if eval_file is not None:
utils.save_variables(eval_file, [details, agg, vocab, imdb],
['details', 'agg', 'vocab', 'imdb'], overwrite = True)
return details
mapping = {'NNS': 'NN', 'NNP': 'NN', 'NNPS': 'NN', 'NN': 'NN', \
'VB': 'VB', 'VBD': 'VB', 'VBN': 'VB', 'VBZ': 'VB', 'VBP': 'VB', 'VBP': 'VB', 'VBG': 'VB', \
'JJR': 'JJ', 'JJS': 'JJ', 'JJ': 'JJ', 'DT': 'DT', 'PRP': 'PRP', 'PRP$': 'PRP', 'IN': 'IN'};
# punctuations to be removed from the sentences
punctuations = ["''", "'", "``", "`", "-LRB-", "-RRB-", "-LCB-", "-RCB-", \
".", "?", "!", ",", ":", "-", "--", "...", ";"]
vocab = preprocess.get_vocab(imset, coco_caps, punctuations, mapping);
sg_utils.save_variables('vocab_' + imset + '.pkl', \
[vocab[x] for x in vocab.keys()], \
vocab.keys(), \
overwrite = True);
##
N_WORDS = 1000;
vocab = preprocess.get_vocab_top_k(vocab, N_WORDS)
image_ids = coco_caps.getImgIds()
counts = preprocess.get_vocab_counts(image_ids, coco_caps, 5, vocab)
P = np.zeros((N_WORDS, 1), dtype = np.float);
R = np.zeros((N_WORDS, 1), dtype = np.float);
for i, w in enumerate(vv['words']):
P[i], R[i] = cap_eval_utils.human_agreement(counts[:,i], 5)
print w, P[i], R[i]
for pos in list(set(vocab['poss'])):
ind = [i for i,x in enumerate(vocab['poss']) if pos == x]
print "{:5s} [{:3d}] : {:.2f} {:.2f} ".format(pos, len(ind), 100*np.mean(P[ind]), 100*np.mean(R[ind]))
def benchmark(imdb,
vocab,
gt_label,
num_references,
detection_file,
eval_file=None):
# Get ground truth
# dt = utils.scio.loadmat(detection_file)
dt = utils.load_variables(detection_file)
mil_prob = dt['mil_prob']
# Benchmark the output, and return a result struct
n_words = len(vocab['words'])
P = np.zeros(mil_prob.shape, dtype=np.float)
R = np.zeros(mil_prob.shape, dtype=np.float)
score = np.zeros(mil_prob.shape, dtype=np.float)
ap = np.zeros((1, n_words), dtype=np.float)
human_prec = np.zeros((1, n_words), dtype=np.float)
human_rec = np.zeros((1, n_words), dtype=np.float)
prec_at_human_rec = np.zeros((1, n_words), dtype=np.float)
rec_at_human_prec = np.zeros((1, n_words), dtype=np.float)
rec_at_half_prec = np.zeros((1, n_words), dtype=np.float)
prec_at_human_rec[...] = np.nan
for i in range(len(vocab['words'])):
utils.tic_toc_print(1,
'benchmarking : {:4d} / {:4d}'.format(i, n_words))
P[:, i], R[:, i], score[:, i], ap[0, i] = cap_eval_utils.calc_pr_ovr(
gt_label[:, i], mil_prob[:, i], num_references)
human_prec[0, i], human_rec[0, i] = cap_eval_utils.human_agreement(
gt_label[:, i], num_references)
ind = np.where(R[:, i] >= human_rec[0, i])[0]
if len(ind) > 0:
ind = np.sort(ind)
prec_at_human_rec[0, i] = P[ind[0], i]
ind = np.where(P[:, i] >= human_prec[0, i])[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_human_prec[0, i] = R[ind[-1], i]
ind = np.where(P[:, i] >= 0.5)[0]
if len(ind) > 0:
ind = np.sort(ind)
rec_at_half_prec[0, i] = R[ind[-1], i]
# # print '{:20s}: {:.3f}'.format(vocab['words'][i], ap[0,i]*100)
details = {'precision': P, 'recall': R, 'ap': ap, 'score': score, \
'prec_at_human_rec': prec_at_human_rec, 'rec_at_human_prec': rec_at_human_prec, \
'human_prec': human_prec, 'human_rec': human_rec, 'rec_at_half_prec': rec_at_half_prec}
# Collect statistics over the POS
agg = []
for pos in list(set(vocab['poss'])):
ind = [i for i, x in enumerate(vocab['poss']) if pos == x]
print " {:5s} [{:4d}] : {:5.2f} {:5.2f} {:5.2f}". \
format(pos, len(ind), 100*np.mean(ap[0, ind]), 100*np.mean(prec_at_human_rec[0, ind]), \
100*np.mean(human_prec[0, ind]))
agg.append({'pos': pos, 'ap': 100*np.mean(ap[0, ind]), \
'prec_at_human_rec': 100*np.mean(prec_at_human_rec[0, ind]), \
'human_prec': 100*np.mean(human_prec[0, ind])})
ind = range(len(vocab['words']))
pos = 'all'
print " {:5s} [{:4d}] : {:5.2f} {:5.2f} {:5.2f}". \
format(pos, len(ind), 100*np.mean(ap[0, ind]), 100*np.mean(prec_at_human_rec[0, ind]), \
100*np.mean(human_prec[0, ind]))
agg.append({'pos': pos, 'ap': 100*np.mean(ap[0, ind]), \
'prec_at_human_rec': 100*np.mean(prec_at_human_rec[0, ind]), \
'human_prec': 100*np.mean(human_prec[0, ind])})
if eval_file is not None:
try:
utils.save_variables(eval_file, [details, agg], ['details', 'agg'],
overwrite=True)
except:
print 'Error trying to save to pickle, will try hdf5 hack'
namespace = globals().copy()
namespace.update(locals())
code.interact(local=namespace)
eval_file_details = eval_file.replace('.pkl', '_details.h5')
eval_file_details_keys = details.keys()
eval_file_details_vals = [
details[x] for x in eval_file_details_keys
]
utils.save_variables(eval_file_details,
eval_file_details_vals,
eval_file_details_keys,
overwrite=True)
eval_file_agg = eval_file.replace('.pkl', '_agg.h5')
eval_file_agg_keys = agg.keys()
eval_file_agg_vals = [agg[x] for x in eval_file_agg_keys]
utils.save_variables(eval_file_agg,
eval_file_agg_vals,
eval_file_agg_keys,
overwrite=True)
return details
'VB': 'VB', 'VBD': 'VB', 'VBN': 'VB', 'VBZ': 'VB', 'VBP': 'VB', 'VBP': 'VB', 'VBG': 'VB', \
'JJR': 'JJ', 'JJS': 'JJ', 'JJ': 'JJ', 'DT': 'DT', 'PRP': 'PRP', 'PRP$': 'PRP', 'IN': 'IN'}
# punctuations to be removed from the sentences
punctuations = ["''", "'", "``", "`", "-LRB-", "-RRB-", "-LCB-", "-RCB-", \
".", "?", "!", ",", ":", "-", "--", "...", ";"]
vocab = preprocess.get_vocab(imset, coco_caps, punctuations, mapping)
sg_utils.save_variables('vocab_' + imset + '.pkl', \
[vocab[x] for x in vocab.keys()], \
vocab.keys(), \
overwrite = True)
##
N_WORDS = 1000
vocab = preprocess.get_vocab_top_k(vocab, N_WORDS)
image_ids = coco_caps.getImgIds()
counts = preprocess.get_vocab_counts(image_ids, coco_caps, 5, vocab)
P = np.zeros((N_WORDS, 1), dtype=np.float)
R = np.zeros((N_WORDS, 1), dtype=np.float)
for i, w in enumerate(vv['words']):
P[i], R[i] = cap_eval_utils.human_agreement(counts[:, i], 5)
print w, P[i], R[i]
for pos in list(set(vocab['poss'])):
ind = [i for i, x in enumerate(vocab['poss']) if pos == x]
print "{:5s} [{:3d}] : {:.2f} {:.2f} ".format(pos, len(ind),
100 * np.mean(P[ind]),
100 * np.mean(R[ind]))
