def evaluate():
selection = sys.argv[1:]
ref_tr_vidnames = None
ref_te_vidnames = None
for cname in selection:
factor = 1.0
if '*' in cname:
factor, cname = cname.split('*')
factor = float(factor)
feature, params = combinations[cname]
print "load feature", cname, "factor", factor
tr_data, tr_labels, tr_vidnames = get_data(feature, 'train_balanced', **params)
te_data, te_labels, te_vidnames = get_data(feature, 'test_balanced', **params)
print "compute kernels train %d*%d test %d*%d" % (
tr_data.shape + te_data.shape)
if ref_tr_vidnames != None:
print "remapping names"
# pdb.set_trace()
te_data = remap_descriptors(te_data, te_vidnames, ref_te_vidnames)
tr_data = remap_descriptors(tr_data, tr_vidnames, ref_tr_vidnames)
Kxx, Kyx = data_to_kernels(tr_data, te_data)
Kxx *= factor
Kyx *= factor
if ref_tr_vidnames == None:
tr_kernel = Kxx
te_kernel = Kyx
ref_te_vidnames = te_vidnames
ref_tr_vidnames = tr_vidnames
else:
tr_kernel += Kxx
te_kernel += Kyx
from fisher_vectors.evaluation import trecvid12_parallel as eval
fit_out = eval.fit(tr_kernel, tr_labels)
print eval.score(te_kernel, te_labels, fit_out)
def vary_nr_negatives():
null_class_idx = 0
feature = 'mbh'
params = {
'dummy': {},
'mbh': {'suffix': '_morenull'},
'sift': {'subsample': 10, 'nr_clusters': 64, 'color': 0}}
tr_data, tr_labels, _ = get_data(feature, 'train', **params[feature])
te_data, te_labels, _ = get_data(feature, 'test', **params[feature])
outfilename = '/home/lear/oneata/data/trecvid12/results/tmp.txt'
with open(outfilename, 'a') as ff:
ff.write('%s %s\n' % (feature, params[feature].__str__()))
ii, nr_repeats = 0, 5
for ii in xrange(nr_repeats):
for proportion in (0.02, 0.04, 0.08, 0.1, 0.2, 0.4, 0.8, 1.):
if proportion < 1.0:
idxs = subsample_null_class(tr_labels, proportion, ii)
else:
idxs = np.arange(len(tr_labels))
random.seed(ii)
random.shuffle(idxs)
_tr_data, _tr_labels = tr_data[idxs], tr_labels[idxs]
tr_kernel, te_kernel = data_to_kernels(tr_data, te_data)
#eval = Evaluation('trecvid12', eval_type='trecvid11')
#score = eval.fit(tr_kernel, tr_labels).score(te_kernel, te_labels)
from fisher_vectors.evaluation import trecvid12_parallel as eval
fit_out = eval.fit(tr_kernel, _tr_labels)
score = eval.score(te_kernel, te_labels, fit_out)
print score
ff.write('%1.2f %d %2.3f \n' % (
proportion,
len(idxs) - len(_tr_labels[_tr_labels != null_class_idx]),
score))