def setup(self):
# These scores are achieved for the following evaluation parameters.
# class_weights='auto', C from 3 ** -2 .. 3 ** 8, and one shuffle
# split with test size of 0.25.
self.expected_scores = {'fv': 55.226, 'fv_sfv': 56.718, 'bow': 42.896}
src_cfg = 'hollywood2_clean'
nr_clusters = 100
self.dataset = Dataset(src_cfg,
ip_type='dense5.track15mbh',
suffix='.original',
nr_clusters=nr_clusters)
sstats_folder = self.dataset.SSTATS_DIR
gmm_fn = self.dataset.GMM
self.tr_fn = os.path.join(sstats_folder, 'train_1_1_1_0.dat')
self.sp_tr_fn = os.path.join(sstats_folder,
'spatial_train_1_1_1_0.dat')
self.tr_labels_fn = os.path.join(sstats_folder, 'labels_train.info')
self.te_fn = os.path.join(sstats_folder, 'test_1_1_1_0.dat')
self.sp_te_fn = os.path.join(sstats_folder, 'spatial_test_1_1_1_0.dat')
self.te_labels_fn = os.path.join(sstats_folder, 'labels_test.info')
self.gmm = load_gmm(gmm_fn)
tr_labels = pickle.load(open(self.tr_labels_fn, 'r'))
te_labels = pickle.load(open(self.te_labels_fn, 'r'))
self.cx = tr_labels
self.cy = te_labels
def setup(self):
# These scores are achieved for the following evaluation parameters.
# class_weights='auto', C from 3 ** -2 .. 3 ** 8, and one shuffle
# split with test size of 0.25.
self.expected_scores = {'fv': 55.226,
'fv_sfv': 56.718,
'bow': 42.896}
src_cfg = 'hollywood2_clean'
nr_clusters = 100
self.dataset = Dataset(src_cfg, ip_type='dense5.track15mbh',
suffix='.original', nr_clusters=nr_clusters)
sstats_folder = self.dataset.SSTATS_DIR
gmm_fn = self.dataset.GMM
self.tr_fn = os.path.join(sstats_folder, 'train_1_1_1_0.dat')
self.sp_tr_fn = os.path.join(sstats_folder,
'spatial_train_1_1_1_0.dat')
self.tr_labels_fn = os.path.join(sstats_folder, 'labels_train.info')
self.te_fn = os.path.join(sstats_folder, 'test_1_1_1_0.dat')
self.sp_te_fn = os.path.join(sstats_folder, 'spatial_test_1_1_1_0.dat')
self.te_labels_fn = os.path.join(sstats_folder, 'labels_test.info')
self.gmm = load_gmm(gmm_fn)
tr_labels = pickle.load(open(self.tr_labels_fn, 'r'))
te_labels = pickle.load(open(self.te_labels_fn, 'r'))
self.cx = tr_labels
self.cy = te_labels
def master(src_cfg, suffix_in ,suffix_out, K, N, nr_processes, double_norm):
D = 64
dataset = Dataset(src_cfg, nr_clusters=K)
samples = [str(sample) for sample in dataset.get_data('train')[0] +
dataset.get_data('test')[0]]
if double_norm:
worker = double_normalization
suffix = '.double_norm'
gmm = load_gmm(
os.path.join(
dataset.FEAT_DIR + suffix_in, 'gmm',
'gmm_%d' % K))
else:
worker = merge
suffix = ''
gmm = None
path_in = os.path.join(
dataset.FEAT_DIR + suffix_in,
'statistics_k_%d' % dataset.VOC_SIZE, 'stats.tmp')
path_out = os.path.join(
dataset.FEAT_DIR + suffix_out,
'statistics_k_%d' % dataset.VOC_SIZE, 'stats.tmp' + suffix)
sstats_in = SstatsMap(path_in)
sstats_out = SstatsMap(path_out)
len_sstats = dataset.VOC_SIZE + 2 * D * dataset.VOC_SIZE
kwargs = {
'N': N,
'sstats_in': sstats_in,
'sstats_out': sstats_out,
'len_sstats': len_sstats,
'gmm': gmm}
if nr_processes > 1:
nr_samples_per_process = len(samples) / nr_processes + 1
for ii in xrange(nr_processes):
mp.Process(target=worker,
args=(samples[
ii * nr_samples_per_process:
(ii + 1) * nr_samples_per_process], ),
kwargs=kwargs).start()
else:
worker(samples, **kwargs)
def master(src_cfg, suffix_in, suffix_out, K, N, nr_processes, double_norm):
D = 64
dataset = Dataset(src_cfg, nr_clusters=K)
samples = [
str(sample) for sample in dataset.get_data('train')[0] +
dataset.get_data('test')[0]
]
if double_norm:
worker = double_normalization
suffix = '.double_norm'
gmm = load_gmm(
os.path.join(dataset.FEAT_DIR + suffix_in, 'gmm', 'gmm_%d' % K))
else:
worker = merge
suffix = ''
gmm = None
path_in = os.path.join(dataset.FEAT_DIR + suffix_in,
'statistics_k_%d' % dataset.VOC_SIZE, 'stats.tmp')
path_out = os.path.join(dataset.FEAT_DIR + suffix_out,
'statistics_k_%d' % dataset.VOC_SIZE,
'stats.tmp' + suffix)
sstats_in = SstatsMap(path_in)
sstats_out = SstatsMap(path_out)
len_sstats = dataset.VOC_SIZE + 2 * D * dataset.VOC_SIZE
kwargs = {
'N': N,
'sstats_in': sstats_in,
'sstats_out': sstats_out,
'len_sstats': len_sstats,
'gmm': gmm
}
if nr_processes > 1:
nr_samples_per_process = len(samples) / nr_processes + 1
for ii in xrange(nr_processes):
mp.Process(target=worker,
args=(samples[ii * nr_samples_per_process:(ii + 1) *
nr_samples_per_process], ),
kwargs=kwargs).start()
else:
worker(samples, **kwargs)
def evaluate_given_dataset(dataset, **kwargs):
model_type = kwargs.get('model_type', 'fv')
sstats_folder = dataset.SSTATS_DIR
tr_fn = os.path.join(sstats_folder, 'train.dat')
tr_labels_fn = os.path.join(sstats_folder, 'labels_train.info')
te_fn = os.path.join(sstats_folder, 'test.dat')
te_labels_fn = os.path.join(sstats_folder, 'labels_test.info')
gmm = load_gmm(dataset.GMM)
tr_labels = pickle.load(open(tr_labels_fn, 'r'))
te_labels = pickle.load(open(te_labels_fn, 'r'))
model = Model(model_type, gmm)
model.compute_kernels([tr_fn], [te_fn])
Kxx, Kyx = model.get_kernels()
evaluation = Evaluation(dataset.DATASET, **kwargs)
print evaluation.fit(Kxx, tr_labels).score(Kyx, te_labels)
def evaluate_given_dataset(dataset, **kwargs):
model_type = kwargs.get('model_type', 'fv')
sstats_folder = dataset.SSTATS_DIR
tr_fn = os.path.join(sstats_folder, 'train.dat')
tr_labels_fn = os.path.join(sstats_folder, 'labels_train.info')
te_fn = os.path.join(sstats_folder, 'test.dat')
te_labels_fn = os.path.join(sstats_folder, 'labels_test.info')
gmm = load_gmm(dataset.GMM)
tr_labels = pickle.load(open(tr_labels_fn, 'r'))
te_labels = pickle.load(open(te_labels_fn, 'r'))
model = Model(model_type, gmm)
model.compute_kernels([tr_fn], [te_fn])
Kxx, Kyx = model.get_kernels()
evaluation = Evaluation(dataset.DATASET, **kwargs)
print evaluation.fit(Kxx, tr_labels).score(Kyx, te_labels)
def discriminative_detection_per_class(class_idx, **kwargs):
max_nr_iter = kwargs.get('max_nr_iter', 1)
#dataset = kwargs.get('dataset', Dataset(
# 'trecvid11_small', nr_clusters=128, suffix='.small.per_slice.delta_240'))
#'trecvid11_small', nr_clusters=128, suffix='.small.per_slice'))
src_cfg = kwargs.get('src_cfg')
nr_clusters = kwargs.get('nr_clusters')
suffix = kwargs.get('suffix')
outfile = kwargs.get('outfile', FILE % (src_cfg, class_idx))
nr_pos = kwargs.get('nr_pos', 10000)
nr_neg = kwargs.get('nr_neg', 10000)
agg_type = kwargs.get('agg_type', 'norm')
use_nr_descs = kwargs.get('use_nr_descs', False)
assert agg_type in ('norm', 'unnorm'), "Unknown aggregation type."
dataset = Dataset(src_cfg, nr_clusters=nr_clusters, suffix=suffix)
gmm = load_gmm(dataset.GMM)
tr_slice_data = get_slice_data_from_file(
dataset, 'train', class_idx, gmm, nr_pos, nr_neg)
te_slice_data = get_slice_data_from_file(
dataset, 'test', class_idx, gmm, nr_pos, nr_neg)
tr_sample_labels = tr_slice_data.get_sample_labels()
te_sample_labels = te_slice_data.get_sample_labels()
for ii in xrange(max_nr_iter):
print 'Iteration %d' % ii
# Feature pooling.
model = Model(gmm)
if ii == 0:
if not os.path.exists(outfile):
print 'Aggregating statistics by the number of descriptors...'
ss = tr_slice_data.get_aggregated_by_nr_descs()
np.array(ss, dtype=np.float32).tofile(outfile)
#tr_sample_sstats = [ss, ss]
tr_sample_sstats = [ss] * 2
else:
print 'Loaded aggregated statistics...'
# Cache results.
ss = np.fromfile(
outfile, dtype=np.float32).reshape((-1, model.D))
#tr_sample_sstats = [ss, ss]
tr_sample_sstats = [ss] * 2
else:
tr_sample_sstats = tr_slice_data.get_aggregated(
agg_type, use_nr_descs)
# Fisher vectors on pooled features.
tr_kernel = model.get_tr_kernel(tr_sample_sstats)
# Train classifier on pooled features.
_eval = Evaluation()
_eval = _eval.fit(tr_kernel, tr_sample_labels)
# Update weights.
tr_slice_data.update_scores(_eval, model)
te_slice_data.update_scores(_eval, model)
# TODO Save data.
tr_slice_data.save_htlist(ii)
te_slice_data.save_htlist(ii)
del _eval
del model
# Final retraining and evaluation.
tr_sample_sstats = tr_slice_data.get_aggregated(agg_type, use_nr_descs)
te_sample_sstats = te_slice_data.get_aggregated(agg_type, use_nr_descs)
model = Model(gmm)
tr_kernel = model.get_tr_kernel(tr_sample_sstats)
te_kernel = model.get_te_kernel(te_sample_sstats)
_eval = Evaluation()
_eval = _eval.fit(tr_kernel, tr_sample_labels)
score = _eval.score(te_kernel, te_sample_labels)
print 'Class %d score %2.3f' % (class_idx, score)
return score
