def __init__(self):
self.batch_size = 16
self.nb_classes = 20
self.prop = 10
class_info = load_ids()
self.id2name = {int(info['id']): info['name'] for info in class_info.values()}
# options = utils.parse_options_file('pv_baseline')
# config_utils.update_config(options)
self.data_dir = '/home/caleml/datasets/pascalvoc/VOCdevkit/VOC2007/'
self.weights_path = '/home/caleml/partial_experiments/exp_20191028_1501_baseline_logits_pv_baseline/weights_100_001.h5'
# self.weights_path = '/home/caleml/partial_experiments/exp_20190725_1516_baseline/weights_100_020.h5'
self.gt_path = os.path.join(self.data_dir, 'Annotations', 'annotations_multilabel_trainval_partial_100_1.csv')
self.test_path = os.path.join(self.data_dir, 'Annotations', 'annotations_multilabel_test.csv')
# 1. load model
self.model = BaselineLogits('%s/partial_experiments/' % os.environ['HOME'], self.nb_classes, self.prop)
self.model.load_weights(self.weights_path)
# load prior
self.load_prior()
# load GT
self.gt = self.load_annotations(self.test_path)
def __init__(self, prior_path, weights_path, comb_method, prop, mode,
use_superclass, allow_marginalization):
self.batch_size = 16
self.nb_classes = 20
self.prop = prop
self.context_length = (self.prop * self.nb_classes / 100) - 1
self.mode = mode
class_info = load_ids()
self.id2name = {
int(info['id']): info['name']
for info in class_info.values()
}
self.data_dir = '/home/caleml/datasets/pascalvoc/VOCdevkit/VOC2007/'
self.weights_path = weights_path
self.gt_path = os.path.join(
self.data_dir, 'Annotations',
'annotations_multilabel_trainval_partial_100_1.csv')
# 1. load model
self.model = BaselineLogits(
'%s/partial_experiments/' % os.environ['HOME'], self.nb_classes,
self.prop)
self.model.load_weights(self.weights_path)
self.prior = self.load_prior(prior_path, comb_method, use_superclass,
allow_marginalization)
# load GT
self.gt = self.load_annotations(self.gt_path)
def __init__(self):
self.class_info = load_ids()
self.nb_classes = len(self.class_info)
assert self.nb_classes == 20
self.id2short = {
data['id']: data['short']
for data in self.class_info.values()
}
def __init__(self,
matrix_path,
nb_classes,
context_length,
comb_method='simple',
use_superclass=False,
alpha=None,
mode='train',
allow_marginalization=True):
'''
context_length: the expected length of the context (nb of classes in the context)
use_superclass means we use a prior computed on superclasses
superclasses are:
- animal
- indoor
- person
- vehicle
combination methods:
- simple -> add the logits of y_v and y_p
- alpha -> weight y_p with alpha first
'''
print(
'loading conditional prior with method %s, superclasses %s, alpha %s, mode %s (context length %s)'
% (comb_method, use_superclass, alpha, mode, context_length))
self.use_superclass = use_superclass
class_info = load_ids()
self.id2superclass = {
v['id']: v['superclass']
for v in class_info.values()
}
self.id2short = {v['id']: v['short'] for v in class_info.values()}
self.prior_matrix = self.load_matrix(matrix_path)
self.nb_classes = nb_classes
self.context_length = context_length
self.threshold = 0.5
self.combination_method = comb_method
self.mode = mode
assert self.mode in ['train', 'test']
if comb_method == 'alpha':
self.alpha = alpha
self.allow_marginalization = allow_marginalization
def create_prior(dataset_path):
'''
create the prior with superclass clustering
order of the keys:
matrix['a1']['i1_p0_v1'] gives the probability of the class animal given the presence of the classes indoor and vehicle and the absence of the classes person in the image
matrix['a1']['i0'] gives the marginalization over all complete-context values with i=0
'''
classes_order = ['animal', 'indoor', 'person', 'vehicle']
cooc_superclass = defaultdict(lambda: defaultdict(lambda: 0.0))
class_info = load_ids()
id2superclass = {
info['id']: info['superclass']
for info in class_info.values()
}
with open(dataset_path, 'r') as f_in:
for line in f_in:
parts = [int(x) for x in line.strip().split(',')]
# ones
ones_idx = [
i - 1 for i in range(len(parts[1:]) + 1) if parts[i] == 1
]
ones_superclasses = set([id2superclass[idx] for idx in ones_idx])
for superclass in ones_superclasses:
context = '_'.join([
'%s%s' % (sc[0], (1 if sc in ones_superclasses else 0))
for sc in classes_order if sc != superclass
])
cooc_superclass['%s1' % superclass[0]][context] += 1
# zeros
zeros_idx = [
i - 1 for i in range(len(parts[1:]) + 1) if parts[i] == -1
]
zeros_superclasses = set([id2superclass[idx] for idx in zeros_idx])
for superclass in zeros_superclasses:
context = '_'.join([
'%s%s' % (sc[0], (1 if sc in ones_superclasses else 0))
for sc in classes_order if sc != superclass
])
cooc_superclass['%s0' % superclass[0]][context] += 1
pprint(cooc_superclass)
# fill missing values, normalize and compute incomplete contexts
for letter in ALL_LETTERS:
full_contexts, partial_contexts = get_all_contexts(letter)
print(full_contexts)
print(partial_contexts)
# normalization: sum over the same context for each version 0 and 1: matrix['a1']['i1_p0_v1'] = matrix['a1']['i1_p0_v1'] / matrix['a1']['i1_p0_v1'] + matrix['a0']['i1_p0_v1']
for context in full_contexts:
print('doing context %s' % context)
one_context = cooc_superclass['%s1' % letter][context]
zero_context = cooc_superclass['%s0' % letter][context]
# both are None: 0.5 for each
if (one_context == 0.0) and (zero_context == 0.0):
cooc_superclass['%s1' % letter][context] = 0.5
cooc_superclass['%s0' % letter][context] = 0.5
# only one is None: epsilon
elif one_context == 0.0:
cooc_superclass['%s1' % letter][context] = cfg.EPSILON
elif zero_context == 0.0:
cooc_superclass['%s0' % letter][context] = cfg.EPSILON
# updated values if needed
one_context = cooc_superclass['%s1' % letter][context]
zero_context = cooc_superclass['%s0' % letter][context]
print('updated one context %s' % one_context)
print('updated zero context %s' % zero_context)
total = one_context + zero_context
cooc_superclass['%s0' % letter][context] /= total
cooc_superclass['%s1' % letter][context] /= total
# precompute the incomplete part
for partial in partial_contexts:
partial_parts = partial.split('_')
# we add all the values of the contexts that match the partial one
sum_one = sum([
val for ctxt, val in cooc_superclass['%s1' % letter].items()
if any([part in ctxt for part in partial_parts])
])
sum_zero = sum([
val for ctxt, val in cooc_superclass['%s0' % letter].items()
if any([part in ctxt for part in partial_parts])
])
cooc_superclass['%s0' %
letter][partial] = sum_zero / (sum_zero + sum_one)
cooc_superclass['%s1' %
letter][partial] = sum_one / (sum_zero + sum_one)
print('\nFinal matrix')
pprint(cooc_superclass)
# saving
save_file = filepath.replace('annotations_multilabel',
'prior_matrix4').replace('.csv', '.json')
with open(save_file, 'w+') as f_json:
json.dump(cooc_superclass, f_json)
def create_full_prior(dataset_path):
'''
create the full matrix (19 class context)
'''
cooc_matrix = defaultdict(lambda: defaultdict(lambda: 0.0))
class_info = load_ids()
id2letter = {info['id']: info['short'] for info in class_info.values()}
all_letters = list(id2letter.values())
# counting of the dataset
with open(dataset_path, 'r') as f_in:
for line in f_in:
parts = [int(x) for x in line.strip().split(',')]
ground_truths = parts[1:]
for i, gt in enumerate(ground_truths):
letter = id2letter[i]
value = (1 if gt == 1 else 0)
context = list()
# '_'.join(['%s%s' % (id2letter[cid], (1 if cid in ones_idx else 0)) for cid in range(len(ground_truths)) if cid != class_id])
for context_j, context_gt in enumerate(ground_truths):
if context_j == i:
continue
context_letter = id2letter[context_j]
context_value = (1 if context_gt == 1 else 0)
context.append('%s%s' % (context_letter, context_value))
full_context = '_'.join(context)
cooc_matrix['%s%s' % (letter, value)][full_context] += 1
# normalization and filling missing full contexts
for letter in all_letters:
context_letters = copy.copy(all_letters)
context_letters.remove(letter)
full_contexts = get_full_contexts(context_letters)
# normalization: sum over the same context for each version 0 and 1: matrix['ae1']['context1'] = matrix['ae1']['context1'] / matrix['a1']['context1'] + matrix['a0']['context1']
for context in full_contexts:
one_context = cooc_matrix['%s1' % letter][context]
zero_context = cooc_matrix['%s0' % letter][context]
# both are None: 0.5 for each
if (one_context == 0.0) and (zero_context == 0.0):
cooc_matrix['%s1' % letter][context] = 0.5
cooc_matrix['%s0' % letter][context] = 0.5
# missing one value: 0.4
elif one_context == 0.0:
cooc_matrix['%s1' % letter][context] = 0.4
elif zero_context == 0.0:
cooc_matrix['%s0' % letter][context] = 0.4
# updated values if needed
one_context = cooc_matrix['%s1' % letter][context]
zero_context = cooc_matrix['%s0' % letter][context]
# actual normalization
total = one_context + zero_context
cooc_matrix['%s0' % letter][context] /= total
cooc_matrix['%s1' % letter][context] /= total
# saving
save_file = filepath.replace('annotations_multilabel',
'prior_matrix_full0').replace(
'.csv', '.json')
with open(save_file, 'w+') as f_json:
json.dump(cooc_matrix, f_json)
print('saved matrix at %s' % save_file)