def create_set_files(n_sets,
input_set_index,
input_set_name,
output_set_index,
output_set_name,
exclusion_files=None,
include_full=False,
set_range=100):
class_images = load_imagelist(set_index=input_set_index,
set_name=input_set_name,
exclusion_files=exclusion_files)
all_classes = np.unique(class_images.values())
n_images = len(class_images)
revalations = [str(r) for r in xrange(0, set_range, 10)]
if include_full: revalations += ['full']
n_revalations = len(revalations)
image_to_index = {}
idx = 0
for c in all_classes:
image_list = [k for k in class_images.keys() if class_images[k] == c]
for image in np.random.permutation(image_list):
image_to_index[image] = idx
idx += 1
for i_set in xrange(n_sets):
set_fn = get_imageset_filename(set_index=output_set_index + i_set,
set_name=output_set_name)
print 'Writing %s' % set_fn
with open(set_fn, 'wt') as fid:
image_list = np.random.permutation(
[k for k in class_images.keys()])
for image in image_list:
rev = revalations[(image_to_index[image] + i_set) %
n_revalations]
fid.write('%s/%s\t%d\n' % (rev, image, class_images[image]))
def load_im2path(self, experiment_ids):
self.im2path = {}
for experiment_id in experiment_ids:
set_index, set_name = util.get_experiment_imageset(experiment_id)
imageset_filename = util.get_imageset_filename(set_index, set_name)
im_names = label_results(imageset_filename)
im_root = util.get_input_image_root(set_index, set_name)
for k in im_names.iterkeys():
self.im2path[k] = im_root
def convert_raw_imageset(set_index, set_name):
fileid_filename = get_imageset_filename(set_index=set_index,
set_name=set_name)
raw_filename = get_imageset_filename_raw(set_index=set_index,
set_name=set_name)
if os.path.isfile(fileid_filename):
raise RuntimeError('File exists: %s' % fileid_filename)
convert_raw_to_fileid(raw_filename, fileid_filename)
print 'Written to %s' % fileid_filename
def write_set_file(set_index, set_name, dst_filename):
# Load set and convert to CSV
src_fn = get_imageset_filename(set_index=set_index, set_name=set_name)
data = open(src_fn, 'rt').read().splitlines()
class_identifiers = ['nonanimal', 'animal'
] # in src_fn, 0 is the distractor; 1 is the tagetg
with open(dst_filename, 'wt') as fid:
fid.write('img,cat\n')
for line in data:
ll = line.split('\t')
fn = ll[0]
class_identifier = class_identifiers[int(ll[1])]
fid.write('%s,%s\n' % (fn, class_identifier))
def eval_by_classes(self, class_idcs, class_names):
self.model_comp_by_class = {}
for class_idx,class_name in zip(class_idcs, class_names):
imageset_filename = util.get_imageset_filename(class_idx, 'set')
class_images = label_results(imageset_filename)
comp = {'animal': [], 'nonanimal': []}
rts = {'animal': [], 'nonanimal': []}
for im,lbl in class_images.iteritems():
if im in self.im2key:
comp[lbl] += [self.model_comp['animal']['human_acc'][0, self.im2key[im]]]
rts[lbl] += [self.data_by_img[im[:-4]]['rt']]
self.model_comp_by_class[class_idx] = comp
print 'Class %s: Found %d animal, %d nonanimal trials' % (class_name, len(comp['animal']), len(comp['nonanimal']))
print 'Class %s percent correct: %.1f (animal=%.1f, nonanimal=%.1f)' % (class_name, np.mean(comp['animal']+comp['nonanimal']), np.mean(comp['animal']), np.mean(comp['nonanimal']))
print 'Class %s RT: %.1fms (animal=%.1fms, nonanimal=%.1fms)' % (class_name, np.mean(rts['animal'] + rts['nonanimal']), np.mean(rts['animal']),np.mean(rts['nonanimal']))
def load_ground_truth(self, set_index, set_name):
imageset_filename = util.get_imageset_filename(set_index, set_name)
# Generate ground truth key
self.im2lab.update(label_results(imageset_filename))
self.loaded_sets += [(set_index, set_name)]