def generate_name_of_result_folder(args):
global_opts = get_global_opts()
results_path = os.path.join(global_opts['result_path'], 'cluster-training')
if 'vis' == args['startnet']:
startnetstr = 'map1'
elif 'cs' == args['startnet']:
startnetstr = 'map0'
elif 'pola' == args['startnet']:
startnetstr = 'map2'
else:
startnetstr = 'other'
cluster_str = 'features%d' % (args['max_features_per_image'])
if args['feature_hinge_loss_weight'] == 0:
result_folder = 'cluster-%s-%s-cn%d-ci%d-vi%d' % (
args['corr_set'], startnetstr, args['n_clusters'],
args['cluster_interval'], args['val_interval'])
else:
result_folder = 'cluster-%s-%s-cn%d-ci%d-vi%d-ws%.5f-wf%.5f-%s-valm' % (
args['corr_set'], startnetstr, args['n_clusters'],
args['cluster_interval'], args['val_interval'],
args['seg_loss_weight'], args['feature_hinge_loss_weight'],
cluster_str)
return os.path.join(results_path, result_folder), os.path.join(
global_opts['result_path'], result_folder)
def get_path_of_startnet(args):
global_opts = get_global_opts()
if args['startnet'] == 'vis':
return os.path.join(global_opts['result_path'], 'base-networks',
'pspnet101_cs_vis.pth')
elif args['startnet'] == 'cs':
return os.path.join(global_opts['result_path'], 'base-networks',
'pspnet101_cityscapes.pth')
def get_path_of_startnet(args):
global_opts = get_global_opts()
if args['include_vistas']:
if args['corr_set'] == 'rc':
return os.path.join(global_opts['result_path'], 'base-networks',
'pspnet101_cs_vis_rc.pth')
elif args['corr_set'] == 'cmu':
return os.path.join(global_opts['result_path'], 'base-networks',
'pspnet101_cs_vis_cmu.pth')
else:
if args['corr_set'] == 'rc':
return os.path.join(global_opts['result_path'], 'base-networks',
'pspnet101_cs_rc.pth')
elif args['corr_set'] == 'cmu':
return os.path.join(global_opts['result_path'], 'base-networks',
'pspnet101_cs_cmu.pth')
def generate_name_of_result_folder(args):
global_opts = get_global_opts()
results_path = os.path.join(global_opts['result_path'], 'corr-training')
if args['classes_to_ignore'] is None:
ignore_classes = 0
else:
ignore_classes = 1
if (args['corr_loss_type'] == 'class') or (args['corr_loss_type'] == 'KL'):
args['feat_dist_threshold_match'] = 0
args['feat_dist_threshold_nomatch'] = 0
result_folder = 'corr-%s-map%d-%s-w%.5f-%.2f-%.2f-%d-%d-%d-seg-w%.5f-%.10flr' % (
args['corr_set'], args['include_vistas'], args['corr_loss_type'],
args['corr_loss_weight'], args['feat_dist_threshold_match'],
args['feat_dist_threshold_nomatch'],
args['n_iterations_before_corr_loss'], ignore_classes,
args['remove_same_class'], args['seg_loss_weight'], args['lr'])
return os.path.join(results_path, result_folder)
]
network_folder = args['network_file'][:slash_inds[-1]]
network_file = args['network_file']
# folder should have same name as for trained network
if args['validation_metric'] == 'miou':
save_folder = os.path.join(network_folder, args['save_folder_name'])
elif args['validation_metric'] == 'acc':
save_folder = os.path.join(network_folder,
args['save_folder_name'] + '_acc')
segment_images_in_folder(network_file, args['img_path'], save_folder, args)
if __name__ == '__main__':
global_opts = get_global_opts()
args = {
'use_gpu':
True,
# 'miou' (miou over classes present in validation set), 'acc'
'validation_metric':
'miou',
'img_set':
'', # ox-vis, cmu-vis, wilddash , ox, cmu, cityscapes overwriter img_path, img_ext and save_folder_name. Set to empty string to ignore
'img_path':
'/semseg/testimg/val',
# 'img_ext': '.png',
'img_ext':
'.jpg',
'save_folder_name':