log = []
for config.train['split_random_seed'] in tqdm(range(args.start,args.num_models) , desc='bagging' , leave=False):
if args.model_name == 'tede':
config.net['name'] = 'tede_resnet18'
config.train['optimizer'] = 'SGD'
config.train['batch_size'] = 64
#config.train['lr_bounds'] = [ 0 , 40 , 60 ,72 , 80 ]
#config.train['lrs'] = [ 1e-1 , 1e-2 , 1e-3 , 1e-4 ]
config.train['lr_bounds'] = [ 0 , 40 , 48 , 52 ]
config.train['lrs'] = [ 1e-1 , 1e-2 , 1e-3 ]
config.train['add_attributes'] = True
config.train['add_class_wordsembeddings'] = True
config.parse_config()
activation_fn = partial( nn.LeakyReLU )
config.net['semantic_mlp_kwargs']['last_activation_fn'] = activation_fn
config.net['visual_mlp_kwargs']['activation_fn'] = activation_fn
feature_layer_dim = 384
config.net['visual_mlp_kwargs']['out_channels'] = feature_layer_dim
config.net['semantic_mlp_kwargs']['out_channels'] = feature_layer_dim
result = main( config )
log.append( [ str(config.train['split_random_seed']) , result['log_path'] , result['log_path'], str(1 - result['non_zero']) , str(1 - result['zero']) ] )
elif args.model_name == 'gcn':
config.train['save_metric'] = 'err'
config.train['optimizer'] = 'SGD'
config.train['batch_size'] = 64
config.train['lr_bounds'] = [ 0 , 40 , 60 ,72 , 80 ]
config.train['lrs'] = [ 3e-1 , 3e-2 , 3e-3 , 1e-4 ]
def main(args):
fc = []
with open(args.model_info) as fp:
first = True
a = fp.readlines()
a = [i for i in map(lambda x: x.strip().split(' '), a)]
#print(a)
if args.top is not None:
a.sort(key=lambda x: float(x[-1]), reverse=False)
a = a[:args.top]
print([i for i in map(lambda x: float(x[-1]), a)])
for l in tqdm(a, leave=False):
if args.model_name == 'gcn':
import test_gcn
random_seed, resume_non_zero_net, resume, nonzero_acc, zero_acc = l[
0], l[1], float(l[2]), float(l[3])
arg = type('', (), {})()
arg.resume = resume
arg.input_list = '../data/test.txt'
arg.resume_non_zero_net = resume_non_zero_net
arg.resume_coarse_net = None
arg.zero_acc = zero_acc
arg.nonzero_acc = nonzero_acc
arg.x_tag = 'feat'
arg.output_list = '../data/output.list'
arg.batch_size = 256
arg.has_label = False
if not args.use_non_zero_net:
arg.resume_non_zero_net = None
import train_config as config
config.train['split_random_seed'] = random_seed
config.train['graph_similarity'] = 'custom'
config.parse_config()
temp_fc = test_gcn.main(arg, config)['fc']
elif args.model_name == 'tede':
import test_tede
arg = type('', (), {})()
random_seed, resume, _, nonzero_acc, zero_acc = l[0], l[1], l[
2], float(l[3]), float(l[4])
arg.resume = resume
arg.input_list = '../data/test.txt'
arg.batch_size = 256
arg.has_label = False
arg.output_list = '../data/output.list'
arg.resume_epoch = args.resume_epoch
import train_config as config
config.train['split_random_seed'] = random_seed
config.net['name'] = 'tede_resnet18'
config.parse_config()
feature_layer_dim = 384
config.net['visual_mlp_kwargs'][
'out_channels'] = feature_layer_dim
config.net['semantic_mlp_kwargs'][
'out_channels'] = feature_layer_dim
temp_fc = test_tede.main(arg, config)['fc']
os.system('rm {}'.format(arg.output_list))
if first:
fc = temp_fc
first = False
else:
for i in range(len(fc)):
for k in fc[i]:
fc[i][k] += temp_fc[i][k]
y = []
for fc_dict in fc:
y.append(max(fc_dict, key=fc_dict.get))
test_list = open('../data/test.txt').read().strip().split('\n')
assert len(y) == len(test_list)
output_list = []
for label, line in zip(y, test_list):
output_list.append(
line.split('\t')[0].strip().split('/')[-1] + '\t' + label)
with open(args.output_list, 'w') as fp:
fp.write('\n'.join(output_list) + '\n')
fp.flush()