def main():
args = parse_args()
torch.cuda.set_device(args.gpu)
if args.dataset == 'domainnet':
num_domain = 6
num_classes = 345
elif args.dataset == 'officehome':
num_domain = 4
num_classes = 65
print(args.dataset, num_classes, num_domain)
_, trg_sup_val = get_dataset(dataset=args.dataset, dataset_root=args.data_root, domain=args.trg_domain,
ssl=False)
trg_num = domain_dict[args.dataset][args.trg_domain]
model = get_model(args.model_name, in_features=num_classes, num_classes=num_classes,
num_domains=num_domain, pretrained=True)
model.load_state_dict(torch.load(args.model_path)['model'])
model = model.cuda(args.gpu)
_, acc = test(args, model, trg_sup_val, domain_dict[args.dataset][args.trg_domain])
print('acc: %0.3f'%(acc))
def test_model(test_path,
model_name,
trained_model,
save_path,
traineddataset,
testdataset,
opt,
lr,
pretrain,
num_classes,
test_split=False):
state = defaultdict()
test_loader = defaultdict()
test_transform = trn.Compose([
trn.ToTensor(),
trn.Normalize([0.566, 0.496, 0.469], [0.266, 0.256, 0.258])
])
img_pixels = (224, 224)
# Load test data
test_set_X, test_set_y, _ = data_utils.process_data(test_path)
test_loader = data_utils.make_dataloader_iter(
test_set_X,
test_set_y,
img_size=img_pixels,
batch_size=10,
transform_test=test_transform)
# Load model
if model_name == 'alexnet':
net = Generalmodels.alexnet(num_classes,
pretrain,
trained_model,
if_test=True)
elif model_name == 'VGG':
net = Generalmodels.VGG16(num_classes,
pretrain,
trained_model,
if_test=True)
elif model_name == 'densenet121':
net = Generalmodels.densenet121(num_classes,
pretrain,
trained_model,
if_test=True)
elif model_name == 'resnet50':
net = Generalmodels.resnet50(num_classes,
pretrain,
trained_model,
if_test=True)
device = torch.device("cuda")
net.load_state_dict(torch.load(trained_model))
net = net.to(device)
# If test general performance
if not test_split:
if not os.path.exists(save_path):
os.makedirs(save_path)
test(net, test_loader, state)
with open(
os.path.join(
save_path, '{}_{}_{}_{}_{}_results.csv'.format(
traineddataset, model_name, opt, lr, testdataset)),
'a') as f:
f.write('%s,%0.6f\n' % (
'(test)',
state['test_accuracy'],
))
# Test performance per age
if test_split:
new_path = os.path.join(save_path, 'split')
if not os.path.exists(new_path):
os.makedirs(new_path)
if testdataset == 'FineTuneData':
testdataset = 'testing'
gt,tp,mae,prediction,classpredicts,regrepredicts,labels=\
test_range(net,test_loader,state,num_classes)
# write predictions
f = open(
os.path.join(
new_path, '{}_{}_{}_{}_{}_allpredictions.csv'.format(
traineddataset, model_name, opt, lr, testdataset)), 'w')
f.write('%s,%s,%s\n' % (
'label',
'classpred',
'regreepred',
))
for i in range(len(classpredicts)):
f.write('%d,%d,%0.2f\n' %
(labels[i], classpredicts[i], regrepredicts[i]))
f.close()
# Write performance
f = open(
os.path.join(
new_path,
'{}_{}_{}_{}_{}_results.csv'.format(traineddataset, model_name,
opt, lr, testdataset)),
'w')
f.write('%s,%s,%s,%s,%s\n' %
('age', 'number', 'accuracy', 'mae', 'perceived'))
for i in range(len(gt)):
f.write('%d,%d,%0.6f,%0.3f,%0.3f\n' %
(i, gt[i], tp[i] / gt[i], mae[i] / gt[i],
prediction[i] / gt[i]))
f.close()