N = int(sys.argv[1])
poisoned_models = glob.glob(
'./poisoned_models/Triggers_01_10/*.pt')[:1000] # Choose 1000 models
clean_models = glob.glob(
'./clean_models/train/*.pt')[:1000] # Choose 1000 models
models = clean_models + poisoned_models
labels = np.concatenate(
[np.zeros((len(clean_models), )),
np.ones((len(poisoned_models), ))])
# split models into train and test set fro ULP training
train_models, test_models, train_labels, test_labels = train_test_split(
models, labels, test_size=.2, random_state=10)
cnn = resnet18_mod(num_classes=nofclasses)
if use_cuda:
device = torch.device('cuda')
cnn.cuda()
else:
device = torch.device('cpu')
X = torch.rand((N, 3, 32, 32), requires_grad=True, device='cuda')
X.data *= 255.
W = torch.randn((200, 2), requires_grad=True, device='cuda')
b = torch.zeros((2, ), requires_grad=True, device='cuda')
optimizerX = optim.SGD(params=[X], lr=1e+2)
optimizerWb = optim.Adam(params=[W, b], lr=1e-4)
os.makedirs(saveDirmeta)
crossentropy=torch.nn.CrossEntropyLoss()
count=0
clean_models = []
partition = int(sys.argv[1])
gpu="0"
runs=0
while runs<50:
n = partition*50+runs
val_temp=0
train_accuracy=0
logging.info('Training model %d'%(n))
cnn = resnet18_mod(num_classes=200)
logging.info(cnn)
# Compute number of parameters
s = sum(np.prod(list(p.size())) for p in cnn.parameters())
print ('Number of params: %d' % s)
if use_cuda:
device='cuda:'+gpu
cnn.to(device)
else:
device=torch.device('cpu')
optimizer = optim.Adam(params=cnn.parameters(), lr=0.001)
for epoch in range(nof_epochs):
cnn.train()
# adjust_learning_rate(optimizer, epoch)