cudnn.benchmark = True
''' Initialize weights from checkpoint '''
if args.resume:
net, best_acc, start_acc = init_from_checkpoint(net)
''' Optimization '''
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
lr_scheduler = ReduceLROnPlateau(optimizer,
factor=0.5,
mode='max',
verbose=True)
''' Define passer '''
if not args.subset:
passer_train = Passer(net, trainloader, criterion, device)
else:
passer_train = Passer(net, subsettrainloader, criterion, device)
passer_test = Passer(net, testloader, criterion, device)
''' Define manipulator '''
manipulator = load_manipulator(args.permute_labels, args.binarize_labels)
''' Make intial pass before any training '''
loss_te, acc_te = passer_test.run()
save_checkpoint(checkpoint={
'net': net.state_dict(),
'acc': acc_te,
'epoch': 0
},
path='./checkpoint/' + ONAME + '/',
fname='ckpt_trial_' + str(args.trial) + '_epoch_0.t7')
''' Prepare criterion '''
if args.dataset in [
'cifar10', 'cifar10_gray', 'vgg_cifar10_adversarial', 'imagenet'
]:
criterion = nn.CrossEntropyLoss()
elif args.dataset in ['mnist', 'mnist_adverarial']:
criterion = F.nll_loss
''' Define label manipulator '''
manipulator = load_manipulator(args.permute_labels, args.binarize_labels)
''' Instead of building graph on the entire set of nodes, pick a subset '''
if args.select_nodes:
''' get activations '''
subsettestloader = loader(args.dataset + '_train',
batch_size=100,
sampling=args.binarize_labels)
passer = Passer(net, subsettestloader, criterion, device)
manipulator = load_manipulator(args.permute_labels, args.binarize_labels)
activs = passer.get_function()
activs = signal_concat(activs)
''' get correct and wrong predictions '''
gts, preds = passer.get_predictions(manipulator=manipulator)
labels = [int(x) for x in gts == preds]
''' compute discriminative nodes '''
nodes = np.concatenate(
get_discriminative_nodes(np.transpose(activs), labels, 0.1))
else:
nodes = np.arange(0, len(activs))
for epoch in args.epochs:
print('==> Loading checkpoint for epoch {}...'.format(epoch))
criterion = nn.CrossEntropyLoss()
''' Define label manipulator '''
manipulator = load_manipulator(args.permute_labels, args.binarize_labels)
for epoch in args.epochs:
print('==> Loading checkpoint for epoch {}...'.format(epoch))
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.net + '_' + args.dataset +
'/ckpt_trial_' + str(args.trial) + '_epoch_' +
str(epoch) + '.t7')
net.load_state_dict(checkpoint['net'])
''' Define passer and get activations '''
functloader = loader(args.dataset + '_test',
batch_size=100,
subset=list(range(0, 1000)))
passer = Passer(net, functloader, criterion, device)
passer_test = Passer(net, functloader, criterion, device)
passer_test.run(manipulator=manipulator)
activs = passer.get_function()
print('activs have shape {}'.format(signal_concat(activs).shape))
start = time.time()
if args.partition == 'hardcoded':
splits = signal_splitting(activs, args.split)
elif args.partition == 'dynamic':
splits = signal_partition(activs, n_part=args.split, binarize_t=0.5)
print('Returning from signal_partition in {} secs'.format(time.time() -
start))
elif args.partition == 'dynamic_from_structure':
sadj = structure_from_view(net.module,
print(net)
''' Prepare criterion '''
criterion = nn.CrossEntropyLoss()
''' Define label manipulator '''
manipulator = load_manipulator(args.permute_labels, args.binarize_labels)
for epoch in args.epochs:
print('==> Loading checkpoint for epoch {}...'.format(epoch))
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.net + '_' + args.dataset +
'/ckpt_trial_' + str(args.trial) + '_epoch_' +
str(epoch) + '.t7')
net.load_state_dict(checkpoint['net'])
''' Define passer and get activations '''
functloader = loader(args.dataset + '_test',
batch_size=100,
subset=list(range(0, 1000)))
passer = Passer(net, functloader, criterion, device)
passer_test = Passer(net, functloader, criterion, device)
passer_test.run(manipulator=manipulator)
activs = passer.get_function()
activs = signal_concat(activs)
adj = adjacency(activs)
print('The dimension of the adjacency matrix is {}'.format(adj.shape))
print('Adj mean {}, min {}, max {}'.format(np.mean(adj), np.min(adj),
np.max(adj)))
''' Write adjacency to binary. To use as DIPHA input for persistence homology '''
save_dipha(SAVE_DIR + 'adj_epc{}_trl{}.bin'.format(epoch, args.trial),
1 - adj)
cudnn.benchmark = True
''' Initialize weights from checkpoint'''
if args.resume:
net, best_acc, start_acc = init_from_checkpoint(net)
''' Optimization '''
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
lr_scheduler = ReduceLROnPlateau(optimizer,
factor=0.5,
mode='max',
verbose=True)
''' Define passer'''
if not args.subset:
passer_train = Passer(net, trainloader, criterion, device)
else:
passer_train = Passer(net, subsettrainloader, criterion, device)
passer_test = Passer(net, testloader, criterion, device)
''' Make intial pass before any training '''
loss_te, acc_te = passer_test.run()
save_checkpoint(checkpoint={
'net': net.state_dict(),
'acc': acc_te,
'epoch': 0
},
path='./checkpoint/' + ONAME,
fname='ckpt_trial_' + str(args.trial) + '_epoch_0.t7')
losses = []
manipulator = load_manipulator(args.permute_labels, args.binarize_labels)
''' Define partitions '''
print(25 * '-')
print('----Perform partition---')
print(25 * '-')
print('==> Loading checkpoint for epoch {}...'.format(args.epochs[-1]))
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.net + '_' + args.dataset +
'/ckpt_trial_' + str(args.trial) + '_epoch_' +
str(args.epochs[-1]) + '.t7')
net.load_state_dict(checkpoint['net'])
''' Define passer and get activations '''
functloader = loader(args.dataset + '_test',
batch_size=100,
subset=list(range(0, 1000)))
passer = Passer(net, functloader, criterion, device)
passer_test = Passer(net, functloader, criterion, device)
passer_test.run()
activs = passer.get_function()
print('activs have shape {}'.format(signal_concat(activs).shape))
start = time.time()
if args.partition == 'hardcoded':
splits = signal_splitting(activs, args.split)
elif args.partition == 'dynamic':
node_splits, _ = signal_partition(activs,
n_part=args.split,
binarize_t=args.thresholds[0])
print('Returning from signal_partition in {} secs'.format(time.time() -
start))
criterion = nn.CrossEntropyLoss()
elif args.dataset in ['mnist', 'mnist_adverarial']:
criterion = F.nll_loss
for epoch in args.epochs:
print('==> Loading checkpoint for epoch {}...'.format(epoch))
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.net + '_' + args.dataset +
'/ckpt_trial_' + str(args.trial) + '_epoch_' +
str(epoch) + '.t7')
net.load_state_dict(checkpoint['net'])
''' Define passer and get activations '''
functloader = loader(args.dataset + '_test',
batch_size=100,
subset=list(range(0, 1000)))
passer = Passer(net, functloader, criterion, device)
activs = passer.get_function()
''' If high number of nodes compute adjacency on layers and chunks'''
''' Treat all network at once or split it into chunks and treat each '''
if not args.split:
activs = signal_concat(activs)
adj = adjacency(activs)
for threshold in THRESHOLDS:
badj = binarize(np.copy(adj), threshold)
print('t={} s={}'.format(threshold, np.sum(badj)))
np.savetxt(SAVE_DIR + 'badj_epc{}_t{:1.2f}_trl{}.csv'.format(
epoch, threshold, args.trial),
badj,
fmt='%d',
delimiter=",")