def train(args):
kwargs = {'num_workers': 4, 'pin_memory': True} if args.cuda else {}
train_loader = data_.DataLoader(datasets.MNIST(
'./data', train=True, download=True, transform=transforms.ToTensor()),
batch_size=args.batch_size,
shuffle=True,
**kwargs)
test_loader = data_.DataLoader(datasets.MNIST(
'./data', train=False, transform=transforms.ToTensor()),
batch_size=args.test_batch_size,
shuffle=True,
**kwargs)
if args.binary == 'connect':
net = model.BinaryConnect(args.in_features, args.out_features)
else:
net = model.BinaryNet(args.in_features, args.out_features)
# net = nn.DataParallel(net)
print(net)
if args.cuda:
net.cuda()
optimizer = optim.Adam(net.parameters(), lr=args.lr)
creterion = nn.NLLLoss() ##negative log likelihood loss
for epoch in range(1, args.epochs + 1):
train_epoch(epoch, net, creterion, optimizer, train_loader, args)
test_epoch(net, creterion, test_loader, args)
def train(args):
kwargs = {'num_workers': 2, 'pin_memory': True} if args.cuda else {}
train_loader = data.DataLoader(
datasets.MNIST('./data', train=True, download=True,
transform=transforms.ToTensor()),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = data.DataLoader(
datasets.MNIST('./data', train=False,
transform=transforms.ToTensor()),
batch_size=args.test_batch_size, shuffle=True, **kwargs)
if args.binary == 'connect':
net = model.BinaryConnect(args.in_features, args.out_features)
elif args.binary == 'bnn':
net = model.BinaryNet(args.in_features, args.out_features)
elif args.binary == 'nn':
net = model.Net(args.in_features, args.out_features)
# net = nn.DataParallel(net)
print(net)
if args.cuda:
net.cuda()
if args.optimizer == 'Adam':
optimizer = optim.Adam(net.parameters(), lr=args.lr)
elif args.optimizer == 'LBFGS':
if args.line_search_fn is None:
optimizer = optim.LBFGS(net.parameters(), lr=args.lr)
elif args.line_search_fn == 'strong_wolfe':
optimizer = optim.LBFGS(net.parameters(), lr=args.lr, line_search_fn='strong_wolfe')
criterion = nn.NLLLoss()
for epoch in range(1, args.epochs+1):
train_epoch(epoch, net, criterion, optimizer, train_loader, args)
test_epoch(net, criterion, test_loader, args)
def train(args):
kwargs = {'num_workers': 2, 'pin_memory': True} if args.cuda else {}
train_loader = data.DataLoader(datasets.MNIST(
'./data', train=True, download=True, transform=transforms.ToTensor()),
batch_size=args.batch_size,
shuffle=True,
**kwargs)
test_loader = data.DataLoader(datasets.MNIST(
'./data', train=False, transform=transforms.ToTensor()),
batch_size=args.test_batch_size,
shuffle=True,
**kwargs)
net = model.BinaryConnect(args.topology, args.batch_norm)
# net = nn.DataParallel(net)
print(net)
output_folder = vars(args).get('output_folder')
if output_folder is not None:
assert not os.path.exists(output_folder)
os.makedirs(output_folder, exist_ok=False)
with open(os.path.join(output_folder, 'command.txt'), 'w') as f:
f.write(' '.join(sys.argv))
if args.cuda:
net.cuda()
optimizer = optim.Adam(net.parameters(), lr=args.lr)
creterion = nn.NLLLoss()
train_accs = []
test_accs = []
assert args.epochs > 0
for epoch in range(1, args.epochs + 1):
train_acc = train_epoch(epoch, net, creterion, optimizer, train_loader,
args)
test_acc = test_epoch(net, creterion, test_loader, args)
train_accs.append(train_acc)
test_accs.append(test_acc)
if output_folder is not None:
state_dict = {
key: value.cpu()
for key, value in net.state_dict().items()
}
torch.save(state_dict,
os.path.join(output_folder, '%s.pth' % (epoch, )))
if output_folder is not None:
np.save(os.path.join(output_folder, 'accuracy_train.npy'),
np.array(train_accs))
np.save(os.path.join(output_folder, 'accuracy_test.npy'),
np.array(test_accs))
def train(args):
kwargs = {'num_workers': 2, 'pin_memory': True} if args.cuda else {}
train_loader = data.DataLoader(
datasets.MNIST('./data', train=True, download=True,
transform=transforms.ToTensor()),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = data.DataLoader(
datasets.MNIST('./data', train=False,
transform=transforms.ToTensor()),
batch_size=args.test_batch_size, shuffle=True, **kwargs)
if args.binary == 'connect':
net = model.BinaryConnect(args.in_features, args.out_features)
else:
net = model.BinaryNet(args.in_features, args.out_features)
# net = nn.DataParallel(net)
print(net)
if args.cuda:
net.cuda()
optimizer = optim.Adam(net.parameters(), lr=args.lr)
decay = (0.000003 / args.lr) ** (1. / args.epochs)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, decay)
creterion = nn.MultiMarginLoss()
input_layer = next(net.net.children())
weights = open('data/weights.csv', 'w')
stats = open('data/stats.csv', 'w')
test_epoch(net, creterion, test_loader, args)
for epoch in range(1, args.epochs+1):
scheduler.step(epoch - 1)
tr_loss, tr_acc = train_epoch(epoch, net, creterion, optimizer, train_loader, args)
tt_loss, tt_acc = test_epoch(net, creterion, test_loader, args)
items = input_layer.weight.cpu().data.numpy().reshape(-1, )
items = np.clip(items, -1, 1)
hist, _ = np.histogram(items, bins=1000, range=(-1., 1.))
weights.write(','.join(map(str, hist)))
weights.write('\n')
weights.flush()
stats.write('{},{},{},{},{},{}\n'.format(epoch, time.clock(), tr_loss, tr_acc, tt_loss, tt_acc))
stats.flush()
weights.close()
stats.close()