def test_swag_cov(self, **kwargs):
model = torch.nn.Linear(300, 3, bias=True)
swag_model = SWAG(torch.nn.Linear, in_features=300, out_features=3, bias=True, subspace_type = 'covariance',
subspace_kwargs = {'max_rank':140})
optimizer = torch.optim.SGD(model.parameters(), lr = 1e-3)
# construct swag model via training
torch.manual_seed(0)
for _ in range(101):
model.zero_grad()
input = torch.randn(100, 300)
output = model(input)
loss = ((torch.randn(100, 3) - output)**2.0).sum()
loss.backward()
optimizer.step()
swag_model.collect_model(model)
# check to ensure parameters have the correct sizes
mean, var = swag_model._get_mean_and_variance()
cov_mat_sqrt = swag_model.subspace.get_space()
true_cov_mat = cov_mat_sqrt.t().matmul(cov_mat_sqrt) + torch.diag(var)
test_cutoff = chi2(df = mean.numel()).ppf(0.95) #95% quantile of p dimensional chi-square distribution
for scale in [0.01, 0.1, 0.5, 1.0, 2.0, 5.0]:
scaled_cov_mat = true_cov_mat * scale
scaled_cov_inv = torch.inverse(scaled_cov_mat)
# now test to ensure that sampling has the correct covariance matrix probabilistically
all_qforms = []
for _ in range(3000):
swag_model.sample(scale=scale)
curr_pars = []
for (module, name, _) in swag_model.base_params:
curr_pars.append(getattr(module, name))
dev = flatten(curr_pars) - mean
#(x - mu)sigma^{-1}(x - mu)
qform = dev.matmul(scaled_cov_inv).matmul(dev)
all_qforms.append(qform.item())
samples_in_cr = (np.array(all_qforms) < test_cutoff).sum()
print(samples_in_cr)
#between 94 and 96% of the samples should fall within the threshold
#this should be very loose
self.assertTrue(0.94*3000 <= samples_in_cr <= 0.96*3000)
def RealNVPTabularSWAG(dim_in,
coupling_layers,
k,
nperlayer=1,
subspace='covariance',
max_num_models=10):
swag_model = SWAG(RealNVPTabular,
subspace_type=subspace,
subspace_kwargs={'max_rank': max_num_models},
num_coupling_layers=coupling_layers,
in_dim=dim_in,
hidden_dim=k,
num_layers=1,
dropout=True)
return swag_model
print('Using model %s' % args.model)
model_cfg = getattr(models, args.model)
print('Loading dataset %s from %s' % (args.dataset, args.data_path))
loaders, num_classes = data.loaders(args.dataset,
args.data_path,
args.batch_size,
args.num_workers,
model_cfg.transform_train,
model_cfg.transform_test,
use_validation=not args.use_test)
print('Preparing model')
swag_model = SWAG(model_cfg.base,
no_cov_mat=not args.cov_mat,
max_num_models=args.swag_rank,
*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
swag_model.to(args.device)
ckpt = torch.load(args.checkpoint)
criterion = losses.cross_entropy
fractions = np.logspace(-np.log10(0.005 * len(loaders['train'].dataset)), 0.0,
args.N)
swa_accuracies = np.zeros(args.N)
swa_nlls = np.zeros(args.N)
swag_accuracies = np.zeros(args.N)
swag_nlls = np.zeros(args.N)
print(*model_cfg.args)
model = model_cfg.base(*model_cfg.args,
num_classes=num_classes,
input_shape=input_shape,
**model_cfg.kwargs)
#model.to(args.device)
if args.cov_mat:
args.no_cov_mat = False
else:
args.no_cov_mat = True
if args.swa:
print("SWAG training")
swag_model = SWAG(model,
no_cov_mat=args.no_cov_mat,
max_num_models=args.max_num_models,
*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
#swag_model.to(args.device)
else:
print("SGD training")
def schedule(epoch):
t = (epoch) / (args.swa_start if args.swa else args.epochs)
lr_ratio = args.swa_lr / args.lr_init if args.swa else 0.01
if t <= 0.5:
factor = 1.0
elif t <= 0.9:
factor = 1.0 - (1.0 - lr_ratio) * (t - 0.5) / 0.4
else:
print(*model_cfg.args)
model = model_cfg.base(*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
wandb.watch(model)
model.to(args.device)
if args.cov_mat:
args.no_cov_mat = False
else:
args.no_cov_mat = True
if args.swa:
print("SWAG training")
swag_model = SWAG(model_cfg.base,
no_cov_mat=args.no_cov_mat,
max_num_models=args.max_num_models,
*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
swag_model.to(args.device)
else:
print("SGD training")
def schedule(epoch):
t = (epoch) / (args.swa_start if args.swa else args.epochs)
lr_ratio = args.swa_lr / args.lr_init if args.swa else 0.01
if t <= 0.5:
factor = 1.0
elif t <= 0.9:
factor = 1.0 - (1.0 - lr_ratio) * (t - 0.5) / 0.4
else:
'max_rank': 20,
'pca_rank': 'mle',
}
}, {
'name': 'freq_dir',
'kwargs': {
'max_rank': 20,
}
}]
for item in subspaces:
name, kwargs = item['name'], item['kwargs']
print('Now running %s %r' % (name, kwargs))
model = model_generator()
small_swag_model = SWAG(base=model_generator,
subspace_type=name,
subspace_kwargs=kwargs)
optimizer = torch.optim.SGD(model.parameters(),
lr=1e-3,
momentum=0.9,
weight_decay=1e-4)
loader = generate_dataloaders(N=10)
state_dict = None
for epoch in range(num_epochs):
model.train()
for x, y in loader:
model.zero_grad()
label_arr = np.load(args.label_arr)
print("Corruption:",
(loaders['train'].dataset.targets != label_arr).mean())
loaders['train'].dataset.targets = label_arr
print('Preparing model')
model = model_cfg.base(*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
model.to(args.device)
print("Model has {} parameters".format(
sum([p.numel() for p in model.parameters()])))
swag_model = SWAG(model_cfg.base,
args.subspace, {'max_rank': args.max_num_models},
*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
swag_model.to(args.device)
columns = ['swag', 'sample', 'te_loss', 'te_acc', 'ens_loss', 'ens_acc']
if args.warm_start:
print('estimating initial PI using K-Means')
kmeans_input = []
for ckpt_i, ckpt in enumerate(args.swag_ckpts):
#print("Checkpoint {}".format(ckpt))
checkpoint = torch.load(ckpt)
swag_model.subspace.rank = torch.tensor(0)
swag_model.load_state_dict(checkpoint['state_dict'])
mean, variance = swag_model._get_mean_and_variance()
torch.cuda.manual_seed(args.seed)
print('Using model %s' % args.model)
model_class = getattr(torchvision.models, args.model)
print('Loading ImageNet from %s' % (args.data_path))
loaders, num_classes = data.loaders(
args.data_path,
args.batch_size,
args.num_workers,
)
print('Preparing model')
swag_model = SWAG(model_class,
no_cov_mat=not args.cov_mat,
loading=True,
max_num_models=20,
num_classes=num_classes)
swag_model.to(args.device)
criterion = losses.cross_entropy
print('Loading checkpoint %s' % args.ckpt)
checkpoint = torch.load(args.ckpt)
swag_model.load_state_dict(checkpoint['state_dict'])
print('SWA')
swag_model.sample(0.0)
print('SWA BN update')
utils.bn_update(loaders['train'], swag_model, verbose=True, subset=0.1)
print('SWA EVAL')
def test_swag_cov(self, **kwargs):
model = torch.nn.Linear(300, 3, bias=True)
swag_model = SWAG(
torch.nn.Linear,
in_features=300,
out_features=3,
bias=True,
no_cov_mat=False,
max_num_models=100,
loading=False,
)
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
# construct swag model via training
torch.manual_seed(0)
for _ in range(101):
model.zero_grad()
input = torch.randn(100, 300)
output = model(input)
loss = ((torch.randn(100, 3) - output) ** 2.0).sum()
loss.backward()
optimizer.step()
swag_model.collect_model(model)
# check to ensure parameters have the correct sizes
mean_list = []
sq_mean_list = []
cov_mat_sqrt_list = []
for (module, name), param in zip(swag_model.params, model.parameters()):
mean = module.__getattr__("%s_mean" % name)
sq_mean = module.__getattr__("%s_sq_mean" % name)
cov_mat_sqrt = module.__getattr__("%s_cov_mat_sqrt" % name)
self.assertEqual(param.size(), mean.size())
self.assertEqual(param.size(), sq_mean.size())
self.assertEqual(
[swag_model.max_num_models, param.numel()], list(cov_mat_sqrt.size())
)
mean_list.append(mean)
sq_mean_list.append(sq_mean)
cov_mat_sqrt_list.append(cov_mat_sqrt)
mean = flatten(mean_list).cuda()
sq_mean = flatten(sq_mean_list).cuda()
cov_mat_sqrt = torch.cat(cov_mat_sqrt_list, dim=1).cuda()
true_cov_mat = (
1.0 / (swag_model.max_num_models - 1)
) * cov_mat_sqrt.t().matmul(cov_mat_sqrt) + torch.diag(sq_mean - mean ** 2)
test_cutoff = chi2(df=mean.numel()).ppf(
0.95
) # 95% quantile of p dimensional chi-square distribution
for scale in [0.01, 0.1, 0.5, 1.0, 2.0, 5.0]:
scaled_cov_mat = true_cov_mat * scale
scaled_cov_inv = torch.inverse(scaled_cov_mat)
# now test to ensure that sampling has the correct covariance matrix probabilistically
all_qforms = []
for _ in range(2000):
swag_model.sample(scale=scale, cov=True)
curr_pars = []
for (module, name) in swag_model.params:
curr_pars.append(getattr(module, name))
dev = flatten(curr_pars) - mean
# (x - mu)sigma^{-1}(x - mu)
qform = dev.matmul(scaled_cov_inv).matmul(dev)
all_qforms.append(qform.item())
samples_in_cr = (np.array(all_qforms) < test_cutoff).sum()
print(samples_in_cr)
# between 94 and 96% of the samples should fall within the threshold
# this should be very loose
self.assertTrue(1880 <= samples_in_cr <= 1920)
criterion = partial(criterion, weight=class_weights)
if args.resume is not None:
print('Resume training from %s' % args.resume)
checkpoint = torch.load(args.resume)
start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
del checkpoint
if args.swa:
print('SWAG training')
swag_model = SWAG(model_cfg.base,
no_cov_mat=False,
max_num_models=20,
*model_cfg.args,
num_classes=num_classes,
use_aleatoric=args.loss == 'aleatoric',
**model_cfg.kwargs)
swag_model.to(args.device)
else:
print('SGD training')
if args.swa and args.swa_resume is not None:
checkpoint = torch.load(args.swa_resume)
swag_model = SWAG(model_cfg.base,
no_cov_mat=False,
max_num_models=20,
*model_cfg.args,
num_classes=num_classes,
use_aleatoric=args.loss == 'aleatoric',
target_transform=model_cfg.target_transform,
)
# criterion = nn.NLLLoss(weight=camvid.class_weight[:-1].cuda(), reduction='none').cuda()
if args.loss == "cross_entropy":
criterion = losses.seg_cross_entropy
else:
criterion = losses.seg_ale_cross_entropy
# construct and load model
if args.swa_resume is not None:
checkpoint = torch.load(args.swa_resume)
model = SWAG(
model_cfg.base,
no_cov_mat=False,
max_num_models=20,
num_classes=num_classes,
use_aleatoric=args.loss == "aleatoric",
)
model.cuda()
model.load_state_dict(checkpoint["state_dict"])
model.sample(0.0)
bn_update(loaders["fine_tune"], model)
else:
model = model_cfg.base(num_classes=num_classes,
use_aleatoric=args.loss == "aleatoric").cuda()
checkpoint = torch.load(args.resume)
start_epoch = checkpoint["epoch"]
print(start_epoch)
model.load_state_dict(checkpoint["state_dict"])
def __init__(self,
base,
epochs,
criterion,
batch_size=50,
lr_init=1e-2,
momentum=0.9,
wd=1e-4,
swag_lr=1e-3,
swag_freq=1,
swag_start=50,
subspace_type='pca',
subspace_kwargs={'max_rank': 20},
use_cuda=False,
use_swag=False,
double_bias_lr=False,
model_variance=True,
num_samples=30,
scale=0.5,
const_lr=False,
*args,
**kwargs):
self.base = base
self.model = base(*args, **kwargs)
num_pars = 0
for p in self.model.parameters():
num_pars += p.numel()
print('number of parameters: ', num_pars)
if use_cuda:
self.model.cuda()
if use_swag:
self.swag_model = SWAG(base,
subspace_type=subspace_type,
subspace_kwargs=subspace_kwargs,
*args,
**kwargs)
if use_cuda:
self.swag_model.cuda()
else:
self.swag_model = None
self.use_cuda = use_cuda
if not double_bias_lr:
pars = self.model.parameters()
else:
pars = []
for name, module in self.model.named_parameters():
if 'bias' in str(name):
print('Doubling lr of ', name)
pars.append({'params': module, 'lr': 2.0 * lr_init})
else:
pars.append({'params': module, 'lr': lr_init})
self.optimizer = torch.optim.SGD(pars,
lr=lr_init,
momentum=momentum,
weight_decay=wd)
self.const_lr = const_lr
self.batch_size = batch_size
# TODO: set up criterions better for classification
if model_variance:
self.criterion = criterion(noise_var=None)
else:
self.criterion = criterion(noise_var=1.0)
if self.criterion.noise_var is not None:
self.var = self.criterion.noise_var
self.epochs = epochs
self.lr_init = lr_init
self.use_swag = use_swag
self.swag_start = swag_start
self.swag_lr = swag_lr
self.swag_freq = swag_freq
self.num_samples = num_samples
self.scale = scale
args.batch_size,
args.num_workers,
)
print('Preparing model')
model = model_class(pretrained=args.pretrained, num_classes=num_classes)
model.to(args.device)
if args.cov_mat:
args.no_cov_mat = False
else:
args.no_cov_mat = True
if args.swa:
print('SWAG training')
args.swa_device = 'cpu' if args.swa_cpu else args.device
swag_model = SWAG(model_class, no_cov_mat=args.no_cov_mat, max_num_models=20,
num_classes=num_classes)
swag_model.to(args.swa_device)
if args.pretrained:
model.to(args.swa_device)
swag_model.collect_model(model)
model.to(args.device)
else:
print('SGD training')
def schedule(epoch):
if args.swa and epoch >= args.swa_start:
return args.swa_lr
else:
return args.lr_init * (0.1 ** (epoch // 30))
args = parser.parse_args()
args.device = None
if torch.cuda.is_available():
args.device = torch.device('cuda')
else:
args.device = torch.device('cpu')
# torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
print('Using model %s' % args.model)
model_class = getattr(torchvision.models, args.model)
num_classes = 1000
print('Preparing model')
model = model_class(pretrained=args.pretrained, num_classes=num_classes)
model.to(args.device)
print('SWAG training')
swag_model = SWAG(model_class,
no_cov_mat=False,
max_num_models=20,
num_classes=num_classes)
swag_model.to('cpu')
for k in range(100):
swag_model.collect_model(model)
print(k + 1)
use_validation=not args.use_test,
split_classes=args.split_classes)
print('Preparing model')
print(*model_cfg.args)
model = model_cfg.base(*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
model.cuda()
swag_model = SWAG(model_cfg.base,
num_classes=num_classes,
subspace_type='pca',
subspace_kwargs={
'max_rank': 140,
'pca_rank': args.rank,
},
*model_cfg.args,
**model_cfg.kwargs)
swag_model.to(args.device)
def checkpoint_num(filename):
num = filename.split("-")[1]
num = num.split(".")[0]
num = int(num)
return num
for file in os.listdir(args.dir):
if torch.cuda.is_available():
args.device = torch.device('cuda')
else:
args.device = torch.device('cpu')
print('Using model %s' % args.model)
model_cfg = getattr(models, args.model)
print('Preparing model')
print(*model_cfg.args)
model = model_cfg.base(*model_cfg.args,
num_classes=args.num_classes,
**model_cfg.kwargs)
model.to(args.device)
swag_model = SWAG(model_cfg.base,
subspace_type=args.subspace,
subspace_kwargs={'max_rank': args.max_num_models},
*model_cfg.args,
num_classes=args.num_classes,
**model_cfg.kwargs)
swag_model.to(args.device)
for path in args.checkpoint:
print(path)
ckpt = torch.load(path)
model.load_state_dict(ckpt['state_dict'])
swag_model.collect_model(model)
torch.save({'state_dict': swag_model.state_dict()}, args.path)
def test_swag_diag(self, **kwargs):
model = torch.nn.Linear(300, 3, bias=True)
swag_model = SWAG(
torch.nn.Linear,
in_features=300,
out_features=3,
bias=True,
no_cov_mat=True,
max_num_models=100,
loading=False,
)
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
# construct swag model via training
torch.manual_seed(0)
for _ in range(101):
model.zero_grad()
input = torch.randn(100, 300)
output = model(input)
loss = ((torch.randn(100, 3) - output) ** 2.0).sum()
loss.backward()
optimizer.step()
swag_model.collect_model(model)
# check to ensure parameters have the correct sizes
mean_list = []
sq_mean_list = []
for (module, name), param in zip(swag_model.params, model.parameters()):
mean = module.__getattr__("%s_mean" % name)
sq_mean = module.__getattr__("%s_sq_mean" % name)
self.assertEqual(param.size(), mean.size())
self.assertEqual(param.size(), sq_mean.size())
mean_list.append(mean)
sq_mean_list.append(sq_mean)
mean = flatten(mean_list).cuda()
sq_mean = flatten(sq_mean_list).cuda()
for scale in [0.01, 0.1, 0.5, 1.0, 2.0, 5.0]:
var = scale * (sq_mean - mean ** 2)
std = torch.sqrt(var)
dist = torch.distributions.Normal(mean, std)
# now test to ensure that sampling has the correct covariance matrix probabilistically
all_qforms = 0
for _ in range(20):
swag_model.sample(scale=scale, cov=False)
curr_pars = []
for (module, name) in swag_model.params:
curr_pars.append(getattr(module, name))
curr_probs = dist.cdf(flatten(curr_pars))
# check if within 95% CI
num_in_cr = ((curr_probs > 0.025) & (curr_probs < 0.975)).float().sum()
# all_qforms.append( num_in_cr )
all_qforms += num_in_cr
# print(all_qforms/(20 * mean.numel()))
# now compute average
avg_prob_in_cr = all_qforms / (20 * mean.numel())
# CLT should hold a bit tighter here
self.assertTrue(0.945 <= avg_prob_in_cr <= 0.955)