def test_determinist_init():
seed1 = set_and_print_random_seed()
DetLin = nn.Linear(32*7*7,10)
BayLin = GaussianLinear(1, 32 * 7 * 7, 10)
set_and_print_random_seed(seed1)
BayLin.reset_parameters()
assert torch.sum(torch.abs(BayLin.mu - DetLin.weight)) == 0
assert torch.sum(torch.abs(BayLin.mu_bias - DetLin.bias)) == 0
def test_determinist_init():
seed1 = set_and_print_random_seed()
DetCNN = nn.Conv2d(1, 16, 3)
BayCNN = GaussianCNN(1, 1, 16, 3)
set_and_print_random_seed(seed1)
BayCNN.reset_parameters()
assert torch.sum(torch.abs(BayCNN.mu - DetCNN.weight)) == 0
assert torch.sum(torch.abs(BayCNN.mu_bias - DetCNN.bias)) == 0
def test_determinist_forward_and_init():
seed1 = set_and_print_random_seed()
DetLin = nn.Linear(32*7*7, 10)
BayLin = GaussianLinear('determinist', 32 * 7 * 7, 10)
set_and_print_random_seed(seed1)
BayLin.reset_parameters()
image = torch.rand(1, 32*7*7)
output1 = BayLin(image)
output2 = DetLin(image)
assert torch.sum(torch.abs(output1 - output2)) == 0
def test_determinist_forward_and_init():
seed1 = set_and_print_random_seed()
DetCNN = nn.Conv2d(1, 16, 3)
BayCNN = GaussianCNN('determinist', 1, 16, 3)
set_and_print_random_seed(seed1)
BayCNN.reset_parameters()
image = torch.rand(1, 1, 28, 28)
output1 = BayCNN(image)
output2 = DetCNN(image)
assert torch.sum(torch.abs(output1 - output2)) == 0
def eval_random(model,
batch_size,
img_channels,
img_dim,
number_of_tests,
random_seed=None,
verbose=False,
device='cpu'):
"""
Args:
model (torch.nn.Module child): the model we evaluate
batch_size (int): The size of the random sample
img_channels (int): dimension of the random sample
img_dim (int): dimension of the random sample
number_of_tests (int): the number of times we do a forward for each input
random_seed (int): the seed of the random generation, for reproducibility
verbose (Bool): whether we want a progress bar or not
device (torch.device || str): which device to compute on (either on GPU or CPU). Either torch.device type or
specific string 'cpu' or 'gpu'.
Returns:
torch.Tensor: size (batch_size): the variation-ratio uncertainty
torch.Tensor: size (batch_size): the predictive entropy uncertainty
torch.Tensor: size (batch_size): the mutual information uncertainty
int: the seed of the random generation, for reproducibility
"""
number_of_classes = model.number_of_classes
seed = set_and_print_random_seed(random_seed)
random_noise = torch.randn(batch_size, img_channels, img_dim,
img_dim).to(device)
output_random = torch.Tensor(number_of_tests, batch_size,
number_of_classes)
if verbose:
iterator = tqdm(range(number_of_tests))
else:
iterator = range(number_of_tests)
for test_idx in iterator:
output_random[test_idx] = model(random_noise).detach()
return output_random, seed
else:
device = 'cpu'
device = torch.device(device)
if trainset == 'mnist':
trainloader, valloader, evalloader = get_mnist(train_labels=range(10),
eval_labels=range(10),
batch_size=batch_size)
dim_input = 28
dim_channels = 1
if trainset == 'cifar10':
trainloader, evalloader = get_cifar10(batch_size=batch_size)
dim_input = 32
dim_channels = 3
seed_model = set_and_print_random_seed()
bay_net = GaussianClassifier(rho=rho,
stds_prior=stds_prior,
dim_input=dim_input,
number_of_classes=10,
dim_channels=dim_channels)
bay_net.to(device)
criterion = CrossEntropyLoss()
if loss_type == 'uniform':
step_function = uniform
loss = BBBLoss(bay_net, criterion, step_function)
elif loss_type == 'exp':
def step_function(batch_idx, number_of_batches):
return 2**(number_of_batches - batch_idx) / (2**number_of_batches - 1)