def forward_hook(module, input, output, module_name=None):
# Compose module id
module_iter = counter[module_name]
module_id = (module_name, module_iter)
# Check if the module has to be altered
if module_id in dirs:
# TODO: handle convolutional layers
# TODO: use canonical basis
if basis[module_id] is None:
basis[module_id] = torch.identity(output)
# Change basis
output = output @ basis[module_id]
# Zero out the directions
mask = torch.ones_like(output)
mask[:, dirs[module_id]] = 0
output = output * mask
# Replace the directions with constant mu
mask = torch.zeros_like(output)
mask[:, dirs[module_id]] = mu
output = output + mask
# Update counter
counter[module_name] += 1
return output
def Likelihood_Y_Z(MuZ,SigmaZ,H,Y,beta):
MuY=H@MuZ
(d,N)=MuY.size()
totalTerm=0
#logdetP=float(0)
for i in range(N):
yi=Y[:,i]
mui=MuY[:,i]
deviation=yi-mui
outerProduct=torch.gres(deviation,deviation.transpose(0,1))
Sigmai=torch.diag(SigmaZ[:,i])
Covi=(H@Sigmai)@H.transpose(0,1)+(1/beta)*torch.identity(d)
Pr=torch.inverse(Covi)
expTerm=torch.sum((outerProduct*Pr))
logdet=torch.log(torch.det(Pr))
#print('Determinant is:', determin)
#print(Covi)
#prod*=1/(math.exp(0.5*expTerm)*math.sqrt(lin.det(Covi)))
totalTerm=totalTerm+0.5*(logdet-expTerm)
return totalTerm
def random_flip(input, decision):
identity = torch.identity()
f1 = identity(input)
f2 = torch.flip(input, dim=3)
return torch.where(torch.less(decision, 0.5), f2, f1)
def batchnorm3d_to_instancenorm3d(model):
conversion_count = 0
for name, module in reversed(model._modules.items()):
if len(list(module.children())) > 0:
# recurse
model._modules[
name], num_converted = batchnorm3d_to_instancenorm3d(module)
conversion_count += num_converted
if isinstance(module, nn.BatchNorm3d):
layer_new = nn.InstanceNorm3d(module.num_features,
eps=module.eps,
momentum=module.momentum,
affine=module.affine,
track_running_stats=False)
layer_new.weight = module.weight
layer_new.bias = module.bias
model._modules[name] = layer_new
conversion_count += 1
return model, conversion_count
if __name__ == '__main__':
model = torch.identity()
model2 = deepcopy(model).cpu()
new_model, count = batchnorm3d_to_instancenorm3d(model2, nn.BatchNorm3d,
nn.InstanceNorm3d)
