def weights_init(self, m):
classname = m.__class__.__name__
if classname == 'Conv2d' or classname == 'Linear':
print(classname, 'apply xavier')
xavier(m.weight.data)
#xavier(m.bias.data)
m.bias.data.fill_(0.1)
def __init__(self):
super().__init__()
self.fc1 = nn.Linear(28 * 28, 100)
self.fc2 = nn.Linear(100, 10)
for m in self.modules():
if isinstance(m, nn.Linear):
xavier(m.weight.data)
def weights_init(m):
import torch.nn.init.xavier_uniform as xavier
if isinstance(m, nn.Linear):
xavier(m.weight.data)
xavier(m.bias.data)
def weights_init(m):
if isinstance(m, nn.Conv2d):
xavier(m.weight.data,gain=args.gain)
m.bias.data.zero_()
criterion_train = nn.MSELoss()
criterion = nn.MSELoss()
elif args.loss=='ce':
criterion = nn.CrossEntropyLoss()
criterion_train = nn.CrossEntropyLoss()
from torch.nn.init import xavier_normal_ as xavier
def weights_init(m):
if isinstance(m, nn.Conv2d):
xavier(m.weight.data,gain=args.gain)
m.bias.data.zero_()
net.apply(weights_init)
FC1 = nn.Linear(512*k, 10).cuda()
FC2 = nn.Linear(512*k, 10).cuda()
xavier(FC1.weight.data, gain=args.gain)
FC1.bias.data.zero_()
net2=copy.deepcopy(net)
# Symmetrize!
FC2.weight.data.copy_(FC1.weight.data)
FC2.bias.data.copy_(FC1.bias.data)
par = list(net.parameters())+list(FC1.parameters())
optimizer = optim.SGD(par, lr=args.lr, momentum=0.9, weight_decay=0)
net_clone = copy.deepcopy(net)
def reset_parameters(self):
xavier(self.weight)
zeros(self.bias)
print("Glorot Initialized Weights")
def weights_init_xavier(m):
if isinstance(m, nn.Conv2d):
if not isinstance(m.weight, torch.HalfTensor):
xavier(m.weight.data)
if m.bias is not None:
m.bias.data.fill_(0)
