def __init__(self):
"""Declare all needed layers."""
torch.nn.Module.__init__(self)
# Convolution and pooling layers of VGG-16.
self.features = torchvision.models.vgg16(pretrained=False).features
self.features = torch.nn.Sequential(*list(
self.features.children())[:-1]) # Remove pool5.
# Linear classifier.
#self.fc = torch.nn.Linear(512**2, 200)
self.d1 = 16
self.d2 = 3
self.w1 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) -
1).float()) #, requires_grad = False)
self.w2 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) -
1).float()) #, requires_grad = False)
self.w3 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) -
1).float()) #, requires_grad = False)
self.w4 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) - 1).float())
self.fc = torch.nn.Linear((512 // self.d1 * self.d2)**2, 67)
self.norm = runLayer()
def __init__(self):
"""Declare all needed layers."""
torch.nn.Module.__init__(self)
# Convolution and pooling layers of VGG-16.
self.features = torchvision.models.vgg16(pretrained=True).features
self.features = torch.nn.Sequential(*list(
self.features.children())[:-1]) # Remove pool5.
self.d1 = 16
self.d2 = 3
self.w1 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) -
1).float()) #, requires_grad = False)
self.w2 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) -
1).float()) #, requires_grad = False)
self.w3 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) -
1).float()) #, requires_grad = False)
self.w4 = torch.nn.Parameter(
(2 * torch.randint(0, 2, (self.d1, self.d2)) - 1).float())
self.fc = torch.nn.Linear((512 // self.d1 * self.d2)**2, 67)
self.norm = runLayer()
for param in self.features.parameters():
param.requires_grad = False
torch.nn.init.kaiming_normal(self.fc.weight.data)
torch.nn.init.kaiming_normal(self.w1.data)
torch.nn.init.kaiming_normal(self.w2.data)
torch.nn.init.kaiming_normal(self.w3.data)
torch.nn.init.kaiming_normal(self.w4.data)
if self.fc.bias is not None:
torch.nn.init.constant(self.fc.bias.data, val=0)
