def reset_parameters(self):
if self.layer_size == -1:
self.layer_size = self.in_size
self.n = self.layer_size
self.LDR1 = sl.StructuredLinear(self.class1, layer_size=self.channels*self.n, r=self.rank1, bias=True)
self.LDR2 = sl.StructuredLinear(self.class2,layer_size=self.channels*self.n, r=self.rank2, bias=True)
self.logits = nn.Linear(self.fc_size, 10)
def reset_parameters(self):
if self.layer_size == -1:
self.layer_size = self.in_size
if self.hidden_size == -1:
self.hidden_size = self.in_size
self.W = sl.StructuredLinear(self.class_type, layer_size=self.layer_size, r=self.r, bias=self.bias,
hidden_size=self.hidden_size)
def reset_parameters(self):
if self.layer_size == -1:
self.layer_size = self.in_size
layers = []
for layer in range(self.num_layers):
layers.append(sl.StructuredLinear(self.class_type,layer_size=self.layer_size, r=self.r, bias=self.bias))
self.layers = nn.ModuleList(layers)
self.W2 = nn.Linear(self.layer_size, self.out_size)
def __init__(self):
super(VAE, self).__init__()
self.fc1 = sl.StructuredLinear(class_type=args.class_type,
layer_size=args.layer_size,
r=args.r,
bias=False)
self.fc21 = nn.Linear(784, 20)
self.fc22 = nn.Linear(784, 20)
self.fc3 = nn.Linear(20, 400)
self.fc4 = nn.Linear(400, 784)
def reset_parameters(self):
self.layer_size = int(self.in_size/3)
if self.hidden_size == -1:
self.hidden_size = self.layer_size
self.d = int(np.sqrt(self.layer_size))
self.conv1 = nn.Conv2d(3, 6, 5, padding=2)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5, padding=2)
self.W = sl.StructuredLinear(self.class_type, layer_size=self.layer_size, r=self.r, bias=self.bias,
hidden_size=self.hidden_size)
self.logits = nn.Linear(self.hidden_size, self.out_size)
def reset_parameters(self):
if self.layer_size == -1:
self.layer_size = self.in_size
if self.hidden_size == -1:
self.hidden_size = self.layer_size
assert self.layer_size == self.in_size
self.d = int(np.sqrt(self.layer_size))
self.conv1 = nn.Conv2d(1, 6, 5, padding=2)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5, padding=2)
layers = []
layers.append(sl.StructuredLinear(self.class_type, layer_size=self.layer_size, r=self.r, bias=self.bias,
hidden_size=self.hidden_size))
self.layers = nn.ModuleList(layers)
self.logits = nn.Linear(self.hidden_size, self.out_size)
def reset_parameters(self):
self.n = 1024
self.fc_size = self.n // 2
if self.channels:
self.LDR1 = ldr.LDR(self.class1, 3, 3, self.rank1, self.n)
else:
self.LDR1 = sl.StructuredLinear(self.class1, layer_size=3*self.n, r=self.rank1)
self.LDR211 = sl.StructuredLinear(self.class2, layer_size=self.fc_size, r=self.rank2)
self.LDR212 = sl.StructuredLinear(self.class2, layer_size=self.fc_size, r=self.rank2)
self.LDR221 = sl.StructuredLinear(self.class2, layer_size=self.fc_size, r=self.rank2)
self.LDR222 = sl.StructuredLinear(self.class2, layer_size=self.fc_size, r=self.rank2)
self.LDR231 = sl.StructuredLinear(self.class2, layer_size=self.fc_size, r=self.rank2)
self.LDR232 = sl.StructuredLinear(self.class2, layer_size=self.fc_size, r=self.rank2)
self.b = Parameter(torch.zeros(self.fc_size))
self.logits = nn.Linear(self.fc_size, 10)
def __init__(self, class_type, r, input_size, hidden_size, use_bias=True):
super(LSTMCell, self).__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.use_bias = use_bias
self.class_type = class_type
self.r = r
# Replace W_ih with structured matrices
self.W_ih = sl.StructuredLinear(class_type,
layer_size=4 * hidden_size,
r=r,
bias=False)
self.W_hh = nn.Parameter(
torch.FloatTensor(hidden_size, 4 * hidden_size))
if use_bias:
self.bias = nn.Parameter(torch.FloatTensor(4 * hidden_size))
else:
self.bias = None
self.reset_parameters()
def reset_parameters(self):
if self.layer_size == -1: self.layer_size = self.in_size
self.W = sl.StructuredLinear(self.class_type, layer_size=self.layer_size, r=self.r)
self.fc = nn.Linear(3*1024, self.fc_size)
self.logits = nn.Linear(self.fc_size, 10)