def forward(self, x):
batch_size = x.shape[0]
self.spiking_linear3.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_linear4.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_linear1.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_linear2.initial_state(batch_size=batch_size,
device=self.device)
spike_out = torch.zeros((batch_size, self.num_classes),
device=self.device)
accuracy_record = torch.zeros((self.seq_length, 2), device='cpu')
for i in range(self.seq_length):
rescale_fac = 1 / (self.max_firing_rate *
self.dt) * (torch.max(x) - torch.min(x))
spike_snapshot = rescale_fac * torch.rand_like(x,
device=self.device)
inp_image = threshold(x, spike_snapshot)
x_spike = self.spiking_linear1.forward(inp_image)
r_spike = self.spiking_linear2.forward(x_spike)
x_spike = self.Weighted * x_spike + r_spike
x_spike = self.spiking_linear3.forward(x_spike)
x_spike = self.spiking_linear4.forward(x_spike)
spike_out += x_spike
pred = spike_out.data.max(1, keepdim=True)[1]
cur_correct = pred.eq(self.y.data.view_as(pred)).cpu().sum()
accuracy_record[i, 0] = cur_correct
accuracy_record[i, 1] = batch_size
return spike_out, accuracy_record
def forward(self, input_current):
self.v_mem += input_current
spike = threshold(self.v_mem, self.if_parameters.v_th)
self.v_mem = (1 -
spike) * self.v_mem + spike * self.if_parameters.v_reset
return spike
def forward(self, input_current):
self.v_mem += input_current * self.rC
spike = threshold(self.v_mem, self.if_parameters.v_th)
delta_rC = 1.0 / self.rC - self.sigma * spike * input_current + self.beta * (
1 - self.sigma * spike) * input_current
# delta_rR = self.tau * (-self.rR + self.sigma * spike - self.beta * (1 - self.sigma * spike))
self.rC += self.tau * delta_rC
# self.rR += delta_rR
return spike
def forward(self, x):
batch_size = x.shape[0]
self.spiking_conv1.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_conv2_0.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_conv2_1.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_conv3.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_conv4.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_linear1.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_linear2.initial_state(batch_size=batch_size,
device=self.device)
spike_out = torch.zeros((batch_size, self.num_classes),
device=self.device)
output_record = torch.zeros((self.seq_length, batch_size),
device='cpu')
for i in range(self.seq_length):
rescale_fac = 1 / (self.max_firing_rate *
self.dt) * (torch.max(x) - torch.min(x))
spike_snapshot = rescale_fac * torch.rand_like(x,
device=self.device)
inp_image = threshold(x, spike_snapshot)
x_spike = self.spiking_conv1.forward(inp_image)
r_spike = self.spiking_conv2_0.forward(x_spike)
r_spike = self.spiking_conv2_1.forward(r_spike)
x_spike = r_spike + self.Weighted * x_spike
x_spike = self.spiking_conv3.forward(x_spike)
x_spike = torch.nn.AvgPool2d(kernel_size=(2, 2),
stride=(2, 2))(x_spike)
x_spike = self.spiking_conv4.forward(x_spike)
x_spike = torch.nn.AvgPool2d(kernel_size=(2, 2),
stride=(2, 2))(x_spike)
x_spike = x_spike.view(-1, self.num_flat_features(x_spike))
x_spike = self.spiking_linear1.forward(x_spike)
x_spike = self.spiking_linear2.forward(x_spike)
spike_out += x_spike
pred = spike_out.data.max(1, keepdim=True)[1]
output_record[i, :] = torch.transpose(pred, 1, 0)
return spike_out, output_record
def forward(self, input_current):
self.v_mem += input_current * self.rC
spike = threshold(self.v_mem, self.if_parameters.v_th)
input_event = torch.where(
input_current >
0.001, # Use a small positive value instead of zero.
torch.zeros_like(input_current),
torch.ones_like(input_current),
)
delta_rC = 1.0 / self.rC - self.sigma * spike * input_current + self.beta * (
1 - self.sigma * spike) * input_current
# delta_rR = self.tau * (-self.rR + self.sigma * spike - self.beta * (1 - self.sigma * spike))
self.rC += self.tau * delta_rC * input_event
# self.rR += delta_rR
return spike
def forward(self, x):
batch_size = x.shape[0]
self.spiking_conv1.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_conv2.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_linear1.initial_state(batch_size=batch_size,
device=self.device)
self.spiking_linear2.initial_state(batch_size=batch_size,
device=self.device)
spike_out = torch.zeros((batch_size, self.num_classes),
device=self.device)
accuracy_record = torch.zeros((self.seq_length, 2), device='cpu')
for i in range(self.seq_length):
rescale_fac = 1 / (self.max_firing_rate *
self.dt) * (torch.max(x) - torch.min(x))
spike_snapshot = rescale_fac * torch.rand_like(x,
device=self.device)
inp_image = threshold(x, spike_snapshot)
x_spike = self.spiking_conv1.forward(inp_image)
x_spike = torch.nn.functional.avg_pool2d(x_spike,
kernel_size=(2, 2),
stride=(2, 2))
x_spike = self.spiking_conv2.forward(x_spike)
x_spike = torch.nn.functional.avg_pool2d(x_spike,
kernel_size=(2, 2),
stride=(2, 2))
x_spike = x_spike.view(-1, self.num_flat_features(x_spike))
x_spike = self.spiking_linear1.forward(x_spike)
x_spike = self.spiking_linear2.forward(x_spike)
spike_out += x_spike
pred = spike_out.data.max(1, keepdim=True)[1]
cur_correct = pred.eq(self.y.data.view_as(pred)).cpu().sum()
accuracy_record[i, 0] = cur_correct
accuracy_record[i, 1] = batch_size
return spike_out, accuracy_record
