m = train_i[np.random.randint(low=0, high=450)] # 随机抽取一个数据
real_class = y_train[m].item() # 真实的类别
for i in range(class_num * per_class_neuron_num):
# 分别送入dec_layer的每一个class
neural_class = i // per_class_neuron_num # 表示这个神经元应该对哪一类响应1
neural_seq = i % per_class_neuron_num
Gi = F.sigmoid(B[i])
dec_layer[i].calculate_membrane_potentials(W[i] * Gi, t_spike[m])
t_max = dec_layer[i].v.argmax()
# 训练第i个分类器
if neural_class == real_class:
# 应该放电
if dec_layer[i].v[t_max] < dec_layer[i].v_thr:
v_error = node.psp_kernel(t_max - t_spike[m],
dec_layer[i].v0,
dec_layer[i].tau,
dec_layer[i].tau_s)
dW = Gi * v_error
dB = W[i] * v_error * Gi * (1 - Gi)
W[i] += learn_rate * dW
B[i] += learn_rate * dB
else:
# 不应该放电
if dec_layer[i].v[t_max] > dec_layer[i].v_thr:
v_error = node.psp_kernel(t_max - t_spike[m],
dec_layer[i].v0,
dec_layer[i].tau,
dec_layer[i].tau_s)
dW = Gi * v_error
dB = W[i] * v_error * Gi * (1 - Gi)
W[i] -= learn_rate * dW
while 1:
m = train_i[np.random.randint(low=0, high=450)] # 随机抽取一个数据
real_class = y_train[m].item() # 真实的类别
for i in range(class_num * per_class_neuron_num):
# 分别送入dec_layer的每一个class
neural_class = i // per_class_neuron_num # 表示这个神经元应该对哪一类响应1
neural_seq = i % per_class_neuron_num
dec_layer[i].calculate_membrane_potentials(W[i], t_spike[m])
t_max = dec_layer[i].v.argmax()
# 训练第i个分类器
if neural_class == real_class:
# 应该放电
if dec_layer[i].v[t_max] < dec_layer[i].v_thr:
W[i] += learn_rate * node.psp_kernel(
t_max - t_spike[m], dec_layer[i].v0, dec_layer[i].tau,
dec_layer[i].tau_s)
else:
# 不应该放电
if dec_layer[i].v[t_max] > dec_layer[i].v_thr:
W[i] -= learn_rate * node.psp_kernel(
t_max - t_spike[m], dec_layer[i].v0, dec_layer[i].tau,
dec_layer[i].tau_s)
# 计算第i个分类器的错误率
if train_times % 128 == 0:
# 测试一次
error_times = 0
for m in test_i:
real_class = y_train[m].item() # 真实的类别
# 比较每一个分类器的输出结果