def lsm(n_inputs, n_classes, n_steps, epoches, x_train, x_test, y_train,
y_test, classifier, data_set):
dim1 = [5, 5, 4]
dim2 = [5, 5, 6]
dim3 = [5, 5, 8]
layers = []
r1 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r2 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r3 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r4 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r5 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r6 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r7 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r8 = reservoir.ReservoirLayer(int(n_inputs / 8),
100,
n_steps,
dim1,
is_input=True,
homeostasis=True)
r9 = reservoir.ReservoirLayer(200,
150,
n_steps,
dim2,
is_input=False,
homeostasis=True)
r10 = reservoir.ReservoirLayer(200,
150,
n_steps,
dim2,
is_input=False,
homeostasis=True)
r11 = reservoir.ReservoirLayer(200,
150,
n_steps,
dim2,
is_input=False,
homeostasis=True)
r12 = reservoir.ReservoirLayer(200,
150,
n_steps,
dim2,
is_input=False,
homeostasis=True)
r13 = reservoir.ReservoirLayer(300,
150,
n_steps,
dim2,
is_input=False,
homeostasis=True)
r14 = reservoir.ReservoirLayer(300,
150,
n_steps,
dim2,
is_input=False,
homeostasis=True)
r15 = reservoir.ReservoirLayer(300,
200,
n_steps,
dim3,
is_input=False,
homeostasis=True)
s1 = feedforward.SpikingLayer(200, n_classes, n_steps)
layers.append(r1)
layers.append(r2)
layers.append(r3)
layers.append(r4)
layers.append(r5)
layers.append(r6)
layers.append(r7)
layers.append(r8)
layers.append(r9)
layers.append(r10)
layers.append(r11)
layers.append(r12)
layers.append(r13)
layers.append(r14)
layers.append(r15)
layers.append(s1)
accuracy = 0
best_acc_e = 0
if classifier == "calcium_supervised":
for e in range(epoches):
record = np.zeros(n_classes)
for i in tqdm(range(len(x_train))): # train phase
reset(layers)
if data_set == "TI46":
x = np.asarray(x_train[i].todense())
elif data_set == "MNIST" and record[y_train[i]] < 80:
x = loadDataset.genrate_poisson_spikes(x_train[i], n_steps)
record[y_train[i]] += 1
else:
continue
o_r1 = r1.forward(x[:,
int(0 * n_inputs / 8):int(1 * n_inputs /
8)])
o_r2 = r2.forward(x[:,
int(1 * n_inputs / 8):int(2 * n_inputs /
8)])
o_r3 = r3.forward(x[:,
int(2 * n_inputs / 8):int(3 * n_inputs /
8)])
o_r4 = r4.forward(x[:,
int(3 * n_inputs / 8):int(4 * n_inputs /
8)])
o_r5 = r5.forward(x[:,
int(4 * n_inputs / 8):int(5 * n_inputs /
8)])
o_r6 = r6.forward(x[:,
int(5 * n_inputs / 8):int(6 * n_inputs /
8)])
o_r7 = r7.forward(x[:,
int(6 * n_inputs / 8):int(7 * n_inputs /
8)])
o_r8 = r8.forward(x[:,
int(7 * n_inputs / 8):int(8 * n_inputs /
8)])
o_r9 = r9.forward(np.concatenate((o_r1, o_r2), 1))
o_r10 = r10.forward(np.concatenate((o_r3, o_r4), 1))
o_r11 = r11.forward(np.concatenate((o_r5, o_r6), 1))
o_r12 = r12.forward(np.concatenate((o_r7, o_r8), 1))
o_r13 = r13.forward(np.concatenate((o_r9, o_r10), 1))
o_r14 = r14.forward(np.concatenate((o_r11, o_r12), 1))
o_r15 = r15.forward(np.concatenate((o_r13, o_r14), 1))
s1.forward(o_r15, e, y_train[i])
correct = 0
for i in tqdm(range(len(x_test))): # test phase
reset(layers)
if data_set == "TI46":
x = np.asarray(x_test[i].todense())
elif data_set == "MNIST" and record[y_test[i]] < 100:
x = loadDataset.genrate_poisson_spikes(x_test[i], n_steps)
record[y_test[i]] += 1
else:
continue
o_r1 = r1.forward(x[:,
int(0 * n_inputs / 8):int(1 * n_inputs /
8)])
o_r2 = r2.forward(x[:,
int(1 * n_inputs / 8):int(2 * n_inputs /
8)])
o_r3 = r3.forward(x[:,
int(2 * n_inputs / 8):int(3 * n_inputs /
8)])
o_r4 = r4.forward(x[:,
int(3 * n_inputs / 8):int(4 * n_inputs /
8)])
o_r5 = r5.forward(x[:,
int(4 * n_inputs / 8):int(5 * n_inputs /
8)])
o_r6 = r6.forward(x[:,
int(5 * n_inputs / 8):int(6 * n_inputs /
8)])
o_r7 = r7.forward(x[:,
int(6 * n_inputs / 8):int(7 * n_inputs /
8)])
o_r8 = r8.forward(x[:,
int(7 * n_inputs / 8):int(8 * n_inputs /
8)])
o_r9 = r9.forward(np.concatenate((o_r1, o_r2), 1))
o_r10 = r10.forward(np.concatenate((o_r3, o_r4), 1))
o_r11 = r11.forward(np.concatenate((o_r5, o_r6), 1))
o_r12 = r12.forward(np.concatenate((o_r7, o_r8), 1))
o_r13 = r13.forward(np.concatenate((o_r9, o_r10), 1))
o_r14 = r14.forward(np.concatenate((o_r11, o_r12), 1))
o_r15 = r15.forward(np.concatenate((o_r13, o_r14), 1))
o_s1 = s1.forward(o_r15)
fire_count = np.sum(o_s1, axis=0)
# print(fire_count)
index = np.argmax(fire_count)
if index == y_test[i]:
correct = correct + 1
acc = correct / len(x_test)
print("test accuracy at epoch %d is %0.2f%%" % (e, acc * 100))
if accuracy < acc:
accuracy = acc
best_acc_e = e
elif classifier == "svmcv":
samples = []
label = []
record = np.zeros(n_classes)
for i in tqdm(range(len(x_train))): # train phase
reset(layers)
if data_set == "TI46":
x = np.asarray(x_train[i].todense())
elif data_set == "MNIST" and record[y_train[i]] < 100:
x = loadDataset.genrate_poisson_spikes(x_train[i], n_steps)
record[y_train[i]] += 1
else:
continue
o_r1 = r1.forward(x[:,
int(0 * n_inputs / 8):int(1 * n_inputs / 8)])
o_r2 = r2.forward(x[:,
int(1 * n_inputs / 8):int(2 * n_inputs / 8)])
o_r3 = r3.forward(x[:,
int(2 * n_inputs / 8):int(3 * n_inputs / 8)])
o_r4 = r4.forward(x[:,
int(3 * n_inputs / 8):int(4 * n_inputs / 8)])
o_r5 = r5.forward(x[:,
int(4 * n_inputs / 8):int(5 * n_inputs / 8)])
o_r6 = r6.forward(x[:,
int(5 * n_inputs / 8):int(6 * n_inputs / 8)])
o_r7 = r7.forward(x[:,
int(6 * n_inputs / 8):int(7 * n_inputs / 8)])
o_r8 = r8.forward(x[:,
int(7 * n_inputs / 8):int(8 * n_inputs / 8)])
o_r9 = r9.forward(np.concatenate((o_r1, o_r2), 1))
o_r10 = r10.forward(np.concatenate((o_r3, o_r4), 1))
o_r11 = r11.forward(np.concatenate((o_r5, o_r6), 1))
o_r12 = r12.forward(np.concatenate((o_r7, o_r8), 1))
o_r13 = r13.forward(np.concatenate((o_r9, o_r10), 1))
o_r14 = r14.forward(np.concatenate((o_r11, o_r12), 1))
o_r15 = r15.forward(np.concatenate((o_r13, o_r14), 1))
fire_count = np.sum(o_r15, axis=0)
samples.append(fire_count)
label.append(y_train[i])
accuracy = svm.cvSVM(samples, label, 5)
best_acc_e = 0
else:
print('Given classifier {} not found'.format(classifier))
sys.exit(-1)
return accuracy, best_acc_e
def lsm(n_inputs, n_classes, n_steps, epoches, x_train, x_test, y_train,
y_test, classifier, data_set):
stdp = False # stdp enabled
dim1 = [3, 3, 15]
r1 = reservoir.ReservoirLayer(n_inputs,
135,
n_steps,
dim1,
is_input=True,
homeostasis=True)
s1 = feedforward.SpikingLayer(135, n_classes, n_steps, homeostasis=True)
accuracy = 0
best_acc_e = 0
# train stdp
if stdp:
# r1.stdp_i = True
r1.stdp_r = True
# s1.stdp_i = True
print("start stdp")
for e_stdp in range(5):
for i in tqdm(range(len(x_train))):
r1.reset()
s1.reset()
x = np.asarray(x_train[i].todense())
o_r1 = r1.forward(x)
s1.forward(o_r1)
print("finish stdp")
r1.stdp_i = False
r1.stdp_r = False
s1.stdp_i = False
if classifier == "calcium_supervised":
for e in range(epoches):
for i in tqdm(range(len(x_train))): # train phase
r1.reset()
s1.reset()
if data_set == "TI46":
x = np.asarray(x_train[i].todense())
elif data_set == "MNIST":
x = loadDataset.genrate_poisson_spikes(x_train[i], n_steps)
else:
continue
o_r1 = r1.forward(x)
s1.forward(o_r1, e, y_train[i])
correct = 0
for i in tqdm(range(len(x_test))): # test phase
r1.reset()
s1.reset()
if data_set == "TI46":
x = np.asarray(x_test[i].todense())
elif data_set == "MNIST":
x = loadDataset.genrate_poisson_spikes(x_test[i], n_steps)
else:
continue
o_r1 = r1.forward(x)
o_s1 = s1.forward(o_r1)
fire_count = np.sum(o_s1, axis=0)
index = np.argmax(fire_count)
if index == y_test[i]:
correct = correct + 1
acc = correct / len(x_test)
print("test accuracy at epoch %d is %0.2f%%" % (e, acc * 100))
if accuracy < acc:
accuracy = acc
best_acc_e = e
elif classifier == "svm":
train_samples = []
test_samples = []
for i in tqdm(range(len(x_train))): # train phase
r1.reset()
if data_set == "TI46":
x = np.asarray(x_train[i].todense())
elif data_set == "MNIST":
x = loadDataset.genrate_poisson_spikes(x_train[i], n_steps)
else:
continue
o_r1 = r1.forward(x)
fire_count = np.sum(o_r1, axis=0)
train_samples.append(fire_count)
for i in tqdm(range(len(x_test))):
r1.reset()
if data_set == "TI46":
x = np.asarray(x_test[i].todense())
elif data_set == "MNIST":
x = loadDataset.genrate_poisson_spikes(x_test[i], n_steps)
else:
continue
o_r1 = r1.forward(x)
fire_count = np.sum(o_r1, axis=0)
test_samples.append(fire_count)
accuracy = svm.traintestSVM(train_samples, y_train, test_samples,
y_test)
best_acc_e = 0
elif classifier == "svmcv":
samples = []
label = []
for i in tqdm(range(len(x_train))):
r1.reset()
if data_set == "TI46":
x = np.asarray(x_train[i].todense())
elif data_set == "MNIST":
x = loadDataset.genrate_poisson_spikes(x_train[i], n_steps)
else:
continue
o_r1 = r1.forward(x)
fire_count = np.sum(o_r1, axis=0)
samples.append(fire_count)
label.append(y_train[i])
accuracy = svm.cvSVM(samples, label, 5)
best_acc_e = 0
else:
print('Given classifier {} not found'.format(classifier))
sys.exit(-1)
return accuracy, best_acc_e
def lsm(n_inputs, n_classes, n_steps, epoches, x_train, x_test, y_train, y_test, classifier, data_set):
dim1 = [10, 10, 20]
r1 = reservoir.ReservoirLayer(n_inputs, 2000, n_steps, dim1, is_input=True, homeostasis=True)
s1 = feedforward.SpikingLayer(2000, n_classes, n_steps)
accuracy = 0
best_acc_e = 0
if classifier == "calcium_supervised":
for e in range(epoches):
record = np.zeros(n_classes)
for i in tqdm(range(len(x_train))): # train phase
r1.reset()
s1.reset()
if data_set == "TI46":
x = np.asarray(x_train[i].todense())
elif data_set == "MNIST" and record[y_train[i]] < 80:
x = loadDataset.genrate_poisson_spikes(x_train[i], n_steps)
record[y_train[i]] += 1
else:
continue
o_r1 = r1.forward(x)
s1.forward(o_r1, e, y_train[i])
correct = 0
num_test = 0
for i in tqdm(range(len(x_test))): # test phase
r1.reset()
s1.reset()
if data_set == "TI46":
x = np.asarray(x_test[i].todense())
elif data_set == "MNIST" and record[y_test[i]] < 100:
x = loadDataset.genrate_poisson_spikes(x_test[i], n_steps)
record[y_test[i]] += 1
else:
continue
o_r1 = r1.forward(x)
o_s1 = s1.forward(o_r1)
fire_count = np.sum(o_s1, axis=0)
# print(fire_count)
index = np.argmax(fire_count)
if index == y_test[i]:
correct = correct + 1
num_test += 1
acc = correct / num_test
print("test accuracy at epoch %d is %0.2f%%" % (e, acc * 100))
if accuracy < acc:
accuracy = acc
best_acc_e = e
elif classifier == "svmcv":
samples = []
label = []
record = np.zeros(n_classes)
for i in tqdm(range(len(x_train))): # train phase
r1.reset()
s1.reset()
if data_set == "TI46":
x = np.asarray(x_train[i].todense())
elif data_set == "MNIST" and record[y_train[i]] < 100:
x = loadDataset.genrate_poisson_spikes(x_train[i], n_steps)
record[y_train[i]] += 1
else:
continue
o_r1 = r1.forward(x)
fire_count = np.sum(o_r1, axis=0)
samples.append(fire_count)
label.append(y_train[i])
accuracy = svm.cvSVM(samples, label, 5)
best_acc_e = 0
else:
print('Given classifier {} not found'.format(classifier))
sys.exit(-1)
return accuracy, best_acc_e