def app(opt):
import numpy as np
np.set_printoptions(threshold=np.inf)
print(opt)
data_loader = n3ml.data.IRISDataLoader()
data = data_loader.run()
summary = data_loader.summarize()
data_encoder = n3ml.encoder.Population(neurons=12,
minimum=summary['min'],
maximum=summary['max'],
max_firing_time=opt.max_firing_time,
not_to_fire=opt.not_to_fire,
dt=opt.dt)
label_encoder = LabelEncoder(opt.num_classes)
model = n3ml.model.Bohte2002()
model.initialize()
optimizer = n3ml.optimizer.Bohte()
for epoch in range(opt.num_epochs):
acc = train(data, model, data_encoder, label_encoder, optimizer, rmse, epoch, opt)
print("epoch: {} - tr. acc: {}".format(epoch, acc))
acc = validate(data, model, data_encoder, label_encoder, opt)
print("epoch: {} - val. acc: {}".format(epoch, acc))
def train(data, model, data_encoder, label_encoder, optimizer, loss, epoch, opt):
total_data = 0
corrects = 0
for i in range(data['train.data'].size(0)):
model.initialize(delay=False)
input = data['train.data'][i]
label = data['train.target'][i]
spiked_input = data_encoder.run(input)
spiked_input = torch.cat((spiked_input.view(-1), torch.zeros(2)))
spiked_label = label_encoder.run(label)
for t in range(opt.num_steps):
model(torch.tensor(t).float(), spiked_input)
o = model.fc2.s
print(model.fc1.s)
print(model.fc2.s)
print("pred: {} - target: {}".format(o, spiked_label))
# l = loss(o, spiked_label)
# print("loss: {}".format(l))
optimizer.step(model, spiked_input, spiked_label, epoch)
total_data += 1
corrects += do_correct(o, spiked_label)
return corrects.float() / total_data
def validate(data, model, data_encoder, label_encoder, loss, opt):
total_data = 0
corrects = 0
total_loss = 0
for i in range(data['test.data'].size(0)):
model.initialize(delay=False)
input = data['test.data'][i]
label = data['test.target'][i]
spiked_input = data_encoder.run(input)
spiked_input = torch.cat((spiked_input.view(-1), torch.zeros(2)))
spiked_label = label_encoder.run(label)
for t in range(opt.num_steps):
model(torch.tensor(t).float(), spiked_input)
o = model.fc2.s
total_data += 1
corrects += do_correct(o, spiked_label)
total_loss += loss(o, spiked_label)
avg_acc = corrects.float() / total_data
avg_loss = total_loss / total_data
return avg_loss, avg_acc
def app(opt):
np.set_printoptions(threshold=np.inf)
print(opt)
data_loader = n3ml.data.IRISDataLoader(ratio=0.8)
data = data_loader.run()
summary = data_loader.summarize()
data_encoder = n3ml.encoder.Population(neurons=12,
minimum=summary['min'],
maximum=summary['max'],
max_firing_time=opt.max_firing_time,
not_to_fire=opt.not_to_fire,
dt=opt.dt)
label_encoder = LabelEncoder(opt.num_classes)
model = n3ml.model.Bohte2002()
model.initialize()
optimizer = n3ml.optimizer.Bohte()
# for plot
plotter = Plot()
meter = {'total_losses': 0.0, 'num_corrects': 0, 'num_images': 0}
acc_buffer = []
loss_buffer = []
for epoch in range(opt.num_epochs):
train(data, model, data_encoder, label_encoder, optimizer, rmse, epoch,
meter, acc_buffer, loss_buffer, plotter, opt)
# print("epoch: {} - tr. loss: {} - tr. acc: {}".format(epoch, loss, acc))
loss, acc = validate(data, model, data_encoder, label_encoder, rmse,
opt)
print("epoch: {} - val. loss: {} - val. acc: {}".format(
epoch, loss, acc))
data = data_loader.run()
def train(data, model, data_encoder, label_encoder, optimizer, loss, epoch,
meter, acc_buffer, loss_buffer, plotter, opt):
for i in range(data['train.data'].size(0)):
model.initialize(delay=False)
input = data['train.data'][i]
label = data['train.target'][i]
spiked_input = data_encoder.run(input)
spiked_input = torch.cat((spiked_input.view(-1), torch.zeros(2)))
spiked_label = label_encoder.run(label)
for t in range(opt.num_steps):
model(torch.tensor(t).float(), spiked_input)
o = model.fc2.s
# print(model.fc1.s)
# print(model.fc2.s)
# print("pred: {} - target: {}".format(o, spiked_label))
# l = loss(o, spiked_label)
# print("loss: {}".format(l))
optimizer.step(model, spiked_input, spiked_label, epoch)
meter['num_images'] += 1
meter['num_corrects'] += do_correct(o, spiked_label)
meter['total_losses'] += loss(o, spiked_label)
if (i + 1) % 30 == 0:
print("label: {} - target: {} - pred: {} - result: {}".format(
label, spiked_label, o, do_correct(o, spiked_label)))
acc_buffer.append(1.0 * meter['num_corrects'] /
meter['num_images'])
loss_buffer.append(meter['total_losses'] / meter['num_images'])
plotter.update(y1=np.array(acc_buffer), y2=np.array(loss_buffer))