def get_error_history(dataset, add_jitter, encoder_dims, max_encoder_rate,
max_decoder_rate):
ripcoder = RIPCoder(encoder_dims, max_encoder_rate, max_decoder_rate)
optimizer = torch.optim.SGD(ripcoder.parameters(), lr=0.01)
mse = nn.MSELoss()
error_hist = list()
for i in range(10000):
error = torch.zeros(1, requires_grad=True)
for j in range(len(dataset)):
if add_jitter:
x = jitter(dataset[j])
else:
x = dataset[j]
# print(x)
x_hat, code = ripcoder.forward(x)
x_hat_polar = py_STDP.cart_to_polar(x_hat)
x_polar = py_STDP.cart_to_polar(x)
error = error + mse.forward(x_hat_polar[0], x_polar[0])
if i % 100 == 0:
print(' {} - ERROR:{:.5f}'.format(i, error.item()))
error_hist.append(error.item())
optimizer.zero_grad()
error.backward(retain_graph=True)
optimizer.step()
return error_hist, ripcoder
def jitter(complex_tensor):
tensor_polar = py_STDP.cart_to_polar(complex_tensor)
rate = tensor_polar[0]
phase = tensor_polar[1]
jitter_tensor = py_STDP.polar_to_cart(
(rate, phase + torch.rand_like(phase) / 20))
return jitter_tensor
def run_sim(index, arg_dict):
N = arg_dict['N']
f_max = arg_dict['f_max']
T = arg_dict['T']
nbins = arg_dict['nbins']
homeostasis = arg_dict['homeostasis']
sparsity = arg_dict['sparsity']
a_param = arg_dict['a_param']
stdp_weight = arg_dict['stdp_weight']
hebb_weight = arg_dict['hebb_weight']
args = inspect.getfullargspec(run_sim)
arg_strings = ['{}={}'.format(arg, eval(arg)) for arg in args]
desc_string = 'index={},'.format(index) + ','.join(arg_strings)
network = py_STDP.RIPLayer(N, N, f_max, homeostasis, sparsity)
x_polar = (f_max * torch.ones(1, N), torch.rand(1, N))
hist_vects = list()
entropy_traj = list()
x = py_STDP.polar_to_cart(x_polar)
for i in range(T + 1):
if i % 100 == 0:
print(' {}'.format(i))
xp = py_STDP.cart_to_polar(x)
hist_vect = np.histogram(xp[1].detach(),
np.linspace(0, 1, nbins + 1),
density=True)[0]
p = hist_vect / sum(hist_vect)
entropy_traj.append(entropy(p))
hist_vects.append(torch.tensor(hist_vect).unsqueeze(0))
y = network.forward(x)
network.rip_learn(x,
y,
a_param,
stdp_weight=stdp_weight,
hebb_weight=hebb_weight)
x = y
dist_hist = torch.cat(hist_vects, dim=0).transpose(0, 1)
return dist_hist, entropy_traj, desc_string
def run_sim(N, f_max, T, nbins):
network = py_STDP.RIPLayer(N, N, f_max, True)
x_polar = (f_max * torch.ones(1, N), torch.rand(1, N))
hist_vects = list()
entropy_traj = list()
x = py_STDP.polar_to_cart(x_polar)
for i in range(T + 1):
if i % 100 == 0:
print(' {}'.format(i))
xp = py_STDP.cart_to_polar(x)
hist_vect = np.histogram(xp[1].detach(),
np.linspace(0, 1, nbins + 1),
density=True)[0]
p = hist_vect / sum(hist_vect)
entropy_traj.append(entropy(p))
hist_vects.append(torch.tensor(hist_vect).unsqueeze(0))
y = network.forward(x)
network.rip_learn(x, y, 4, stdp_weight=100, hebb_weight=10)
x = y
dist_hist = torch.cat(hist_vects, dim=0).transpose(0, 1)
return dist_hist, entropy_traj
N = 100
T = 10000
nbins = 25
f_max = 1
network = py_STDP.RIPLayer(N, N, f_max, True)
x_polar = (f_max * torch.ones(1, N), torch.rand(1, N))
# plt.hist(x_polar[1].detach())
hist_vects_1 = list()
entropy_1 = list()
x = py_STDP.polar_to_cart(x_polar)
for i in range(T + 1):
if i % 100 == 0:
print(i)
xp = py_STDP.cart_to_polar(x)
hist_vect = np.histogram(xp[1].detach(),
np.linspace(0, 1, nbins + 1),
density=True)[0]
p = hist_vect / sum(hist_vect)
entropy_1.append(entropy(p))
# hist_vect = plt.hist(xp[1].detach(), np.linspace(0, 1, nbins + 1), density=True)
hist_vects_1.append(torch.tensor(hist_vect).unsqueeze(0))
x = network.forward(x)
hist_hist_1 = torch.cat(hist_vects_1, dim=0).transpose(0, 1)
# plt.clf()
# plt.subplot(2,1,1)
hist_vects_2 = list()
entropy_2 = list()
x = py_STDP.polar_to_cart(x_polar)