def gen_bound(n_train, n_test, base_param, k):
n_gate = 1
values = np.random.uniform(-k, +k, n_train)
ticks = np.random.uniform(0, 1, (n_train, n_gate)) < 0.01
train_data = generate_data(values, ticks)
values = smoothen(np.random.uniform(-1, +1, n_test))
ticks = np.random.uniform(0, 1, (n_test, n_gate)) < 0.01
test_data = generate_data(values, ticks, last=train_data["output"][-1])
return base_param, train_data, test_data
def gen_discrete(n_train, n_test, base_param, k):
n_gate = 1
values = np.random.uniform(-1, +1, n_train)
ticks = np.random.uniform(0, 1, (n_train, n_gate)) < 0.01
discrete_values = np.random.uniform(-1, 1, k)
idx = np.where(ticks == 1)[0]
values[idx] = np.random.choice(discrete_values, len(idx))
train_data = generate_data(values, ticks)
values = smoothen(np.random.uniform(-1, +1, n_test))
ticks = np.random.uniform(0, 1, (n_test, n_gate)) < 0.01
test_data = generate_data(values, ticks, last=train_data["output"][-1])
return base_param, train_data, test_data
def gen_gate(n_train, n_test, base_param, k):
n_gate = k
values = np.random.uniform(-1, +1, n_train)
ticks = np.random.uniform(0, 1, (n_train, n_gate)) < 0.01
train_data = generate_data(values, ticks)
values = smoothen(np.random.uniform(-1, +1, n_test))
ticks = np.random.uniform(0, 1, (n_test, n_gate)) < 0.01
test_data = generate_data(values, ticks, last=train_data["output"][-1])
param = base_param.copy()
param["shape"] = (1 + n_gate, param["shape"][1], n_gate)
return param, train_data, test_data
def gen_trigger(n_train, n_test, base_param, k):
n_gate = 1
values = np.random.uniform(-1, +1, n_train)
ticks_interval = np.random.randint(1, k + 1, size=(n_train))
ticks_time = np.cumsum(ticks_interval)
i_max = np.max(np.where(ticks_time < n_train)[0])
ticks = np.zeros((n_train, ))
ticks[ticks_time[:i_max]] = 1
train_data = generate_data(values, ticks)
values = smoothen(np.random.uniform(-1, +1, n_test))
ticks = np.random.uniform(0, 1, (n_test, n_gate)) < 0.01
test_data = generate_data(values, ticks, last=train_data["output"][-1])
return base_param, train_data, test_data
def gen_value(n_train, n_test, base_param, k):
n_gate = 1
values = np.random.uniform(-1, +1, (n_train, k))
ticks = np.random.uniform(0, 1, (n_train, n_gate)) < 0.01
train_data = generate_data(values, ticks)
values = np.empty((n_test, k))
for i in range(k):
values[:, i] = smoothen(np.random.uniform(-1, +1, n_test))
ticks = np.random.uniform(0, 1, (n_test, n_gate)) < 0.01
test_data = generate_data(values, ticks, last=train_data["output"][-1])
param = base_param.copy()
param["shape"] = (k + n_gate, param["shape"][1], n_gate)
return param, train_data, test_data
# Random generator initialization
np.random.seed(1)
# Build memoryticks
n_gate = 1
# Training data
n = 25000
values = np.random.uniform(-1, +1, n)
ticks = np.random.uniform(0, 1, (n, n_gate)) < 0.01
train_data = generate_data(values, ticks)
# Testing data
n = 2500
values = smoothen(np.random.uniform(-1, +1, n))
ticks = np.random.uniform(0, 1, (n, n_gate)) < 0.01
test_data = generate_data(values, ticks, last = train_data["output"][-1])
base_param = {"shape":(1+n_gate,1000,n_gate), "sparsity":0.5, "radius":0.1, "scaling":(1.0,1.0), "leak":1.0, "noise":(0, 1e-4, 0), "seed":None}
n_res = 20
n_sample = 20
noise_values = np.stack([np.zeros(n_sample), np.logspace(-8, 0, n_sample), np.zeros(n_sample)], axis = 1)
noise_test_errors = sharedmem.empty((len(noise_values),n_res))
noise_test_errors[...] = np.nan
radius_values = np.concatenate([[0], np.logspace(-2, 1, n_sample)])
radius_test_errors = sharedmem.empty((len(radius_values),n_res))
radius_test_errors[...] = np.nan
sparsity_values = np.logspace(-2, 0, n_sample)
model = generate_model(shape=(1 + n_gate, 1000, n_gate),
sparsity=0.5,
radius=0.1,
scaling=(1.0, 1.0),
leak=1.0,
noise=(0.0000, 0.0001, 0.0001))
# Training data
n = 25000 # 300000
values = np.random.uniform(-1, +1, n)
ticks = np.random.uniform(0, 1, (n, n_gate)) < 0.01
train_data = generate_data(values, ticks)
# Testing data
n = 2500
values = smoothen(np.random.uniform(-1, +1, n))
ticks = np.random.uniform(0, 1, (n, n_gate)) < 0.01
test_data = generate_data(values, ticks, last=train_data["output"][-1])
error = train_model(model, train_data)
print("Training error : {0}".format(error))
error = test_model(model, test_data)
print("Testing error : {0}".format(error))
np.save(files[0], test_data)
np.save(files[1], model["output"])
np.save(files[2], model["state"])
else:
test_data = np.load(files[0])
model = {}
model["output"] = np.load(files[1])
sparsity=0.5,
radius=0.1,
scaling=1.0,
leak=1.0,
noise=(0, 1e-4, 0))
# Sparsify W_in & W_fb
model["W_in"] *= np.random.uniform(0, 1, model["W_in"].shape) < sparsity
model["W_fb"] *= np.random.uniform(0, 1, model["W_fb"].shape) < sparsity
# Training data
n = 50000
values = np.random.uniform(-1, +1, (n, n_values))
for i in range(n_values):
values[:, i] = smoothen(values[:, i])
gates = np.random.uniform(0, 1, (n, n_gates)) < 0.01
train_data = generate_data(values, gates)
# Testing data
n = 2500
values = np.random.uniform(-1, +1, (n, n_values))
for i in range(n_values):
values[:, i] = smoothen(values[:, i])
gates = np.random.uniform(0, 1, (n, n_gates)) < 0.01
test_data = generate_data(values, gates, last=train_data["output"][-1])
print("Training")
rmse_train = train_model(model, train_data)
print("Training error : {0:.5f}".format(rmse_train))
