def test_SCCircuit():
EPS = 1e-12
DTYPE = tf.float64
np.random.seed(0)
M = 100
sess = tf.Session()
# true_sys = v1_circuit(T=T, dt=dt)
Z = tf.placeholder(dtype=DTYPE, shape=(1, M, None))
p = 0.7
std = 0.05
var = std**2
inact_str = "NI"
N = 500
param_dict = {"behavior_type": "WTA",
"p": p,
"var": var,
"inact_str": inact_str,
"N":N}
system = get_system_from_template("SCCircuit", param_dict)
print(system.behavior_str)
assert (
system.behavior_str
== "WTA_mu=7.00E-01_7.00E-01_2.50E-03_2.50E-03_0.00E+00_0.00E+00_1.00E+00_1.00E+00"
)
assert system.D == 8
assert system.num_suff_stats == 8
# Test simulation
_Z = np.random.normal(0.0, 3.0, (1, M, system.D))
r_t = system.simulate(Z)
_r_t = sess.run(r_t, {Z: _Z})
true_sys = sc_circuit()
_sW_P = _Z[0, :, 0]
_sW_A = _Z[0, :, 1]
_vW_PA = _Z[0, :, 2]
_vW_AP = _Z[0, :, 3]
_dW_PA = _Z[0, :, 4]
_dW_AP = _Z[0, :, 5]
_hW_P = _Z[0, :, 6]
_hW_A = _Z[0, :, 7]
W = np.array(
[
[_sW_P, _vW_PA, _dW_PA, _hW_P],
[_vW_AP, _sW_A, _hW_A, _dW_AP],
[_dW_AP, _hW_A, _sW_A, _vW_AP],
[_hW_P, _dW_PA, _vW_PA, _sW_P],
]
)
W = np.transpose(W, [2, 0, 1])
E_constant = 0.0
E_Pbias = 0.0
E_Prule = 10.0
E_Arule = 10.0
E_choice = 2.0
E_light = 1.0
Es = [E_constant, E_Pbias, E_Prule, E_Arule, E_choice, E_light]
_w = system.w
N = _w.shape[4]
eta_NI = np.ones((system.T,))
r_t_true = np.zeros((2, system.T, M, 4, system.N))
for i in range(M):
for j in range(N):
_r_t_true_ij = true_sys.simulate(W[i], Es, _w[:, 0, 0, :, j], eta_NI)
r_t_true[:, :, i, :, j] = _r_t_true_ij
r_t_true = np.transpose(r_t_true, [1, 0, 2, 3, 4])
assert approx_equal(_r_t, r_t_true, EPS)
return None
sysname = "SCCircuit"
dir_str = 'SC_WTA_%s' % inact_str
std = 0.15
var = std**2
N = 500
param_dict = {
"behavior_type": "WTA",
"p": p,
"var": var,
"inact_str": inact_str,
"N": N,
}
system = get_system_from_template(sysname, param_dict)
# Get DSN architecture
repeats = 1
nlayers = 2
arch_params = {
'D': system.D,
'repeats': repeats,
'nlayers': nlayers,
'upl': upl,
'sigma_init': sigma_init,
}
param_dict.update(arch_params)
arch_dict = get_arch_from_template(system, param_dict)
def test_V1Circuit():
os.chdir('dsn/')
np.random.seed(0)
M = 100
sysname = 'V1Circuit'
T = 100
dt = 0.005
W_allen = read_W('allen')
# Test rates behavior
#TODO at some point this should update to incorporate s into the input function
behavior_type = 'rates'
fac = 1.0
s = [5]
C = len(s)
param_dict = {'behavior_type': behavior_type, 'fac': fac, 's': s}
system = get_system_from_template(sysname, param_dict)
assert (len(system.z_labels) == system.D)
# Test parameter setting
Z = tf.placeholder(tf.float64, (1, None, system.D))
W, b, h_FF, h_LAT, h_RUN, tau, n, s_0, a, c_50 = system.filter_Z(Z)
_Z = np.zeros((1, M, system.D))
Z_a, Z_b = system.density_network_bounds
for i in range(system.D):
_Z[0, :, i] = np.random.uniform(Z_a[i], Z_b[i], (M, ))
tf_vars = [W, b, h_FF, h_LAT, h_RUN, tau, n]
with tf.Session() as sess:
_tf_vars = sess.run(tf_vars, {Z: _Z})
_W, _b, _h_FF, _h_LAT, _h_RUN, _tau, _n = _tf_vars
for i in range(M):
assert (approx_equal(np.abs(_W[0, i, :, :]), W_allen, EPS))
assert (approx_equal(_b[0, i, :, 0], _Z[0, i, :4], EPS))
assert (approx_equal(_n[0, i, 0], _Z[0, i, 4], EPS))
assert (approx_equal(_h_FF, 0.0, EPS))
assert (approx_equal(_h_LAT, 0.0, EPS))
assert (approx_equal(_h_RUN, 0.0, EPS))
assert (approx_equal(_tau, 0.02, EPS))
# Test input calculation
h = system.compute_h(b, h_FF, h_LAT, h_RUN, s_0)
with tf.Session() as sess:
_h = sess.run(h, {Z: _Z})
assert (approx_equal(_h[0, :, :, 0], _Z[0, :, :4], EPS))
# Test suff_stat calculation
r_t = system.simulate(Z)
T_x = system.compute_suff_stats(Z)
with tf.Session() as sess:
_r_t = sess.run(r_t, {Z: _Z})
_T_x = sess.run(T_x, {Z: _Z})
v1_circuit_true = v1_circuit(T=T, dt=dt, init_conds=system.init_conds)
r_ss = np.zeros((M, 4))
for i in range(M):
r_t_i = v1_circuit_true.simulate(_W[0, i, :, :], _h[0, i, :, :],
_tau[0, i, 0, 0], _n[0, i, 0, 0])
r_ss[i, :] = r_t_i[-1, :]
#assert(approx_equal(_r_t[:,0,i,:,0], r_t_i, 1e-6)) # should make exact
mean_true = r_ss
var_true = np.square(r_ss - np.expand_dims(system.mu[:4], 0))
T_x_true = np.concatenate((mean_true, var_true), axis=1)
assert (approx_equal(_T_x, np.expand_dims(T_x_true, 0), 1e-6))
# Check that mu matches data
# Local test. Cant put this data on github.
s_list = [5, [5], [10], np.array([60])]
fac_list = [1.0, 0.5, 100.0]
num_s = len(s_list)
num_fac = len(fac_list)
for i in range(num_s):
s = s_list[i]
means, stds = read_V1_responses(s)
for j in range(num_fac):
fac = fac_list[j]
means_true = fac * means
stds_true = fac * stds
vars_true = np.square(stds_true)
mu_true = np.concatenate((means_true, vars_true), axis=0)
param_dict = {'behavior_type': behavior_type, 'fac': fac, 's': s}
system = get_system_from_template(sysname, param_dict)
assert (len(system.T_x_labels) == system.num_suff_stats)
assert (system.mu.shape[0] == system.num_suff_stats)
approx_equal(mu_true, system.mu, EPS)
# Test differences behavior
#TODO at some point this should update to incorporate s into the input function
behavior_type = 'difference'
alphas = ['E', 'P', 'S', 'V']
num_alphas = len(alphas)
for k in range(num_alphas):
alpha = alphas[k]
param_dict = {
"behavior_type": behavior_type,
"alpha": alpha,
"inc_val": 0.5,
}
system = get_system_from_template(sysname, param_dict)
assert (len(system.z_labels) == system.D)
# Test parameter setting
Z = tf.placeholder(tf.float64, (1, None, system.D))
W, b, h_FF, h_LAT, h_RUN, tau, n, s_0, a, c_50 = system.filter_Z(Z)
_Z = np.zeros((1, M, system.D))
Z_a, Z_b = system.density_network_bounds
for i in range(system.D):
_Z[0, :, i] = np.random.uniform(Z_a[i], Z_b[i], (M, ))
tf_vars = [W, b, h_FF, h_LAT, h_RUN, tau, n]
with tf.Session() as sess:
_tf_vars = sess.run(tf_vars, {Z: _Z})
_W, _b, _h_FF, _h_LAT, _h_RUN, _tau, _n = _tf_vars
for i in range(M):
assert (approx_equal(np.abs(_W[0, i, :, :]), W_allen, EPS))
assert (approx_equal(_h_RUN[0, i, :, 0], _Z[0, i, :], EPS))
assert (approx_equal(_n, 2.0, EPS))
assert (approx_equal(_h_FF, 0.0, EPS))
assert (approx_equal(_h_LAT, 0.0, EPS))
assert (approx_equal(_b, 1.0, EPS))
assert (approx_equal(_tau, 0.02, EPS))
# Test input calculation
h = system.compute_h(b, h_FF, h_LAT, h_RUN, s_0)
with tf.Session() as sess:
_h = sess.run(h, {Z: _Z})
# Test suff_stat calculation
r_t = system.simulate(Z)
T_x = system.compute_suff_stats(Z)
with tf.Session() as sess:
_r_t = sess.run(r_t, {Z: _Z})
_T_x = sess.run(T_x, {Z: _Z})
v1_circuit_true = v1_circuit(T=T, dt=dt, init_conds=system.init_conds)
r_ss = np.zeros((M, 2))
for i in range(M):
r_t_i_NORUN = v1_circuit_true.simulate(_W[0, i, :, :], _h[0,
i, :, :],
_tau[0, i, 0, 0], _n[0, i,
0, 0])
r_t_i_RUN = v1_circuit_true.simulate(_W[0, i, :, :], _h[1,
i, :, :],
_tau[0, i, 0, 0], _n[0, i, 0,
0])
r_ss[i, 0] = r_t_i_NORUN[-1, k]
r_ss[i, 1] = r_t_i_RUN[-1, k]
#assert(approx_equal(_r_t[:,0,i,:,0], r_t_i, 1e-6)) # should make exact
dr_ss = r_ss[:, 1] - r_ss[:, 0]
mean_true = np.expand_dims(dr_ss, 1)
var_true = np.square(mean_true - system.mu[0])
T_x_true = np.concatenate((mean_true, var_true), axis=1)
assert (approx_equal(_T_x, np.expand_dims(T_x_true, 0), 1e-6))
return None
