def run_test(m,
output_dir_name,
show_connectivity=True,
repeats=1,
n_show_only=None,
add_noise=True,
dropouts=[{
'E': 0,
'I': 0
}]):
n_cells_driven = m.N_DRIVING_CELLS * m.PROJECTION_NUM
w_u_e_drive = np.zeros(m.N_DRIVING_CELLS)
w_u_e_drive[0] = m.W_U_E
w_u_e = np.repeat(np.diag(w_u_e_drive), m.PROJECTION_NUM, axis=0)
## input weights
w_u = {
# localized inputs to trigger activation from start of chain
'E':
np.block([
[w_u_e, np.zeros([n_cells_driven, m.N_EXC + m.N_INH])],
[
np.zeros([
m.N_EXC + m.N_INH - n_cells_driven,
m.N_EXC + m.N_INH + m.N_DRIVING_CELLS
])
],
]),
'I':
np.zeros((m.N_EXC + m.N_INH, m.N_DRIVING_CELLS + m.N_EXC + m.N_INH)),
'A':
np.zeros((m.N_EXC + m.N_INH, m.N_DRIVING_CELLS + m.N_EXC + m.N_INH)),
}
e_i_r = np.stack(
[rand_n_ones_in_vec_len_l(10, m.N_EXC) for i in range(m.N_INH)])
i_e_r = np.stack(
[rand_n_ones_in_vec_len_l(20, m.N_INH) for i in range(m.N_EXC)])
w_e_e_r = generate_chain_weight_mat(m)
w_e_e_r[:m.N_EXC, :m.N_EXC] += m.RAND_WEIGHT_MAX * np.random.rand(
m.N_EXC, m.N_EXC)
np.fill_diagonal(w_e_e_r, 0.)
connectivity = np.ones((m.N_EXC, m.N_EXC))
## recurrent weights
w_r_base = {
'E':
np.block([
[w_e_e_r, np.zeros((m.N_EXC, m.N_INH))],
[e_i_r * m.W_E_I_R,
np.zeros((m.N_INH, m.N_INH))],
]),
'I':
np.block([
[np.zeros((m.N_EXC, m.N_EXC)), i_e_r * m.W_I_E_R],
[np.zeros((m.N_INH, m.N_EXC)),
np.zeros((m.N_INH, m.N_INH))],
]),
'A':
np.block([
[
m.W_A * np.diag(np.ones((m.N_EXC))),
np.zeros((m.N_EXC, m.N_INH))
],
[np.zeros((m.N_INH, m.N_EXC)),
np.zeros((m.N_INH, m.N_INH))],
]),
}
w_r_for_dropouts = []
for dropout in dropouts:
w_r = copy(w_r_base)
# w_r['E'][:, :m.N_EXC] = dropout_on_mat(w_r['E'][:, :m.N_EXC], dropout['E'])
# w_r['I'][:, m.N_EXC:] = dropout_on_mat(w_r['I'][:, m.N_EXC:], dropout['I'])
w_r_for_dropouts.append(w_r)
# generate timesteps and initial excitatory input window
t = np.arange(0, S.T, S.DT)
all_rsps = []
# run simulation for same set of parameters
for rp_idx in range(repeats):
show_trial = (type(n_show_only) is int and rp_idx < n_show_only)
rsps_for_trial = []
## external currents
if add_noise:
i_ext = m.SGM_N / S.DT * np.random.randn(
len(t), m.N_EXC + m.N_INH) + m.I_EXT_B
else:
i_ext = m.I_EXT_B * np.ones((len(t), m.N_EXC + m.N_INH))
## inp spks
spks_u_base = np.zeros((len(t), m.N_DRIVING_CELLS + m.N_EXC + m.N_INH),
dtype=int)
# trigger inputs
activation_times = np.concatenate([
np.random.poisson(m.DRIVING_HZ * S.DT, size=(len(t), 1))
for i in range(m.N_DRIVING_CELLS)
],
axis=1)
spks_u = copy(spks_u_base)
spks_u[:, :m.N_DRIVING_CELLS] = np.zeros((len(t), m.N_DRIVING_CELLS))
burst_t = np.arange(0, 4 * int(m.BURST_T / S.DT),
int(m.BURST_T / S.DT))
for t_idx, driving_cell_idx in zip(*activation_times.nonzero()):
try:
spks_u[burst_t + t_idx, driving_cell_idx] = 1
except IndexError as e:
pass
rsps_for_trial = []
for d_idx, dropout in enumerate(dropouts):
w_r_for_dropout = w_r_for_dropouts[d_idx]
w_max = m.W_MAX / m.M
if show_connectivity:
graph_weights(copy(w_r_for_dropout), copy(w_u), v_max=w_max)
ntwk = LIFNtwkG(
c_m=m.C_M_E,
g_l=m.G_L_E,
e_l=m.E_L_E,
v_th=m.V_TH_E,
v_r=m.E_R_E,
t_r=m.T_R_E,
e_s={
'E': M.E_E,
'I': M.E_I,
'A': M.E_A
},
t_s={
'E': M.T_E,
'I': M.T_E,
'A': M.T_A
},
w_r=copy(w_r_for_dropout),
w_u=copy(w_u),
plasticity_indices=np.arange(m.N_EXC),
connectivity=connectivity,
W_max=m.W_MAX,
m=m.M,
eta=m.ETA,
epsilon=m.EPSILON,
dt=S.DT,
gamma=m.GAMMA,
alpha=m.ALPHA,
fr_set_points=m.FR_SET_POINTS,
output_freq=1000,
homeo=True,
)
clamp = Generic(v={0: np.repeat(m.E_L_E, m.N_EXC + m.N_INH)},
spk={})
# run smln
rsp = ntwk.run(
dt=S.DT,
clamp=clamp,
i_ext=i_ext,
output_dir_name=f'{output_dir_name}_{rp_idx}_{d_idx}',
spks_u=spks_u,
dropouts=[(m.DROPOUT_TIME, dropouts[d_idx])],
)
if show_connectivity:
graph_weights(rsp.ntwk.w_r, rsp.ntwk.w_u, v_max=w_max)
w_r_e = rsp.ntwk.w_r['E'][:m.N_EXC, :m.N_EXC]
graph_weight_matrix(w_r_e, 'Exc->Exc Weights\n', v_max=w_max)
graph_weight_matrix(np.sqrt(np.dot(w_r_e, w_r_e.T)),
'W * W.T \n',
v_max=w_max)
rsps_for_trial.append({
'spks_t':
copy(rsp.spks_t),
'spks_c':
copy(rsp.spks_c),
'spks_u':
spks_u.nonzero(),
'w_r':
copy(rsp.ntwk.w_r),
'activation_times':
activation_times.nonzero(),
})
all_rsps.append(rsps_for_trial)
return all_rsps
def run_test(m, output_dir_name, show_connectivity=True, repeats=1, n_show_only=None,
add_noise=True, dropouts=[{'E': 0, 'I': 0}], w_r_e=None, w_r_i=None, epochs=500):
output_dir = f'./figures/{output_dir_name}'
os.makedirs(output_dir)
robustness_output_dir = f'./robustness/{output_dir_name}'
os.makedirs(robustness_output_dir)
w_u_e = np.diag(np.ones(m.N_DRIVING_CELLS)) * m.W_U_E
## input weights
w_u = {
# localized inputs to trigger activation from start of chain
'E': np.block([
[ w_u_e, np.zeros([m.N_DRIVING_CELLS, m.N_EXC + m.N_SILENT + m.N_INH]) ],
[ np.zeros([m.N_EXC + m.N_SILENT + m.N_INH - m.N_DRIVING_CELLS, m.N_EXC + m.N_SILENT + m.N_INH + m.N_DRIVING_CELLS]) ],
]),
'I': np.zeros((m.N_EXC + m.N_SILENT + m.N_INH , m.N_DRIVING_CELLS + m.N_EXC + m.N_SILENT + m.N_INH)),
'A': np.zeros((m.N_EXC + m.N_SILENT + m.N_INH, m.N_DRIVING_CELLS + m.N_EXC + m.N_SILENT + m.N_INH)),
}
connectivity = np.ones((m.N_EXC, m.N_EXC))
def solid_unit_func():
return np.ones((m.PROJECTION_NUM, m.PROJECTION_NUM))
def rand_unit_func():
return np.random.rand(m.PROJECTION_NUM, m.PROJECTION_NUM)
if w_r_e is None:
w_e_e_r = generate_exc_ff_chain(m)
scaled_down_mask = np.random.rand(m.N_EXC + m.N_SILENT, 1) < m.SCALE_DOWN_PROB
scaled_down_amp = np.random.rand(m.N_EXC + m.N_SILENT, 1) * m.W_E_E_SCALE_DOWN_FACTOR
scaled_down_amp_e_i = np.random.rand(m.N_EXC + m.N_SILENT, 1) * m.W_E_I_SCALE_DOWN_FACTOR
w_e_e_r = np.where(scaled_down_mask, scaled_down_amp, 1) * w_e_e_r
np.fill_diagonal(w_e_e_r, 0.)
con_per_i = m.E_I_CON_PER_LINK * m.N_EXC / m.PROJECTION_NUM
e_i_r = rand_per_row_mat(int(con_per_i), (m.N_INH, m.N_EXC))
e_i_r = np.where(scaled_down_mask.T, scaled_down_amp_e_i.T, 1) * e_i_r
s_e_r = rand_per_row_mat(int(0.1 * m.N_SILENT), (m.N_EXC, m.N_SILENT))
w_r_e = np.block([
[ w_e_e_r, s_e_r * m.W_E_E_R / m.PROJECTION_NUM, np.zeros((m.N_EXC, m.N_INH)) ],
[ np.zeros((m.N_SILENT, m.N_EXC + m.N_SILENT + m.N_INH)) ],
[ e_i_r * m.W_E_I_R, np.zeros((m.N_INH, m.N_INH + m.N_SILENT)) ],
])
if w_r_i is None:
i_e_r = mat_1_if_under_val(m.I_E_CON_PROB, (m.N_EXC, m.N_INH))
w_r_i = np.block([
[ np.zeros((m.N_EXC, m.N_EXC + m.N_SILENT)), i_e_r * m.W_I_E_R ],
[ np.zeros((m.N_SILENT + m.N_INH, m.N_EXC + m.N_SILENT + m.N_INH)) ],
])
## recurrent weights
w_r = {
'E': w_r_e,
'I': w_r_i,
'A': np.block([
[ m.W_A * np.diag(np.ones((m.N_EXC))), np.zeros((m.N_EXC, m.N_SILENT + m.N_INH)) ],
[ np.zeros((m.N_SILENT + m.N_INH, m.N_EXC + m.N_SILENT + m.N_INH)) ],
]),
}
all_rsps = []
# run simulation for same set of parameters
for rp_idx in range(repeats):
show_trial = (type(n_show_only) is int and rp_idx < n_show_only)
rsps_for_trial = []
for d_idx, dropout in enumerate(dropouts):
e_cell_fr_setpoints = None
i_cell_fr_setpoints = None
e_cell_pop_fr_setpoint = None
active_cells_pre_dropout_mask = None
surviving_cell_indices = None
w_r_copy = copy(w_r)
for i_e in range(epochs):
if i_e == m.DROPOUT_ITER:
w_r_copy['E'][:, :(m.N_EXC + m.N_SILENT)], surviving_cell_indices = dropout_on_mat(w_r_copy['E'][:, :(m.N_EXC + m.N_SILENT)], dropout['E'], min_idx=m.DROPOUT_MIN_IDX, max_idx=m.DROPOUT_MAX_IDX)
t = np.arange(0, S.T, S.DT)
## external currents
if add_noise:
i_ext = m.SGM_N/S.DT * np.random.randn(len(t), m.N_EXC + m.N_SILENT + m.N_INH) + m.I_EXT_B
else:
i_ext = m.I_EXT_B * np.ones((len(t), m.N_EXC + m.N_SILENT + m.N_INH))
## inp spks
spks_u_base = np.zeros((len(t), m.N_DRIVING_CELLS + m.N_EXC + m.N_SILENT + m.N_INH), dtype=int)
# trigger inputs
activation_times = np.zeros((len(t), m.N_DRIVING_CELLS))
for t_ctr in np.arange(0, S.T, 1./m.DRIVING_HZ):
activation_times[int(t_ctr/S.DT), :] = 1
np.concatenate([np.random.poisson(m.DRIVING_HZ * S.DT, size=(len(t), 1)) for i in range(m.N_DRIVING_CELLS)], axis=1)
spks_u = copy(spks_u_base)
spks_u[:, :m.N_DRIVING_CELLS] = np.zeros((len(t), m.N_DRIVING_CELLS))
burst_t = np.arange(0, 5 * int(m.BURST_T / S.DT), int(m.BURST_T / S.DT))
for t_idx, driving_cell_idx in zip(*activation_times.nonzero()):
input_noise_t = np.array(np.random.normal(scale=m.INPUT_STD / S.DT), dtype=int)
try:
spks_u[burst_t + t_idx + input_noise_t, driving_cell_idx] = 1
except IndexError as e:
pass
def create_prop(prop_exc, prop_inh):
return cc([prop_exc * np.ones(m.N_EXC), prop_inh * np.ones(m.N_INH)])
c_m = create_prop(m.C_M_E, m.C_M_I)
g_l = create_prop(m.G_L_E, m.G_L_I)
e_l = create_prop(m.E_L_E, m.E_L_I)
v_th = create_prop(m.V_TH_E, m.V_TH_I)
e_r = create_prop(m.E_R_E, m.E_R_I)
t_r = create_prop(m.T_R_E, m.T_R_I)
ntwk = LIFNtwkG(
c_m=c_m,
g_l=g_l,
e_l=e_l,
v_th=v_th,
v_r=e_r,
t_r=t_r,
e_s={'E': M.E_E, 'I': M.E_I, 'A': M.E_A},
t_s={'E': M.T_E, 'I': M.T_E, 'A': M.T_A},
w_r=w_r_copy,
w_u=w_u,
plasticity_indices=np.arange(m.N_EXC),
connectivity=connectivity,
W_max=m.W_E_E_R_MAX,
m=m.M,
output=False,
output_freq=100000,
weight_update=False,
)
clamp = Generic(v={0: np.repeat(m.E_L_E, m.N_EXC + m.N_SILENT + m.N_INH)}, spk={})
# run smln
rsp = ntwk.run(dt=S.DT, clamp=clamp, i_ext=i_ext,
output_dir_name=f'{output_dir_name}_{rp_idx}_{d_idx}', spks_u=spks_u,
dropouts=[], m=m, repairs=[],
)
scale = 0.8
gs = gridspec.GridSpec(3, 1)
fig = plt.figure(figsize=(9 * scale, 9 * scale), tight_layout=True)
axs = [fig.add_subplot(gs[:2]), fig.add_subplot(gs[2])]
spks_for_e_cells = rsp.spks[:, :(m.N_EXC + m.N_SILENT)]
spks_for_i_cells = rsp.spks[:, (m.N_EXC + m.N_SILENT):(m.N_EXC + m.N_SILENT + m.N_INH)]
if surviving_cell_indices is not None:
spks_for_e_cells[:, ~(surviving_cell_indices.astype(bool))] = 0
spk_bins, freqs = bin_occurrences(spks_for_e_cells.sum(axis=0), max_val=200, bin_size=1)
if surviving_cell_indices is not None:
freqs[0] -= np.sum(np.where(~(surviving_cell_indices.astype(bool)), 1, 0))
axs[1].bar(spk_bins, freqs, alpha=0.5)
axs[1].set_xlabel('Spks per neuron')
axs[1].set_ylabel('Frequency')
axs[1].set_xlim(-0.5, 20.5)
axs[1].set_ylim(0, m.N_EXC + m.N_SILENT)
raster = np.stack([rsp.spks_t, rsp.spks_c])
inh_raster = raster[:, raster[1, :] > (m.N_EXC + m.N_SILENT)]
spk_bins_i, freqs_i = bin_occurrences(spks_for_i_cells.sum(axis=0), max_val=100, bin_size=1)
axs[1].bar(spk_bins_i, freqs_i, color='black', alpha=0.5)
if active_cells_pre_dropout_mask is not None:
exc_cells_initially_active = copy(spks_for_e_cells)
exc_cells_initially_active[:, ~active_cells_pre_dropout_mask] = 0
exc_cells_initially_active = np.stack(exc_cells_initially_active.nonzero())
exc_cells_newly_active = copy(spks_for_e_cells)
exc_cells_newly_active[:, active_cells_pre_dropout_mask] = 0
exc_cells_newly_active = np.stack(exc_cells_newly_active.nonzero())
axs[0].scatter(exc_cells_initially_active[0, :] * S.DT * 1000, exc_cells_initially_active[1, :], s=1, c='black', zorder=0, alpha=0.2)
axs[0].scatter(exc_cells_newly_active[0, :] * S.DT * 1000, exc_cells_newly_active[1, :], s=1, c='green', zorder=1, alpha=1)
else:
exc_raster = raster[:, raster[1, :] < (m.N_EXC + m.N_SILENT)]
axs[0].scatter(exc_raster[0, :] * 1000, exc_raster[1, :], s=1, c='black', zorder=0, alpha=1)
axs[0].scatter(inh_raster[0, :] * 1000, inh_raster[1, :], s=1, c='red', zorder=0, alpha=1)
axs[0].set_ylim(-1, m.N_EXC + m.N_INH)
axs[0].set_xlim(0, 0.15 * 1000)
axs[0].set_ylabel('Cell Index')
axs[0].set_xlabel('Time (ms)')
for i in range(len(axs)):
set_font_size(axs[i], 14)
fig.savefig(f'{output_dir}/{d_idx}_{zero_pad(i_e, 4)}.png')
first_spk_times = process_single_activation(exc_raster, m)
if i_e == 0:
sio.savemat(robustness_output_dir + '/' + f'title_{title}_dropout_{d_idx}_eidx_{zero_pad(i_e, 4)}', {
'first_spk_times': first_spk_times,
'w_r_e_summed': np.sum(rsp.ntwk.w_r['E'][:m.N_EXC, :m.N_EXC], axis=1),
'w_r_e_i_summed': np.sum(rsp.ntwk.w_r['E'][m.N_EXC:, :m.N_EXC], axis=1),
'spk_bins': spk_bins,
'freqs': freqs,
'exc_raster': exc_raster,
'inh_raster': inh_raster,
})
else:
if i_e >= m.DROPOUT_ITER:
spks_for_e_cells[:, ~surviving_cell_indices.astype(int)] = 0
# filter e cell spks for start of bursts
def burst_kernel(spks):
if spks.shape[0] > 1 and np.count_nonzero(spks[:-1]) > 0:
return 0
else:
return spks[-1]
filtered_spks_for_e_cells = np.zeros(spks_for_e_cells.shape)
t_steps_in_burst = int(4e-3/S.DT)
for i_c in range(spks_for_e_cells.shape[1]):
for i_t in range(spks_for_e_cells.shape[0]):
idx_filter_start = (i_t - t_steps_in_burst) if (i_t - t_steps_in_burst) > 0 else 0
idx_filter_end = (i_t + 1)
filtered_spks_for_e_cells[i_t, i_c] = burst_kernel(spks_for_e_cells[idx_filter_start: idx_filter_end, i_c])
stdp_burst_pair = 0
# E SINGLE-CELL FIRING RATE RULE
fr_update_e = 0
if i_e >= m.E_SINGLE_FR_TRIALS[0] and i_e < m.E_SINGLE_FR_TRIALS[1]:
if e_cell_fr_setpoints is None:
e_cell_fr_setpoints = np.sum(spks_for_e_cells > 0, axis=0)
else:
e_cell_fr_setpoints += np.sum(spks_for_e_cells > 0, axis=0)
elif i_e == m.E_SINGLE_FR_TRIALS[1]:
e_cell_fr_setpoints = e_cell_fr_setpoints / (m.E_SINGLE_FR_TRIALS[1] - m.E_SINGLE_FR_TRIALS[0])
where_fr_is_0 = (e_cell_fr_setpoints == 0)
if not m.SINGLE_CELL_FR_SYM:
e_cell_fr_setpoints[where_fr_is_0] = np.random.normal (
loc=m.SINGLE_CELL_FR_SETPOINT_MIN,
scale=m.SINGLE_CELL_FR_SETPOINT_MIN_STD,
size=e_cell_fr_setpoints[where_fr_is_0].shape[0]
)
elif i_e > m.E_SINGLE_FR_TRIALS[1]:
e_diffs = e_cell_fr_setpoints - np.sum(spks_for_e_cells > 0, axis=0)
# STDP FOR E CELLS: put in pairwise STDP on filtered_spks_for_e_cells
for i_t in range(spks_for_e_cells.shape[0]):
stdp_start = i_t - m.CUT_IDX_TAU_PAIR if i_t - m.CUT_IDX_TAU_PAIR > 0 else 0
stdp_spk_hist = filtered_spks_for_e_cells[stdp_start:i_t, :]
t_points_for_stdp = stdp_spk_hist.shape[0]
curr_spks = filtered_spks_for_e_cells[i_t, :]
if t_points_for_stdp > 0:
sparse_curr_spks = csc_matrix(curr_spks)
sparse_spks = csc_matrix(np.flip(stdp_spk_hist, axis=0))
# compute sparse pairwise correlations with curr_spks and spikes in stdp pairwise time window & dot into pairwise kernel
stdp_burst_pair += kron(curr_spks, sparse_spks).T.dot(m.KERNEL_PAIR[:t_points_for_stdp]).reshape(spks_for_e_cells.shape[1], spks_for_e_cells.shape[1])
else:
stdp_burst_pair += 0.
stdp_burst_pair[e_diffs <= 0, :] = 0
if not m.SINGLE_CELL_FR_SYM:
e_diffs[e_diffs >= 0] = 0
fr_update_e = e_diffs.reshape(e_diffs.shape[0], 1) * np.ones((m.N_EXC + m.N_SILENT, m.N_EXC + m.N_SILENT)).astype(float)
# E POPULATION-LEVEL FIRING RATE RULE
fr_pop_update = 0
if i_e >= m.POP_FR_TRIALS[0] and i_e < m.POP_FR_TRIALS[1]:
if e_cell_pop_fr_setpoint is None:
e_cell_pop_fr_setpoint = np.sum(spks_for_e_cells)
else:
e_cell_pop_fr_setpoint += np.sum(spks_for_e_cells)
elif i_e == m.POP_FR_TRIALS[1]:
e_cell_pop_fr_setpoint = e_cell_pop_fr_setpoint / (m.POP_FR_TRIALS[1] - m.POP_FR_TRIALS[0])
elif i_e > m.POP_FR_TRIALS[1]:
fr_pop_update = e_cell_pop_fr_setpoint - np.sum(spks_for_e_cells)
# I SINGLE-CELL FIRING RATE RULE
fr_update_i = 0
if i_e >= m.I_SINGLE_FR_TRIALS[0] and i_e < m.I_SINGLE_FR_TRIALS[1]:
if i_cell_fr_setpoints is None:
i_cell_fr_setpoints = np.sum(spks_for_i_cells > 0, axis=0)
else:
i_cell_fr_setpoints += np.sum(spks_for_i_cells > 0, axis=0)
elif i_e == m.I_SINGLE_FR_TRIALS[1]:
i_cell_fr_setpoints = i_cell_fr_setpoints / (m.I_SINGLE_FR_TRIALS[1] - m.I_SINGLE_FR_TRIALS[0])
elif i_e > m.I_SINGLE_FR_TRIALS[1]:
i_diffs = i_cell_fr_setpoints - np.sum(spks_for_i_cells > 0, axis=0)
i_diffs[(i_diffs <= 1) & (i_diffs >= -1)] = 0
fr_update_i = i_diffs.reshape(i_diffs.shape[0], 1) * np.ones((m.N_INH, m.N_EXC + m.N_SILENT)).astype(float)
print('fr_pop_update')
report_mat(fr_pop_update)
print('fr_update_i')
report_mat(fr_update_i)
e_total_potentiation = m.ETA * (m.ALPHA_1 * fr_update_e + m.BETA * stdp_burst_pair + m.GAMMA * fr_pop_update)
i_total_potentiation = m.ETA * (m.ALPHA_2 * fr_update_i)
if type(e_total_potentiation) is not float:
e_total_potentiation[:m.DROPOUT_MIN_IDX, :] = 0
e_total_potentiation[m.DROPOUT_MAX_IDX:, :] = 0
e_total_potentiation[:, :m.DROPOUT_MIN_IDX] = 0
e_total_potentiation[:, m.DROPOUT_MAX_IDX:] = 0
if type(i_total_potentiation) is not float:
try:
i_total_potentiation[:, :m.DROPOUT_MIN_IDX] = 0
i_total_potentiation[:, m.DROPOUT_MAX_IDX:] = 0
except TypeError as e:
breakpoint()
print('delta_e_fr')
mat_delta_e_fr = fr_update_e * w_r_copy['E'][:(m.N_EXC + m.N_SILENT), :(m.N_EXC + m.N_SILENT)]
report_mat(m.ALPHA_1 * mat_delta_e_fr)
print('delta_e_stdp')
mat_delta_e_stdp = stdp_burst_pair * w_r_copy['E'][:(m.N_EXC + m.N_SILENT), :(m.N_EXC + m.N_SILENT)]
report_mat(m.BETA * mat_delta_e_stdp)
w_r_copy['E'][:(m.N_EXC + m.N_SILENT), :(m.N_EXC + m.N_SILENT)] += (e_total_potentiation * w_r_copy['E'][:(m.N_EXC + m.N_SILENT), :(m.N_EXC + m.N_SILENT)])
w_r_copy['E'][(m.N_EXC + m.N_SILENT):, :(m.N_EXC + m.N_SILENT)] += (i_total_potentiation * w_r_copy['E'][(m.N_EXC + m.N_SILENT):, :(m.N_EXC + m.N_SILENT)])
w_r_copy['E'][w_r_copy['E'] < 0] = 0
w_e_e_hard_bound = m.W_E_E_R_MAX / m.PROJECTION_NUM
w_r_copy['E'][:(m.N_EXC + m.N_SILENT), :(m.N_EXC + m.N_SILENT)][w_r_copy['E'][:(m.N_EXC + m.N_SILENT), :(m.N_EXC + m.N_SILENT)] > w_e_e_hard_bound] = w_e_e_hard_bound
w_e_i_hard_bound = m.W_E_I_R_MAX
w_r_copy['E'][(m.N_EXC + m.N_SILENT):, :(m.N_EXC + m.N_SILENT)][w_r_copy['E'][(m.N_EXC + m.N_SILENT):, :(m.N_EXC + m.N_SILENT)] > m.W_E_I_R_MAX] = m.W_E_I_R_MAX
if i_e == m.DROPOUT_ITER - 1:
active_cells_pre_dropout_mask = np.where(spks_for_e_cells.sum(axis=0) > 0, True, False)
if i_e % 10 == 0:
if i_e < m.DROPOUT_ITER:
if i_e % 250 == 0:
sio.savemat(robustness_output_dir + '/' + f'title_{title}_dropout_{d_idx}_eidx_{zero_pad(i_e, 4)}', {
'first_spk_times': first_spk_times,
'w_r_e_summed': np.sum(rsp.ntwk.w_r['E'][:m.N_EXC, :m.N_EXC], axis=1),
'w_r_e_i_summed': np.sum(rsp.ntwk.w_r['E'][m.N_EXC:, :m.N_EXC], axis=1),
'w_r_e': rsp.ntwk.w_r['E'],
'w_r_i': rsp.ntwk.w_r['I'],
'spk_bins': spk_bins,
'freqs': freqs,
'exc_raster': exc_raster,
'inh_raster': inh_raster,
})
else:
sio.savemat(robustness_output_dir + '/' + f'title_{title}_dropout_{d_idx}_eidx_{zero_pad(i_e, 4)}', {
'first_spk_times': first_spk_times,
'w_r_e_summed': np.sum(rsp.ntwk.w_r['E'][:m.N_EXC, :m.N_EXC], axis=1),
'w_r_e_i_summed': np.sum(rsp.ntwk.w_r['E'][m.N_EXC:, :m.N_EXC], axis=1),
'spk_bins': spk_bins,
'freqs': freqs,
'exc_raster': exc_raster,
'inh_raster': inh_raster,
})
else:
if i_e % 250 == 0:
sio.savemat(robustness_output_dir + '/' + f'title_{title}_dropout_{d_idx}_eidx_{zero_pad(i_e, 4)}', {
'first_spk_times': first_spk_times,
'w_r_e_summed': np.sum(rsp.ntwk.w_r['E'][:m.N_EXC, :m.N_EXC], axis=1),
'w_r_e_i_summed': np.sum(rsp.ntwk.w_r['E'][m.N_EXC:, :m.N_EXC], axis=1),
'w_r_e': rsp.ntwk.w_r['E'],
'w_r_i': rsp.ntwk.w_r['I'],
'spk_bins': spk_bins,
'freqs': freqs,
'exc_cells_initially_active': exc_cells_initially_active,
'exc_cells_newly_active': exc_cells_newly_active,
'inh_raster': inh_raster,
'surviving_cell_indices': surviving_cell_indices,
})
else:
sio.savemat(robustness_output_dir + '/' + f'title_{title}_dropout_{d_idx}_eidx_{zero_pad(i_e, 4)}', {
'first_spk_times': first_spk_times,
'w_r_e_summed': np.sum(rsp.ntwk.w_r['E'][:m.N_EXC, :m.N_EXC], axis=1),
'w_r_e_i_summed': np.sum(rsp.ntwk.w_r['E'][m.N_EXC:, :m.N_EXC], axis=1),
'spk_bins': spk_bins,
'freqs': freqs,
'exc_cells_initially_active': exc_cells_initially_active,
'exc_cells_newly_active': exc_cells_newly_active,
'inh_raster': inh_raster,
'surviving_cell_indices': surviving_cell_indices,
})
