def load_decoder(self):
'''
Create a 'seed' decoder for the simulation which is simply randomly initialized
'''
from riglib.bmi import state_space_models
ssm = state_space_models.StateSpaceEndptVel2D()
self.decoder = train.rand_KFDecoder(ssm, self.encoder.get_units())
def test_zero_velocity_goal(self):
ssm = state_space_models.StateSpaceEndptVel2D()
goal_calc = goal_calculators.ZeroVelocityGoal(ssm)
target_pos = np.array([0, 0, 0], dtype=np.float64)
(goal_state, error), _ = goal_calc(target_pos)
self.assertTrue(np.array_equal(goal_state.ravel(), np.array([0, 0, 0, 0, 0, 0, 1])))
def make_FACosEnc(num):
from riglib.bmi import state_space_models as ssm
import pickle
num_neurons = 20;
SSM = ssm.StateSpaceEndptVel2D()
for n in range(num):
kwargs = {}
kwargs['n_neurons'] = num_neurons
C = np.random.rand(num_neurons, SSM.n_states)
enc = FACosEnc(C, SSM, return_ts=True, **kwargs)
enc.psi_unt_std[:4] /= 40
pickle.dump(enc, open('/storage/preeya/grom_data/sims/test_obs_vs_co_overlap/encoder_param_matched_'+str(n)+'.pkl', 'wb'))
def from_file_to_FACosEnc(plot=False):
from riglib.bmi import state_space_models as ssm
import pickle
import os
import matplotlib.pyplot as plt
dat = pickle.load(
open(
os.path.expandvars(
'/home/lab/preeya/fa_analysis/grom_data/co_obs_SNR_w_coefficients.pkl'
)))
SSM = ssm.StateSpaceEndptVel2D()
snr = {}
eps = 10**-10
if plot:
f, ax = plt.subplots(nrows=3, ncols=3)
for j, i in enumerate(np.sort(list(dat.keys()))):
snr[i] = []
d = dat[i]
kwargs = {}
kwargs['n_neurons'] = len(list(d.keys()))
C = np.random.rand(kwargs['n_neurons'], SSM.n_states)
kwargs['wt_sources'] = [1, 1, 0, 0]
enc = FACosEnc(C, SSM, return_ts=True, **kwargs)
for n in range(len(list(d.keys()))):
#For individual units:
enc.psi_tun[n, [3, 5, 6]] = d[n][3][0, :] #Terrible construction.
enc.mu[n] = 0
#Now set the standard deviation: Draw from VFB distribution of commands
data, enc = sim_enc(enc)
U = np.vstack((data['unt']))
T = np.vstack((data['tun']))
spk = U + T
vel = np.hstack((data['ctl']))[[3, 5], :].T
vel = np.hstack((np.array(vel), np.ones((len(vel), 1))))
# Fit encoder:
n_units = spk.shape[1]
snr_act = []
for n in range(n_units):
snr_des = d[n][2]
if np.isnan(snr_des):
snr_des = .3
print('sucdess')
snr_act.append(snr_des)
s2 = spk[:, n] #Spikes:
x = np.linalg.lstsq(
vel, s2[:, np.newaxis]) #Regress Spikes against Velocities
qnoise = np.var(s2[:, np.newaxis] - vel * np.mat(x[0])) #Residuals
#Explained Variance vs. Residual Variance:
qsig = np.var(vel * np.mat(x[0]))
k = qsig / (snr_des)
enc.psi_unt_std[n] = np.sqrt(k) # + eps
#Fit simulation:
data, enc = sim_enc(enc)
U = np.vstack((data['unt']))
T = np.vstack((data['tun']))
spk = U + T
vel = np.hstack((data['ctl']))[[3, 5], :].T
vel = np.hstack((np.array(vel), np.ones((len(vel), 1))))
snr_sim = []
for n in range(n_units):
s2 = spk[:, n] #Spikes:
x = np.linalg.lstsq(
vel, s2[:, np.newaxis]) #Regress Spikes against Velocities
qnoise = np.var(s2[:, np.newaxis] - vel * np.mat(x[0])) #Residuals
#Explained Variance vs. Residual Variance:
qsig = np.var(vel * np.mat(x[0]))
snr_sim.append(qsig / qnoise)
if plot:
axi = ax[j / 3, j % 3]
axi.plot(snr_sim, snr_act, '.')
axi.set_title(list(dat.keys())[j])
#kwargs['psi_unt_std'] = psi_unt_std
#kwargs['psi_tun'] = psi_tun
pickle.dump(
enc,
open(
os.path.expandvars(
'$FA_GROM_DATA/sims/test_obs_vs_co_overlap/encoder_param_matched_'
+ str(i) + '.pkl'), 'wb'))