def test(omegas=[0.5,0.0,1.0]):
'''
test reservoir on an unseen combination of omegas
'''
#predict
tot_in_scores = np.zeros([nT,n_neu, 2])
tot_out_scores = np.zeros([nT,n_neu, 2])
print "this gesture's omega ", omegas
inputs, outputs = win.run(nT, framerate=framerate, omegas = omegas)
d,e,f = np.shape(win.teach_signals)
teach_sign = np.zeros([nT,n_neu])
teach_sign[:,0:e*f] = np.reshape(win.teach_signals, [nT,e*f])
#scaling patching in 256
#from 6x3 input signals to matrix of nTx256 signals
#X = np.zeros([nT,n_neu])
#a,b,c = np.shape(win.inputs_signals)
#X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
# Convert input and output spikes to analog signals
inputs = convert_input(inputs)
X = L.ts2sig(timev, membrane, inputs[0][:,0], inputs[0][:,1], n_neu = 256)
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
zh = liquid.RC_predict (X,Y)
#pred_in_scores = liquid.RC_score(zh["input"], teach_sign)
#pred_out_scores = liquid.RC_score(zh["output"], teach_sign)
return X, Y, teach_sign, zh
def plot_svd():
inputs, outputs, time_exc = win.run(duration, framerate=framerate, neu_sync = nsync, speed=speed)
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
small_inputs = convert_input(inputs)
X = L.ts2sig(timev, membrane, small_inputs[0][:,0], small_inputs[0][:,1], n_neu = 256)
figure()
plot_inputs_small(small_inputs)
figure()
plot_inputs_small(outputs)
figure()
ac=np.mean(Y**2,axis=0)
aci=np.mean(X**2,axis=0)
max_pos = np.where(ac == np.max(ac))[0]
max_posi = np.where(aci == np.max(aci))[0]
subplot(3,1,1)
plot(X[:,max_posi])
xlabel('time (ds)')
ylabel('freq (Hz)')
subplot(3,1,2)
plot(Y[:,max_pos])
xlabel('time (ds)')
ylabel('freq (Hz)')
subplot(3,1,3)
CO = np.dot(Y.T,Y)
CI = np.dot(X.T,X)
si = np.linalg.svd(CI, full_matrices=True, compute_uv=False)
so = np.linalg.svd(CO, full_matrices=True, compute_uv=False)
semilogy(so/so[0], 'bo-', label="outputs")
semilogy(si/si[0], 'go-', label="inputs")
xlabel('Singular Value number')
ylabel('value')
legend(loc="best")
def test(omegas=[0.5, 0.0, 1.0]):
'''
test reservoir on an unseen combination of omegas
'''
#predict
tot_in_scores = np.zeros([nT, n_neu, 2])
tot_out_scores = np.zeros([nT, n_neu, 2])
print "this gesture's omega ", omegas
inputs, outputs = win.run(nT, framerate=framerate, omegas=omegas)
d, e, f = np.shape(win.teach_signals)
teach_sign = np.zeros([nT, n_neu])
teach_sign[:, 0:e * f] = np.reshape(win.teach_signals, [nT, e * f])
#scaling patching in 256
#from 6x3 input signals to matrix of nTx256 signals
#X = np.zeros([nT,n_neu])
#a,b,c = np.shape(win.inputs_signals)
#X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
# Convert input and output spikes to analog signals
inputs = convert_input(inputs)
X = L.ts2sig(timev, membrane, inputs[0][:, 0], inputs[0][:, 1], n_neu=256)
Y = L.ts2sig(timev,
membrane,
outputs[0][:, 0],
outputs[0][:, 1],
n_neu=256)
zh = liquid.RC_predict(X, Y)
#pred_in_scores = liquid.RC_score(zh["input"], teach_sign)
#pred_out_scores = liquid.RC_score(zh["output"], teach_sign)
return X, Y, teach_sign, zh
def test_determinism():
inputs, outputs = win.run(nT, framerate=framerate, neu_sync = nsync, omegas = omegas)
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
print "we are plotting outputs"
figure(fig_h.number)
for i in range(256):
subplot(16,16,i)
plot(Y[:,i])
axis('off')
def test_determinism(ntrials):
for i in range(ntrials):
inputs, outputs, time_exc = win.run(nT, framerate=framerate, neu_sync = nsync)
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
print "we are plotting outputs"
figure(fig_h.number)
for i in range(255):
subplot(16,16,i)
plot(Y[:,i])
axis('off')
def plot_svd():
inputs, outputs, time_exc = win.run(duration,
framerate=framerate,
neu_sync=nsync,
speed=speed)
Y = L.ts2sig(timev,
membrane,
outputs[0][:, 0],
outputs[0][:, 1],
n_neu=256)
small_inputs = convert_input(inputs)
X = L.ts2sig(timev,
membrane,
small_inputs[0][:, 0],
small_inputs[0][:, 1],
n_neu=256)
figure()
plot_inputs_small(small_inputs)
figure()
plot_inputs_small(outputs)
figure()
ac = np.mean(Y**2, axis=0)
aci = np.mean(X**2, axis=0)
max_pos = np.where(ac == np.max(ac))[0]
max_posi = np.where(aci == np.max(aci))[0]
subplot(3, 1, 1)
plot(X[:, max_posi])
xlabel('time (ds)')
ylabel('freq (Hz)')
subplot(3, 1, 2)
plot(Y[:, max_pos])
xlabel('time (ds)')
ylabel('freq (Hz)')
subplot(3, 1, 3)
CO = np.dot(Y.T, Y)
CI = np.dot(X.T, X)
si = np.linalg.svd(CI, full_matrices=True, compute_uv=False)
so = np.linalg.svd(CO, full_matrices=True, compute_uv=False)
semilogy(so / so[0], 'bo-', label="outputs")
semilogy(si / si[0], 'go-', label="inputs")
xlabel('Singular Value number')
ylabel('value')
legend(loc="best")
def test_determinism(ntrials):
ion()
fig_h = figure()
for i in range(ntrials):
inputs, outputs, time_exc = win.run(duration, framerate=framerate, neu_sync = nsync, speed=speed)
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
print "we are plotting outputs"
figure(fig_h.number)
for i in range(255):
subplot(16,16,i)
plot(Y[:,i])
axis('off')
def test(omegas=[0.5,0.0,1.0], teach_orth = True, ind = 99, this_t = 99):
'''
test reservoir on an unseen combination of omegas
'''
#predict
tot_in_scores = np.zeros([nT,n_neu, 2])
tot_out_scores = np.zeros([nT,n_neu, 2])
print "this gesture's omega ", omegas
inputs, outputs, eff_time = win.run(nT, framerate=framerate, omegas = omegas)
X = np.zeros([nT,n_neu])
a,b,c = np.shape(win.inputs_signals)
X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
if teach_orth:
teach_ortogonal_sign = L.orth_signal(X)#np.zeros([nT,n_neu])
teach_ortogonal = np.zeros([nT,n_neu])
for i in range(n_neu):
teach_ortogonal[:,i] = teach_ortogonal_sign
teach_sign = teach_ortogonal
else:
d,e,f = np.shape(win.teach_signals)
teach_sign = np.zeros([nT,n_neu])
teach_sign[:,0:e*f] = np.reshape(win.teach_signals, [nT,e*f])
#scaling patching in 256
#from 6x3 input signals to matrix of nTx256 signals
#X = np.zeros([nT,n_neu])
#a,b,c = np.shape(win.inputs_signals)
#X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
# Convert input and output spikes to analog signals
#inputs = convert_input(inputs)
#X = L.ts2sig(timev, membrane, inputs[0][:,0], inputs[0][:,1], n_neu = 256)
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
#pred_in_scores = liquid.RC_score(zh["input"], teach_sign)
#pred_out_scores = liquid.RC_score(zh["output"], teach_sign)
if save_data_to_disk:
np.savetxt("lsm_ret/omegas_gesture_"+str(ind)+"_trial_"+str(this_t)+".txt", omegas)
print "save data"
np.savetxt("lsm_ret/inputs_gesture_"+str(ind)+"_trial_"+str(this_t)+".txt", inputs[0])
np.savetxt("lsm_ret/outputs_gesture_"+str(ind)+"_trial_"+str(this_t)+".txt", outputs[0])
for i in range(e*f):
np.savetxt("lsm_ret/teaching_signals_gesture_"+str(ind)+"_teach_input_"+str(i)+"_trial_"+str(this_t)+".txt", teach_sign[:,i])
return X, Y, teach_sign
def test_determinism():
inputs, outputs, eff_time = win.run(nT, framerate=framerate, neu_sync = nsync, omegas = omegas)
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
#X = L.ts2sig(timev, membrane, inputs[0][:,0], inputs[0][:,1], n_neu = 128*128)
X = np.zeros([nT,n_neu])
a,b,c = np.shape(win.inputs_signals)
X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
print "we are plotting outputs"
figure(fig_h.number)
for i in range(256):
subplot(16,16,i)
plot(Y[:,i])
axis('off')
print "we are plotting svd and inputs/outputs"
plot_svd(fig_hh, X, Y)
def test_determinism():
inputs, outputs, eff_time = win.run(nT,
framerate=framerate,
neu_sync=nsync,
omegas=omegas)
Y = L.ts2sig(timev,
membrane,
outputs[0][:, 0],
outputs[0][:, 1],
n_neu=256)
#X = L.ts2sig(timev, membrane, inputs[0][:,0], inputs[0][:,1], n_neu = 128*128)
X = np.zeros([nT, n_neu])
a, b, c = np.shape(win.inputs_signals)
X[:, 0:b * c] = np.reshape(win.inputs_signals, [nT, b * c])
print "we are plotting outputs"
figure(fig_h.number)
for i in range(256):
subplot(16, 16, i)
plot(Y[:, i])
axis('off')
print "we are plotting svd and inputs/outputs"
plot_svd(fig_hh, X, Y)
def train(num_gestures,ntrials):
'''
train reservoir with retina inputs
'''
for ind in range(num_gestures):
print "sample omegas from a 3d sphere surface"
#N = np.random.randint(10000)
#v = np.random.uniform(size=(3,N))
#vn = v / np.sqrt(np.sum(v**2, 0))
#omegas = [vn[0,N-1], vn[1,N-1], vn[2,N-1]]
theta = np.linspace(0,2.0*pi,num_gestures)
phi = np.linspace(0,pi/10.0,350) #bigger smaller angle
T,P = np.meshgrid(theta,phi)
wX = sin(P)*cos(T)
wY = sin(P)*sin(T)
wZ = cos(P)
omegas = [wX[ntrials-1,ind], wY[ntrials-1,ind], wZ[ntrials-1,ind]]
#print "this gesture's omega ", omegas
#just poke it for the first time... to start in same configuration all the times..
inputs, outputs = win.run(np.round(nT/10), framerate=framerate, neu_sync = nsync, omegas = omegas)
for this_t in xrange(ntrials):
#omegas = [wX[ind,this_t]+0.5, wY[ind,this_t], wZ[ind,this_t]]
omegas = [wX[ind,this_t]+0.5, 0.0, 1.0]
print "this gesture's omega ", omegas
inputs, outputs = win.run(nT, framerate=framerate, neu_sync = nsync, omegas = omegas)
#from 6x3 input signals to matrix of nTx256 signals
#X = np.zeros([nT,n_neu])
#a,b,c = np.shape(win.inputs_signals)
#X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
d,e,f = np.shape(win.teach_signals)
teach_sign = np.zeros([nT,n_neu])
teach_sign[:,0:e*f] = np.reshape(win.teach_signals, [nT,e*f])
# Convert input and output spikes to analog signals
if real_time_learn:
#inputs = convert_input(inputs)
#X = L.ts2sig(timev, membrane, inputs[0][:,0], inputs[0][:,1], n_neu = 256)
X = np.zeros([nT,n_neu])
a,b,c = np.shape(win.inputs_signals)
X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
#learn.. what? tip or something ortogonal to it?
if teach_orth:
teach_ortogonal_sign = L.orth_signal(X)#np.zeros([nT,n_neu])
teach_ortogonal = np.zeros([nT,n_neu])
for i in range(n_neu):
teach_ortogonal[:,i] = teach_ortogonal_sign
teach_sign = teach_ortogonal
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1], n_neu = 256)
print "teaching non orthogonal stuff to the input signals..."
liquid._realtime_learn (X,Y,teach_sign)
print np.sum(liquid.CovMatrix['input']), np.sum(liquid.CovMatrix['output'])
#evaluate
zh = liquid.RC_predict (X,Y)
score_in = liquid.RC_score(zh["input"], teach_sign)
score_out = liquid.RC_score(zh["output"], teach_sign)
tot_scores_in.append(score_in[0:e*f,:])
tot_scores_out.append(score_out[0:e*f,:])
#print "we are scoring...", scores
if save_data_to_disk:
print "save data"
np.savetxt("lsm_ret/inputs_gesture_"+str(ind)+"_trial_"+str(this_t)+".txt", inputs[0])
np.savetxt("lsm_ret/outputs_gesture_"+str(ind)+"_trial_"+str(this_t)+".txt", outputs[0])
for i in range(e*f):
np.savetxt("lsm_ret/teaching_signals_gesture_"+str(ind)+"_teach_input_"+str(i)+"_trial_"+str(this_t)+".txt", teach_sign[:,i])
return X, Y, teach_sign, zh
def train(num_gestures, ntrials):
'''
train reservoir with retina inputs
'''
for ind in range(num_gestures):
print "sample omegas from a 3d sphere surface"
#N = np.random.randint(10000)
#v = np.random.uniform(size=(3,N))
#vn = v / np.sqrt(np.sum(v**2, 0))
#omegas = [vn[0,N-1], vn[1,N-1], vn[2,N-1]]
theta = np.linspace(0, 2.0 * pi, num_gestures)
phi = np.linspace(0, pi / 10.0, 350) #bigger smaller angle
T, P = np.meshgrid(theta, phi)
wX = sin(P) * cos(T)
wY = sin(P) * sin(T)
wZ = cos(P)
omegas = [
wX[ntrials - 1, ind], wY[ntrials - 1, ind], wZ[ntrials - 1, ind]
]
#print "this gesture's omega ", omegas
#just poke it for the first time... to start in same configuration all the times..
inputs, outputs = win.run(np.round(nT / 10),
framerate=framerate,
neu_sync=nsync,
omegas=omegas)
for this_t in xrange(ntrials):
#omegas = [wX[ind,this_t]+0.5, wY[ind,this_t], wZ[ind,this_t]]
omegas = [wX[ind, this_t] + 0.5, 0.0, 1.0]
print "this gesture's omega ", omegas
inputs, outputs = win.run(nT,
framerate=framerate,
neu_sync=nsync,
omegas=omegas)
#from 6x3 input signals to matrix of nTx256 signals
#X = np.zeros([nT,n_neu])
#a,b,c = np.shape(win.inputs_signals)
#X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
d, e, f = np.shape(win.teach_signals)
teach_sign = np.zeros([nT, n_neu])
teach_sign[:, 0:e * f] = np.reshape(win.teach_signals, [nT, e * f])
# Convert input and output spikes to analog signals
if real_time_learn:
#inputs = convert_input(inputs)
#X = L.ts2sig(timev, membrane, inputs[0][:,0], inputs[0][:,1], n_neu = 256)
X = np.zeros([nT, n_neu])
a, b, c = np.shape(win.inputs_signals)
X[:, 0:b * c] = np.reshape(win.inputs_signals, [nT, b * c])
#learn.. what? tip or something ortogonal to it?
if teach_orth:
teach_ortogonal_sign = L.orth_signal(
X) #np.zeros([nT,n_neu])
teach_ortogonal = np.zeros([nT, n_neu])
for i in range(n_neu):
teach_ortogonal[:, i] = teach_ortogonal_sign
teach_sign = teach_ortogonal
Y = L.ts2sig(timev,
membrane,
outputs[0][:, 0],
outputs[0][:, 1],
n_neu=256)
print "teaching non orthogonal stuff to the input signals..."
liquid._realtime_learn(X, Y, teach_sign)
print np.sum(liquid.CovMatrix['input']), np.sum(
liquid.CovMatrix['output'])
#evaluate
zh = liquid.RC_predict(X, Y)
score_in = liquid.RC_score(zh["input"], teach_sign)
score_out = liquid.RC_score(zh["output"], teach_sign)
tot_scores_in.append(score_in[0:e * f, :])
tot_scores_out.append(score_out[0:e * f, :])
#print "we are scoring...", scores
if save_data_to_disk:
print "save data"
np.savetxt(
"lsm_ret/inputs_gesture_" + str(ind) + "_trial_" +
str(this_t) + ".txt", inputs[0])
np.savetxt(
"lsm_ret/outputs_gesture_" + str(ind) + "_trial_" +
str(this_t) + ".txt", outputs[0])
for i in range(e * f):
np.savetxt(
"lsm_ret/teaching_signals_gesture_" + str(ind) +
"_teach_input_" + str(i) + "_trial_" + str(this_t) +
".txt", teach_sign[:, i])
return X, Y, teach_sign, zh
def test(omegas=[0.5, 0.0, 1.0], teach_orth=True, ind=99, this_t=99):
'''
test reservoir on an unseen combination of omegas
'''
#predict
tot_in_scores = np.zeros([nT, n_neu, 2])
tot_out_scores = np.zeros([nT, n_neu, 2])
print "this gesture's omega ", omegas
inputs, outputs, eff_time = win.run(nT, framerate=framerate, omegas=omegas)
X = np.zeros([nT, n_neu])
a, b, c = np.shape(win.inputs_signals)
X[:, 0:b * c] = np.reshape(win.inputs_signals, [nT, b * c])
if teach_orth:
teach_ortogonal_sign = L.orth_signal(X) #np.zeros([nT,n_neu])
teach_ortogonal = np.zeros([nT, n_neu])
for i in range(n_neu):
teach_ortogonal[:, i] = teach_ortogonal_sign
teach_sign = teach_ortogonal
else:
d, e, f = np.shape(win.teach_signals)
teach_sign = np.zeros([nT, n_neu])
teach_sign[:, 0:e * f] = np.reshape(win.teach_signals, [nT, e * f])
#scaling patching in 256
#from 6x3 input signals to matrix of nTx256 signals
#X = np.zeros([nT,n_neu])
#a,b,c = np.shape(win.inputs_signals)
#X[:,0:b*c] = np.reshape(win.inputs_signals, [nT,b*c])
# Convert input and output spikes to analog signals
#inputs = convert_input(inputs)
#X = L.ts2sig(timev, membrane, inputs[0][:,0], inputs[0][:,1], n_neu = 256)
Y = L.ts2sig(timev,
membrane,
outputs[0][:, 0],
outputs[0][:, 1],
n_neu=256)
#pred_in_scores = liquid.RC_score(zh["input"], teach_sign)
#pred_out_scores = liquid.RC_score(zh["output"], teach_sign)
if save_data_to_disk:
np.savetxt(
"lsm_ret/omegas_gesture_" + str(ind) + "_trial_" + str(this_t) +
".txt", omegas)
print "save data"
np.savetxt(
"lsm_ret/inputs_gesture_" + str(ind) + "_trial_" + str(this_t) +
".txt", inputs[0])
np.savetxt(
"lsm_ret/outputs_gesture_" + str(ind) + "_trial_" + str(this_t) +
".txt", outputs[0])
for i in range(e * f):
np.savetxt(
"lsm_ret/teaching_signals_gesture_" + str(ind) +
"_teach_input_" + str(i) + "_trial_" + str(this_t) + ".txt",
teach_sign[:, i])
return X, Y, teach_sign
figure()
signal_trace_good = [raw_data_times[index_ord],signal]
plot(signal_trace_good[0], signal_trace_good[1])
#X = L.ts2sig(timev, membrane, raw_data_times, raw_data_amp, n_neu = 256)
#X = np.loadtxt("bioamp/XXspikes.txt")
# Time vector for analog signals
Fs = 1000/1e3 # Sampling frequency (in kHz)
T = 12000
nT = np.round (Fs*T)
timev = np.linspace(0,T,nT)
#Conversion from spikes to analog
membrane = lambda t,ts: np.atleast_2d(np.exp((-(t-ts)**2)/(2*1**2)))
X = L.ts2sig(timev, membrane, raw_data_times, raw_data_amp, n_neu = 256)
figure()
for i in range(2):
plot(X[:,i])
#SVD
figure()
ac=np.mean(Y**2,axis=0)
aci=np.mean(X**2,axis=0)
max_pos = np.where(ac == np.max(ac))[0]
max_posi = np.where(aci == np.max(aci))[0]
subplot(3,1,1)
plot(X[:,max_posi])
subplot(3,1,2)
# c = c,
# duration=duration,
# delay_sync=delay_sync,
# max_freq= 2800, min_freq = 400)
#nsetup.chips['mn256r1'].load_parameters('biases/biases_reservoir.biases')
#time.sleep(0.2)
#stimulate
if not use_retina:
inputs, outputs = liquid.RC_poke(stimulus)
else:
inputs, outputs = win.run(counts, framerate=framerate)
# Convert input and output spikes to analog signals
X = L.ts2sig(timev, membrane, inputs[0][:,0], np.floor(inputs[0][:,1]))
Y = L.ts2sig(timev, membrane, outputs[0][:,0], outputs[0][:,1])
#if(learn_real_time == True):
# Calculate activity of current inputs.
# As of now the reservoir can only give answers during activity
tmp_ac = np.mean(func_avg(timev[:,None], outputs[0][:,0][None,:]), axis=1)
tmp_ac = tmp_ac / np.max(tmp_ac)
if (np.sum(tmp_ac) > np.sum(ac)) or (this_t==0):
ac = tmp_ac[:,None]
#teach_sig = L.orth_signal(X)*ac.T**4 #
#teach_sig = L.orth_signal(X)[None,:]
teach_sig = gesture_teach * ac**4 # Windowed by activity
#learn