def run():
def c2(c):
return np.array([[1., c], [c, 1.]])
def cov(n, w, dr=0):
distance = np.array([[((i - j + (n / 2.)) % n) - n / 2.
for i in range(n)] for j in range(n)])
gauss = np.exp(-(distance**2) / (2. * w * w)) - dr
out = np.dot(gauss, gauss)
return out / out[0, 0]
IN = np.array([[1., .1, .1]])
out = []
for s1 in np.linspace(.1, 5, 10):
for s2 in np.linspace(.1, 5, 10):
CC = s1 * cov(3, .1)
NC = s2 * cov(3, .1)
out.append([
GSM.gexp(0, IN, CC, NC, precom=False),
np.linalg.det(CC),
np.linalg.det(NC),
np.linalg.norm(CC),
np.linalg.norm(NC)
])
np.savetxt("./NCORtest.csv", out)
def make_GSM_resp(filt, quenched, K, mgs, tag, COR, ncor, n_trial=1):
#make sure everything is the right type
assert quenched in (True, False)
assert type(ncor) == np.ndarray
GV = []
GVnn = []
#generate samples for each filter set
if quenched:
nftemp = np.tile(filt, [n_trial, 1])
print(nftemp.shape)
noise = np.reshape(
np.random.multivariate_normal(np.zeros(np.prod(nftemp.shape[1:])),
ncor, n_trial * filt.shape[0]),
tuple([n_trial * filt.shape[0]]) + filt.shape[1:])
nftemp += noise
else:
nftemp = filt
#run it for each value of k
for k in K:
GV.append(
np.array([
MGSM.gexp(i, nftemp, COR, k * k * ncor, precom=True)
for i in range(filt.shape[1])
]).transpose())
GVnn.append(MGSM.gnn(filt, COR))
print(GV[-1].shape)
GV = np.concatenate(GV)
GVnn = np.concatenate(GVnn)
print(tag)
print(GV.shape)
print(GVnn.shape)
np.savetxt(tag + "_noisy.csv", GV)
np.savetxt(tag + "_clean.csv", GVnn)
def get_seg_probs(X, CC, NC, P):
print("segprob")
SEG1 = MGSM.NPShared(X, CC, NC) * P[0]
SEG21 = MGSM.NPShared(np.reshape(X[:, 0], [-1, 1]), np.array([[CC[0, 0]]]),
np.array([[NC[0, 0]]]))
SEG22 = MGSM.NPShared(np.reshape(X[:, 1], [-1, 1]), np.array([[CC[1, 1]]]),
np.array([[NC[1, 1]]]))
SEG2 = SEG21 * SEG22 * P[1]
NORM = SEG1 + SEG2
nnSEG1 = MGSM.LPShared(X, CC) * P[0]
nnSEG21 = MGSM.LPShared(np.reshape(X[:, 0], [-1, 1]), np.array([[CC[0,
0]]]))
nnSEG22 = MGSM.LPShared(np.reshape(X[:, 1], [-1, 1]), np.array([[CC[1,
1]]]))
nnSEG2 = nnSEG21 * nnSEG22 * P[1]
nnNORM = nnSEG1 + nnSEG2
O1 = np.reshape(np.array([SEG1 / NORM, SEG2 / NORM]), [2, -1])
O2 = np.reshape(np.array([nnSEG1 / nnNORM, nnSEG2 / nnNORM]), [2, -1])
print(np.max(O1))
print(np.max(nnSEG2))
print(np.min(O1))
print(np.min(nnSEG2))
return O1, O2
def run_KW(k, w):
resp1.append([])
resp2.append([])
resp3.append([])
nresp1.append([])
nresp2.append([])
nresp3.append([])
W = w
CC = cov(N, .6)
NC = cov(N, .6 * w)
NC = NC * np.linalg.norm(CC) / np.linalg.norm(NC)
NN = 0 * np.random.multivariate_normal(np.zeros(len(NC)), NC,
[3, NUM_CON])
stim1 = np.array([cc * I1 for cc in con])
stim2 = np.array([cc * I2 for cc in con])
stim3 = np.array([cc * I3 for cc in con])
resp1[-1].append(MGSM.gexp(0, stim1 + NN[0] / (k), CC, NC / (k * k)))
resp2[-1].append(MGSM.gexp(0, stim2 + NN[1] / (k), CC, NC / (k * k)))
resp3[-1].append(MGSM.gexp(0, stim3 + NN[2] / (k), CC, NC / (k * k)))
nresp1[-1].append(MGSM.gnn(stim1 + 0 * NN[0] / (k), CC)[:, 0])
nresp2[-1].append(MGSM.gnn(stim2 + 0 * NN[0] / (k), CC)[:, 0])
nresp3[-1].append(MGSM.gnn(stim3 + 0 * NN[0] / (k), CC)[:, 0])
def make_resp(filt,
model,
quenched,
K,
mgs,
tag,
COR,
ncor,
n_trial=1,
NOISE=True):
P = COR[0]
CNS = COR[1]
C1 = COR[2]
C2 = COR[3]
print(P.shape)
print(CNS.shape)
print(C1.shape)
print(C2.shape)
GVnn = MGSM.MGSM_gnn(filt, CNS, C1, C2, P, model)
return GVnn
def run():
def c2(c):
return np.array([[1., c], [c, 1.]])
def cov(n, w, dr=0):
distance = np.array([[((i - j + (n / 2.)) % n) - n / 2.
for i in range(n)] for j in range(n)])
gauss = np.exp(-(distance**2) / (2. * w * w)) - dr
out = np.dot(gauss, gauss)
return out / out[0, 0]
def inp(n, w, c1, c2, i):
stim1 = np.array([((j + (n / 2.)) % n) - (n / 2.) for j in range(n)])
stim2 = np.array([(((j - (i % n)) + (n / 2.)) % n) - (n / 2.)
for j in range(n)])
gauss1 = np.exp(-(stim1**2) / (2. * w * w))
gauss2 = np.exp(-(stim2**2) / (2. * w * w))
return c1 * gauss1 + c2 * gauss2
LO = []
HI = []
nnLO = []
nnHI = []
ip = np.array([[1, 0], [0, 1], [1, 1]])
for c in np.linspace(-.9, .9, 20):
cov = c2(c)
nc = np.identity(2) * (1 + c)
sHI = GSM.gexp(0, 20 * ip, cov, nc * (2 * 2), precom=True)
sLO = GSM.gexp(0, ip, cov, nc * (2 * 2), precom=False)
snnLO = GSM.gnn(ip, cov)[:, 0]
snnHI = GSM.gnn(20 * ip, cov)[:, 0]
HI.append([c, sHI[2] / (sHI[0])])
LO.append([c, sLO[2] / (sLO[0])])
nnHI.append([c, snnHI[2] / (snnHI[0])])
nnLO.append([c, snnLO[2] / (snnLO[0])])
HI = np.array(HI)
LO = np.array(LO)
nnHI = np.array(nnHI)
nnLO = np.array(nnLO)
plt.plot(HI[:, 0], [1 for x in HI], 'k--', linewidth=1)
plt.plot(HI[:, 0], HI[:, 1], 'k')
plt.plot(nnHI[:, 0], nnHI[:, 1], 'k--')
plt.plot(LO[:, 0], LO[:, 1], 'r')
plt.plot(nnLO[:, 0], nnLO[:, 1], 'r--')
plt.xlabel("Correlation")
plt.ylabel("Modulation Ratio")
plt.xlim(-1, 1)
plt.ylim(0, 2)
plt.xticks([-.5, 0, .5])
plt.tight_layout()
plt.savefig("./2DAIparam.pdf")
print("done with AI")
#what it is that we want to do here? We want to look at COS
con = 20 * np.logspace(-2, 0, 50)
NN = 2
WW = np.array([.15])
out1 = []
out2 = []
nnout1 = []
nnout2 = []
out12 = []
out22 = []
nnout12 = []
nnout22 = []
NCOV = np.identity(2)
k = .75
I1 = np.array([[c1, c1 / 10] for c1 in con])
I2 = np.array([[c1 + con[-1] / 10, con[-1] + c1 / 10] for c1 in con])
I12 = np.array([[c1, 0] for c1 in con])
I22 = np.array([[c1, con[-1]] for c1 in con])
print(I1.shape)
print(I2.shape)
CC = c2(.6)
NCOV = np.identity(2)
out1 = GSM.gexp(0, I1, CC, NCOV / (k * k), precom=False)
out2 = GSM.gexp(0, I2, CC, NCOV / (k * k), precom=False)
nnout1 = GSM.gnn(I1, CC).T
nnout2 = GSM.gnn(I2, CC).T
out12 = GSM.gexp(0, I12, CC, NCOV / (k * k), precom=False)
out22 = GSM.gexp(0, I22, CC, NCOV / (k * k), precom=False)
nnout12 = GSM.gnn(I12, CC).T
nnout22 = GSM.gnn(I22, CC).T
print(nnout2)
plt.figure()
plt.plot(con / con[-1], out1, 'r')
plt.plot(con / con[-1], out2, 'k')
plt.plot(con / con[-1], nnout1[0], 'r--')
plt.plot(con / con[-1], nnout2[0], 'k--')
plt.xlabel("contrast")
plt.ylabel("Respose")
# plt.yscale("log")
plt.xscale("log")
plt.tight_layout()
plt.savefig("./2Dssfig_0.pdf")
plt.figure()
plt.plot(con / con[-1], out12, 'r')
plt.plot(con / con[-1], out22, 'k')
plt.plot(con / con[-1], nnout12[0], 'r--')
plt.plot(con / con[-1], nnout22[0], 'k--')
plt.xlabel("contrast")
plt.ylabel("Respose")
# plt.yscale("log")
plt.xscale("log")
plt.tight_layout()
plt.savefig("./2Dssfig_1.pdf")
return out
elif t < 1.75 and TARGET:
out += np.array([1.,.5])
elif t > .125 and MASK:
out += np.array([.5,1.])
return out
def stimfunc(STIM,t,nt,MASK = True,TARGET = True):
if STIM == 1:
return np.array([S1(T,MASK,TARGET) for T in np.linspace(0,t,nt)])
elif STIM == 2:
return np.array([S2(T,MASK,TARGET) for T in np.linspace(0,t,nt)])
elif STIM == 3:
return np.array([S3(T,MASK,TARGET) for T in np.linspace(0,t,nt)])
nta = 80
tot = .4
ta = tot/(nta)
Fa = np.exp(-ta)*np.identity(2)
stima = np.array([10*stimfunc(s,.4,nta + 1,MASK = M[1],TARGET = M[0])[:k] for k in range(1,81,5) for s in [1,2,3] for M in [[True,False],[False,True],[True,True]]])
stimp = np.array([10*stimfunc(s,.4,nta + 1,MASK = M[1],TARGET = M[0]) for s in [1,2,3] for M in [[True,False],[False,True],[True,True]]])
cor = [.1]
resp = np.array([inference.att_gexp(0,np.array([s]),cov(c),cov(0),inference.Q_self_con(cov(c),Fa),Fa) for c in cor for s in stima])
np.savetxt("./att_2d_resp.csv",resp)
np.savetxt("./att_2d_stim.csv",np.reshape(stimp,[-1,2]))
def run_w():
#what it is that we want to do here? We want to look at COS
resp1 = []
resp2 = []
resp3 = []
nresp1 = []
nresp2 = []
nresp3 = []
NUM_CURVE = 20
NUM_CON = 50
CMAX = 1
CMIN = -2
N = 8
I1 = inp(N, .6, 1, 1, -1 + N / 2)
I2 = inp(N, .6, 1, 0, -1 + N / 2)
I3 = inp(N, .6, 0, 1, -1 + N / 2)
KK = np.linspace(.27, .75, NUM_CURVE)
con = np.logspace(CMIN, CMAX, NUM_CON)
if 0:
for k in KK:
print(k)
W = k
CC = cov(N, W)
NC = cov(N, W)
np.savetxt("./covariance_test.csv", CC)
stim1 = np.array(
[cc * I1 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim2 = np.array(
[cc * I2 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim3 = np.array(
[cc * I3 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
resp1.append(MGSM.gexp(0, stim1, CC, NC / (2 * 2), precom=False))
resp2.append(MGSM.gexp(0, stim2, CC, NC / (2 * 2), precom=False))
resp3.append(MGSM.gexp(0, stim3, CC, NC / (2 * 2), precom=False))
nresp1.append(MGSM.gnn(stim1, CC)[:, 0])
nresp2.append(MGSM.gnn(stim2, CC)[:, 0])
nresp3.append(MGSM.gnn(stim3, CC)[:, 0])
np.savetxt("./param_files/wresp1.csv", resp1)
np.savetxt("./param_files/wresp2.csv", resp2)
np.savetxt("./param_files/wresp3.csv", resp3)
np.savetxt("./param_files/wnnresp1.csv", nresp1)
np.savetxt("./param_files/wnnresp2.csv", nresp2)
np.savetxt("./param_files/wnnresp3.csv", nresp3)
else:
resp1 = np.loadtxt("./param_files/wresp1.csv")
resp2 = np.loadtxt("./param_files/wresp2.csv")
resp3 = np.loadtxt("./param_files/wresp3.csv")
nresp1 = np.loadtxt("./param_files/wnnresp1.csv")
nresp2 = np.loadtxt("./param_files/wnnresp2.csv")
nresp3 = np.loadtxt("./param_files/wnnresp3.csv")
resp1 = np.array(resp1)
resp2 = np.array(resp2)
resp3 = np.array(resp3)
nresp1 = np.array(nresp1)
nresp2 = np.array(nresp2)
nresp3 = np.array(nresp3)
AI = resp1 / (resp2 + resp3)
nAI = nresp1 / (nresp2 + nresp3)
fig = plt.figure()
ax = fig.add_subplot(111)
cc = con / np.max(con)
for k in range(0, len(AI), 2):
ax.plot(cc, AI[k], label=str(KK[k]))
ax.plot(cc, nAI[k], '--')
ax.plot(con / np.max(con), [1 for k in range(len(cc))], 'k--')
plt.xlabel("contrast")
plt.ylabel("A.I.")
plt.ylim(.5, 1.25)
plt.tight_layout()
fig.savefig("./paramfig_W.pdf")
MAX = np.array([[KK[k] * 180 / np.pi, np.max(AI[k])]
for k in range(len(AI))])
plt.figure()
plt.plot(MAX[:, 0], MAX[:, 1], "r")
plt.xscale('log')
plt.tight_layout()
plt.xlabel("w")
plt.ylabel("Max AI")
plt.savefig("./max_AI_by_W.pdf")
pts = []
for k in range(len(AI)):
pts.append([KK[k], get_trans(AI[k], con)])
pts = np.array(pts)
plt.figure()
plt.plot(pts[:, 0], pts[:, 1])
plt.xscale('log')
plt.yscale('log')
plt.tight_layout()
plt.xlabel("w")
plt.ylabel("Transition (con)")
plt.savefig("./trans_by_W.pdf")
plt.figure()
plt.plot(KK, AI[:, -6])
plt.tight_layout()
plt.title("con. = {}".format(KK[-6]))
plt.xlabel("w")
plt.ylabel("High Contrast AI")
plt.savefig("./HighAI_by_W.pdf")
def on_off_response(data,SNR,dt,nframes):
par = data["params"]
#I want to make a bunch of gratings in 6 orientations
#I want to simulate them many times, with noise, and record the full temporal response for a while(until stability?)
#lets go for 100 presentations of each of the 6 gratings.
#I need enough to do a numerical estimation of the mutual information
if par["segmentation"] != "gsm":
print("On off is only for GSM models!")
exit()
f_pos = model_tools.get_f_pos(par["filter_position"],par["filter_distance"]*np.max(par["wavelengths"]),par["n_surrounds"])
indices = np.concatenate([[[a,b,c] for a in range(par["n_angles"]) for b in range(len(par["wavelengths"])) for c in range(2)] for p in f_pos])
positions = np.concatenate([[p for a in range(par["n_angles"]) for b in range(len(par["wavelengths"])) for c in range(2)] for p in f_pos])
fullsize = int(5*max(par["wavelengths"]) + 2*np.max(f_pos))
minwav = np.min(par["wavelengths"])
out = []
grats = stim.make_grating(.5,0,par["wavelengths"][0],fullsize/2,fullsize)
print("Getting Coefficients")
#get filters
coeffs = make_data.get_filter_maps(grats,data["kernels"])
print("DT: {}".format(dt))
path = [coeffs.shape[-1]/2 for k in range(nframes)]
print(path)
rundat = make_data.sample_path(coeffs,path,indices,positions)/np.array([data["fac"]])
print(rundat.shape)
#add a few zeroes for before and after stimulus onset
Z = np.zeros([2] + list(rundat.shape[1:]))
#2 zeros before, 6 after
rundat = np.concatenate([Z,rundat,Z,Z,Z,Z,Z],axis = 0)
#get all filters
ind = [k for k in range(len(indices))]
feps = [[linalg.logm(m)/data["params"]["walk_dt"] for m in f] for f in data["F"]]
FF = [[np.float32(linalg.expm(dt * data["params"]["walk_dt"] * m)) for m in f] for f in feps]
#print(FF[0][0])
QQ = [[inference.Q_self_con(data["C"][k][m],FF[k][m]) for m in range(len(data["F"][k]))] for k in range(len(data["F"]))]
NC = [[m/(SNR**2) for m in f] for f in data["C"]]
runf = rundat
#run the response analysis
out = []
for k in range(1,runf.shape[0]):
print("{}\t{}".format(k,runf.shape))
responses = inference.general_MGSM_g_att(np.array([runf[:k]]),data["segs"],data["C"],NC,QQ,FF,data["P"],ind,stable = True,op = False)
out.append(responses)
return np.array(out)
def run():
#what it is that we want to do here? We want to look at COS
resp1 = []
resp2 = []
resp3 = []
nresp1 = []
nresp2 = []
nresp3 = []
NUM_CURVE = 20
NUM_CON = 50
CMAX = 2.5
CMIN = -2
N = 8
I1 = inp(N, float(N) / 5, 1, 1, -1 + N / 2)
I2 = inp(N, float(N) / 5, 1, 0, -1 + N / 2)
I3 = inp(N, float(N) / 5, 0, 1, -1 + N / 2)
KK = np.logspace(-1, 1, NUM_CURVE)
for k in KK:
print(k)
W = float(N) / (5)
CC = cov(N, W)
NC = CC
np.savetxt("./covariance_test.csv", CC)
stim1 = np.array([cc * I1 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim2 = np.array([cc * I2 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim3 = np.array([cc * I3 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
resp1.append(MGSM.gexp(0, stim1, CC, NC / (k * k), precom=False))
resp2.append(MGSM.gexp(0, stim2, CC, NC / (k * k), precom=False))
resp3.append(MGSM.gexp(0, stim3, CC, NC / (k * k), precom=False))
nresp1.append(MGSM.gnn(stim1, CC)[:, 0])
nresp2.append(MGSM.gnn(stim2, CC)[:, 0])
nresp3.append(MGSM.gnn(stim3, CC)[:, 0])
resp1 = np.array(resp1)
resp2 = np.array(resp2)
resp3 = np.array(resp3)
nresp1 = np.array(nresp1)
nresp2 = np.array(nresp2)
nresp3 = np.array(nresp3)
AI = resp1 / (resp2 + resp3)
nAI = nresp1 / (nresp2 + nresp3)
fig = plt.figure()
ax = fig.add_subplot(111)
ax.set_prop_cycle("color",
[cm(1. * i / NUM_CURVE) for i in range(NUM_CURVE)])
K = np.linspace(.5, 5, NUM_CURVE)
cc = np.logspace(CMIN, CMAX, NUM_CON) / (10**CMAX)
for k in range(0, len(AI), 1):
ax.plot(cc, AI[k], label=str(K[k]))
ax.plot(cc, nAI[k], '--')
handles, labels = ax.get_legend_handles_labels()
# ax.legend(bbox_to_anchor=(0., 1.02, 1., .102), loc=3,
# ncol=2, mode="expand", borderaxespad=0.)
ax.plot(cc, [1 for k in range(len(cc))], 'k--')
plt.xscale('log')
fig.savefig("./paramfig.pdf")
MAX = np.array([[KK[k], np.max(AI[k])] for k in range(len(AI))])
plt.figure()
plt.plot(MAX[:, 0], MAX[:, 1], "r")
plt.xscale('log')
plt.savefig("./max_AI_by_k.pdf")
pts = []
for k in range(len(AI)):
print(np.argmax(AI[k]))
for j in range(np.argmax(AI[k]), len(AI[k]) - 1):
if AI[k][j] >= 1 and AI[k][j + 1] <= 1:
pts.append([KK[k], cc[j]])
break
pts = np.array(pts)
plt.figure()
plt.plot(pts[:, 0], pts[:, 1])
plt.xscale('log')
plt.yscale('log')
plt.savefig("./trans_by_K.pdf")
if args["fexp"]:
feps = [[linalg.logm(m) / data["params"]["walk_dt"] for m in f]
for f in data["F"]]
FF = [[
np.float32(linalg.expm(args["dt"] * data["params"]["walk_dt"] * m))
for m in f
] for f in feps]
else:
FF = [[np.eye(len(m)) + args["dt"] * (m - np.eye(len(m))) for m in f]
for f in data["F"]]
#print(FF[0][0])
QQ = [[args["dt"] * m for m in f] for f in data["Q"]]
NC = [[m / (args["snr"]**2) for m in f] for f in data["C"]]
print(rundat.shape)
test_max = [
inference.find_GSM_pia_max(np.array(k), data["C"][0][0], NC[0][0],
QQ[0][0], FF[0][0], 0, np.inf, 10., 1.)
for k in rundat[0]
]
test_gapp = [
inference.att_egia(0, np.array(rundat[0][k]), test_max[k][0],
data["C"][0][0], NC[0][0], QQ[0][0], FF[0][0])
for k in range(len(test_max))
]
print(test_gapp)
print([k[0] for k in test_max])
def run_k():
#what it is that we want to do here? We want to look at COS
resp1 = []
resp2 = []
resp3 = []
nresp1 = []
nresp2 = []
nresp3 = []
NUM_CURVE = 10
NUM_CON = 50
NUM_W = 20
CMAX = 1.
CMIN = -2.
N = 8
I1 = inp(N, .6, 1, 1, -1 + N / 2)
I2 = inp(N, .6, 1, 0, -1 + N / 2)
I3 = inp(N, .6, 0, 1, -1 + N / 2)
KK = np.logspace(0, 1, NUM_CURVE)
con = np.logspace(CMIN, CMAX, NUM_CON)
WW = np.linspace(.1, 2, NUM_W)
for w in range(len(WW)):
resp1.append([])
resp2.append([])
resp3.append([])
nresp1.append([])
nresp2.append([])
nresp3.append([])
for k in KK:
print(k)
W = WW[w]
CC = cov(N, W)
NC = CC
np.savetxt("./covariance_test.csv", CC)
stim1 = np.array(
[cc * I1 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim2 = np.array(
[cc * I2 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim3 = np.array(
[cc * I3 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
resp1[w].append(MGSM.gexp(0, stim1, CC,
NC / (k * k))) #,precom = False))
resp2[w].append(MGSM.gexp(0, stim2, CC,
NC / (k * k))) #,precom = False))
resp3[w].append(MGSM.gexp(0, stim3, CC,
NC / (k * k))) #,precom = False))
nresp1[w].append(MGSM.gnn(stim1, CC)[:, 0])
nresp2[w].append(MGSM.gnn(stim2, CC)[:, 0])
nresp3[w].append(MGSM.gnn(stim3, CC)[:, 0])
resp1 = np.array(resp1)
resp2 = np.array(resp2)
resp3 = np.array(resp3)
nresp1 = np.array(nresp1)
nresp2 = np.array(nresp2)
nresp3 = np.array(nresp3)
AI = resp1 / (resp2 + resp3)
nAI = nresp1 / (nresp2 + nresp3)
fig = plt.figure()
ax = fig.add_subplot(111)
ax.set_prop_cycle("color",
[cm(1. * i / NUM_CURVE) for i in range(NUM_CURVE)])
K = np.linspace(.5, 5, NUM_CURVE)
cc = 100 * np.logspace(CMIN, CMAX, NUM_CON) / (10**CMAX)
for w in [1]:
for k in range(0, len(AI[w]), len(AI[w]) / 5):
ax.plot(100 * con / np.max(con), AI[w, k], label=str(K[k]))
ax.plot(100 * con / np.max(con), nAI[w, k], '--')
handles, labels = ax.get_legend_handles_labels()
# ax.legend(bbox_to_anchor=(0., 1.02, 1., .102), loc=3,
# ncol=2, mode="expand", borderaxespad=0.)
ax.plot(cc, [1 for k in range(len(cc))], 'k--')
plt.xlabel("Contrast")
plt.ylabel("A.I.")
plt.tight_layout()
fig.savefig("./paramfig_K.pdf")
fig = plt.figure()
ax = fig.add_subplot(111)
ax.set_prop_cycle("color",
[cm(1. * i / NUM_CURVE) for i in range(NUM_CURVE)])
WW = np.linspace(.1, 2, NUM_W)
cc = 100 * np.logspace(CMIN, CMAX, NUM_CON) / (10**CMAX)
for k in [5]:
for w in range(0, len(AI), len(AI) / 5):
ax.plot(100 * con / np.max(con), AI[w, k], label=str(WW[w]))
ax.plot(100 * con / np.max(con), nAI[w, k], '--')
handles, labels = ax.get_legend_handles_labels()
# ax.legend(bbox_to_anchor=(0., 1.02, 1., .102), loc=3,
# ncol=2, mode="expand", borderaxespad=0.)
ax.plot(cc, [1 for k in range(len(cc))], 'k--')
plt.xlim([0, 20])
plt.xlabel("Contrast")
plt.ylabel("A.I.")
plt.tight_layout()
fig.savefig("./paramfig_W.pdf")
plt.figure()
for w in [5]:
MAX = np.array([[1. / KK[k], np.max(AI[w, k])]
for k in range(len(AI[w]))])
plt.plot(MAX[:, 0], MAX[:, 1], colortab[0])
# plt.xscale('log')
plt.xlabel("1/SNR")
plt.ylabel("A.I.")
plt.tight_layout()
plt.savefig("./max_AI_by_k.pdf")
plt.figure()
for k in [5]:
MAX = np.array([[WW[w] * 180. / np.pi,
np.max(AI[w, k])] for w in range(len(AI))])
plt.plot(MAX[:, 0], MAX[:, 1], colortab[0])
plt.xlabel("Width (deg.)")
plt.ylabel("A.I.")
plt.tight_layout()
plt.savefig("./max_AI_by_W.pdf")
plt.figure()
for w in [5]:
pts = []
for k in range(len(AI[w])):
pts.append(
[1. / KK[k],
get_trans(AI[w, k], 100 * con / np.max(con))])
pts = np.array(pts)
plt.plot(pts[:, 0], pts[:, 1], colortab[0])
plt.xlabel("1/SNR")
plt.ylabel("Contrast")
plt.xscale('log')
plt.yscale('linear')
plt.tight_layout()
plt.savefig("./trans_by_K.pdf")
plt.figure()
for k in [5]:
pts = []
for w in range(len(AI)):
pts.append([
WW[w] * 180. / np.pi,
get_trans(AI[w, k], 100 * con / np.max(con))
])
pts = np.array(pts)
plt.plot(pts[:, 0], pts[:, 1], colortab[0])
plt.xlabel("Width (deg.)")
plt.ylabel("Contrast")
plt.yscale('linear')
plt.tight_layout()
plt.savefig("./trans_by_W.pdf")
def run_cov():
#what it is that we want to do here? We want to look at COS
resp1 = []
resp2 = []
resp3 = []
nresp1 = []
nresp2 = []
nresp3 = []
NUM_CURVE = 10
NUM_CON = 50
NUM_W = 10
CMAX = 1.
CMIN = -1.
N = 8
I1 = inp(N, .6, 1, 1, -1 + N / 2)
I2 = inp(N, .6, 1, 0, -1 + N / 2)
I3 = inp(N, .6, 0, 1, -1 + N / 2)
KK = np.logspace(0, 1, NUM_CURVE)
con = np.logspace(CMIN, CMAX, NUM_CON)
WW = [.1, .5, 1.]
for k in WW:
print(k)
CC = cov(N, .5)
if k == WW[0]:
NC = np.identity(N)
else:
NC = cov(N, k)
NC = NC * np.linalg.norm(CC) / (np.linalg.norm(NC) * 4)
sn = 1
stim1 = np.array([cc * I1 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim2 = np.array([cc * I2 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim3 = np.array([cc * I3 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
resp1.append(sn * MGSM.gexp(0, stim1 / sn, CC, NC, precom=False))
resp2.append(sn * MGSM.gexp(0, stim2 / sn, CC, NC, precom=False))
resp3.append(sn * MGSM.gexp(0, stim3 / sn, CC, NC, precom=False))
nresp1.append(MGSM.gnn(stim1, CC)[:, 0])
nresp2.append(MGSM.gnn(stim2, CC)[:, 0])
nresp3.append(MGSM.gnn(stim3, CC)[:, 0])
resp1 = np.array(resp1)
resp2 = np.array(resp2)
resp3 = np.array(resp3)
nresp1 = np.array(nresp1)
nresp2 = np.array(nresp2)
nresp3 = np.array(nresp3)
AI = resp1 / (resp2 + resp3)
nAI = nresp1 / (nresp2 + nresp3)
print(AI[-1])
fig = plt.figure()
ax = fig.add_subplot(111)
ax.set_prop_cycle("color",
[cm(1. * i / NUM_CURVE) for i in range(NUM_CURVE)])
K = np.linspace(.5, 5, NUM_CURVE)
cc = 100 * np.logspace(CMIN, CMAX, NUM_CON) / (10**CMAX)
for k in range(0, len(AI)):
ax.plot(100 * con / np.max(con), AI[k], colortab[k], label=str(K[k]))
ax.plot(100 * con / np.max(con), nAI[k], colortab[k] + '--')
handles, labels = ax.get_legend_handles_labels()
# ax.legend(bbox_to_anchor=(0., 1.02, 1., .102), loc=3,
# ncol=2, mode="expand", borderaxespad=0.)
ax.plot(cc, [1 for k in range(len(cc))], 'k--')
plt.xlabel("Contrast")
plt.ylabel("A.I.")
plt.tight_layout()
fig.savefig("./paramfig_NC.pdf")
grat = stim.make_SS_filters(.5,
0,
0,
0,
16,
0,
5 * 16,
int(np.max(np.linalg.norm(sites, axis=1))),
get_grat=True)
inp = np.array([get_grating_data(g, sites, RF) / fac for g in grat])
nc = CNS
f = .5 * np.identity(len(CNS))
q = inf.Q_self_con(CNS, f)
inp += np.random.multivariate_normal(np.zeros_like(inp[0]), nc, len(inp))
aI = np.tile(np.expand_dims(inp, 1), [1, 3, 1])
resp = inf.gnn(inp, CNS)
iresp = inf.att_gexp(0, aI, CNS, nc, q, f)
f = .75 * np.identity(len(CNS))
q = inf.Q_self_con(CNS, f)
aresp = inf.att_gexp(0, aI, CNS, nc, q, f)
print(oris)
def run_k():
#what it is that we want to do here? We want to look at COS
resp1 = []
resp2 = []
resp3 = []
nresp1 = []
nresp2 = []
nresp3 = []
NUM_CURVE = 10
NUM_CON = 50
CMAX = 1.
CMIN = -2.
N = 8
I1 = inp(N, .6, 1, 1, -1 + N / 2)
I2 = inp(N, .6, 1, 0, -1 + N / 2)
I3 = inp(N, .6, 0, 1, -1 + N / 2)
KK = np.logspace(0, 1, NUM_CURVE)
con = np.logspace(CMIN, CMAX, NUM_CON)
if 0:
for k in KK:
print(k)
W = .6
CC = cov(N, W)
NC = CC
np.savetxt("./covariance_test.csv", CC)
stim1 = np.array(
[cc * I1 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim2 = np.array(
[cc * I2 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
stim3 = np.array(
[cc * I3 for cc in np.logspace(CMIN, CMAX, NUM_CON)])
resp1.append(MGSM.gexp(0, stim1, CC, NC / (k * k), precom=False))
resp2.append(MGSM.gexp(0, stim2, CC, NC / (k * k), precom=False))
resp3.append(MGSM.gexp(0, stim3, CC, NC / (k * k), precom=False))
nresp1.append(MGSM.gnn(stim1, CC)[:, 0])
nresp2.append(MGSM.gnn(stim2, CC)[:, 0])
nresp3.append(MGSM.gnn(stim3, CC)[:, 0])
np.savetxt("./param_files/kresp1.csv", resp1)
np.savetxt("./param_files/kresp2.csv", resp2)
np.savetxt("./param_files/kresp3.csv", resp3)
np.savetxt("./param_files/knnresp1.csv", nresp1)
np.savetxt("./param_files/knnresp2.csv", nresp2)
np.savetxt("./param_files/knnresp3.csv", nresp3)
else:
resp1 = np.loadtxt("./param_files/kresp1.csv")
resp2 = np.loadtxt("./param_files/kresp2.csv")
resp3 = np.loadtxt("./param_files/kresp3.csv")
nresp1 = np.loadtxt("./param_files/knnresp1.csv")
nresp2 = np.loadtxt("./param_files/knnresp2.csv")
nresp3 = np.loadtxt("./param_files/knnresp3.csv")
resp1 = np.array(resp1)
resp2 = np.array(resp2)
resp3 = np.array(resp3)
nresp1 = np.array(nresp1)
nresp2 = np.array(nresp2)
nresp3 = np.array(nresp3)
AI = resp1 / (resp2 + resp3)
nAI = nresp1 / (nresp2 + nresp3)
fig = plt.figure()
ax = fig.add_subplot(111)
K = np.linspace(.5, 5, NUM_CURVE)
cc = np.logspace(CMIN, CMAX, NUM_CON) / (10**CMAX)
for k in range(0, len(AI), 2):
ax.plot(con / np.max(con), AI[k], label=str(K[k]))
ax.plot(con / np.max(con), nAI[k], '--')
ax.plot(cc, [1 for k in range(len(cc))], 'k--')
plt.tight_layout()
plt.xlabel("contrast")
plt.ylabel("AI")
plt.ylim(0, 1.5)
fig.savefig("./paramfig_K.pdf")
MAX = np.array([[1. / KK[k], np.max(AI[k])] for k in range(len(AI))])
plt.figure()
plt.plot(MAX[:, 0], MAX[:, 1], "r")
plt.xscale('log')
plt.tight_layout()
plt.xlabel("1/SNR")
plt.ylabel("Max AI")
plt.savefig("./max_AI_by_k.pdf")
plt.figure()
plt.plot(con / np.max(con), resp2[NUM_CURVE / 2, :], "r")
plt.plot(con / np.max(con), nresp2[NUM_CURVE / 2, :], "r--")
plt.xscale('log')
plt.tight_layout()
plt.xlabel("Contrast")
plt.ylabel("Response")
plt.savefig("./param_CRF.pdf")
pts = []
for k in range(len(AI)):
pts.append([KK[k], get_trans(AI[k], con)])
pts = np.array(pts)
plt.figure()
plt.plot(1. / pts[:, 0], pts[:, 1])
plt.xscale('log')
# plt.yscale('log')
plt.tight_layout()
plt.xlabel("1/SNR")
plt.ylabel("Transition (con.)")
plt.savefig("./trans_by_K.pdf")
import numpy as np
import GSM.MGSM_inference as MGSM
cov = np.array([[1, .1], [.1, 1]])
ncov = np.array([[1, 0], [0, 1]])
qcov = np.array([[.01, 0], [0, .01]])
print("cov", cov)
print("ncov", ncov)
print("qcov", qcov)
ff = np.array([[1, 0], [1, 1]])
res = MGSM.att_gexp(0, ff, ff, cov, ncov, qcov)
print(res)
def self_con(Q):
return [
MGSM.F_self_con(CNS, Q[0]),
[MGSM.F_self_con(C1[k], Q[1][k]) for k in range(4)],
[MGSM.F_self_con(C2[k], Q[2][k]) for k in range(4)]
]
#S = np.random.rand(m,m)
#S = np.dot(S,S.transpose())
S = np.array([[
np.exp(1 + np.cos(x - y))
for y in np.linspace(0, 2 * np.pi - (2. * np.pi / m), m)
] for x in np.linspace(0, 2 * np.pi - (2. * np.pi / m), m)])
print(S.shape)
b = np.random.multivariate_normal(np.zeros(m), S, n)
dd = a * b
chtest = MGSM.CHtov(
np.fliplr(
np.flipud(
np.transpose(np.linalg.cholesky(np.cov(np.transpose(dd)) / 2)))))
print(MGSM.log_likelihood_center(dd, S))
out, test = fit.fit_center(dd, GRAD=True, init=chtest)
print("TRUE", MGSM.log_likelihood_center(dd, S))
print(np.mean(np.abs(S)))
print(np.mean(np.abs(S - out)))
#for k in range(len(test)):
# print(np.mean(np.abs(S-test[k])))
np.savetxt("./truecov.csv", S)
np.savetxt("./fitcov.csv", out)
def make_resp(filt, model, quenched, K, mgs, tag, COR, ncor, n_trial=1):
P = COR[0]
CNS = COR[1]
C1 = COR[2]
C2 = COR[3]
NCNS = ncor[0]
NC1 = ncor[1]
NC2 = ncor[2]
fullcor = ncor[3]
print(CNS.shape)
print(C1.shape)
#make sure everything is the right type
assert model in ("ours", "coen_cagli")
assert quenched in (True, False)
assert type(ncor) == list
GV = []
GVnn = []
#generate samples for each filter set
if quenched:
nftemp = np.tile(filt, [n_trial, 1, 1])
noise = np.reshape(
np.random.multivariate_normal(np.zeros(np.prod(nftemp.shape[1:])),
fullcor, n_trial * filt.shape[0]),
tuple([n_trial * filt.shape[0]]) + filt.shape[1:])
nftemp += noise
else:
nftemp = filt
#run it for each value of k
SP = []
SPnn = []
for k in K:
print(k)
print(nftemp.shape)
GV.append(
MGSM.MGSM_g(nftemp, [CNS, C1, C2],
[NCNS * (k * k), NC1 * (k * k), NC2 * (k * k)], P))
GVnn.append(MGSM.MGSM_gnn(filt, [CNS, C1, C2], P))
if model == "ours":
SP.append(
MGSM.get_noisy_seg_weight(filt, CNS, C1, C2, NCNS * (k * k),
NC1 * (k * k), NC2 * (k * k), P))
SPnn.append(MGSM.get_seg_weight(filt, CNS, C1, C2, P))
elif model == "coen_cagli":
SP.append(
MGSM.get_noisy_CC_seg_weight(filt, CNS, C1, C2, NCNS * (k * k),
NC1 * (k * k), NC2 * (k * k), P))
SPnn.append(MGSM.get_CC_seg_weight(filt, CNS, C1, C2, P))
GV = np.concatenate(GV)
GVnn = np.concatenate(GVnn)
SP = np.concatenate(SP)
SPnn = np.concatenate(SPnn)
np.savetxt(tag + "_noisy.csv", GV)
np.savetxt(tag + "_clean.csv", GVnn)
np.savetxt(tag + "_SP_noisy.csv", SP)
np.savetxt(tag + "_SP_clean.csv", SPnn)
return GV, GVnn
def make_att_resp(filt1,
filt2,
model,
quenched,
K,
mgs,
tag,
COR,
ncor,
qcor,
fcor,
n_trial=1):
P = COR[0]
CNS = COR[1]
C1 = COR[2]
C2 = COR[3]
NCNS = ncor[0]
NC1 = ncor[1]
NC2 = ncor[2]
QCNS = qcor[0]
QC1 = qcor[1]
QC2 = qcor[2]
FCNS = fcor[0]
FC1 = fcor[1]
FC2 = fcor[2]
fullcor = ncor[3]
fullqcor = qcor[3]
print(CNS.shape)
print(C1.shape)
#make sure everything is the right type
assert model in ("ours", "coen_cagli")
assert quenched in (True, False)
assert type(ncor) == list
GV = []
GVnn = []
#run it for each value of k
SP = []
SPnn = []
for k in K:
print(k)
GV.append(
MGSM.MGSM_att_g(filt1, filt2, [CNS, C1, C2],
[NCNS * (k * k), NC1 * (k * k), NC2 * (k * k)],
[QCNS, QC1, QC2], [FCNS, FC1, FC2], P))
GVnn.append(MGSM.MGSM_gnn(filt2, [CNS, C1, C2], P))
SP.append(
MGSM.get_att_CC_seg_weight(filt1, filt2, CNS, C1, C2,
NCNS * (k * k), NC1 * (k * k),
NC2 * (k * k), QCNS, QC1, QC2, FCNS,
FC1, FC2, P))
SPnn.append(MGSM.get_CC_seg_weight(filt2, CNS, C1, C2, P))
GV = np.concatenate(GV)
GVnn = np.concatenate(GVnn)
SP = np.concatenate(SP)
SPnn = np.concatenate(SPnn)
np.savetxt(tag + "_noisy.csv", GV)
np.savetxt(tag + "_clean.csv", GVnn)
np.savetxt(tag + "_SP_noisy.csv", SP)
np.savetxt(tag + "_SP_clean.csv", SPnn)
return GV, GVnn
def mutual_information_data(data,dt,snr = [.5,1.],n_samples=1000,use_grat = True):
#get the BSDS file location
F = open("./CONFIG","r")
for l in F:
BSDSloc = l.split("=")[1]
break
F.close()
img_names = glob.glob(BSDSloc + "*.jpg")
###########################
maxframe = 20
par = data["params"]
#I want to make a bunch of gratings in 3 orientations
#I want to simulate them many times, with noise, and record the full temporal response for a while(until stability?)
#lets go for 100 presentations of each of the 3 gratings.
#I need enough to do a numerical estimation of the mutual information
if par["segmentation"] != "gsm":
print("Mutual Information is only for GSM models!")
exit()
f_pos = model_tools.get_f_pos(par["filter_position"],par["filter_distance"]*np.max(par["wavelengths"]),par["n_surrounds"])
indices = np.concatenate([[[a,b,c] for a in range(par["n_angles"]) for b in range(len(par["wavelengths"])) for c in range(2)] for p in f_pos])
positions = np.concatenate([[p for a in range(par["n_angles"]) for b in range(len(par["wavelengths"])) for c in range(2)] for p in f_pos])
fullsize = int(5*max(par["wavelengths"]) + 2*np.max(f_pos))
minwav = np.min(par["wavelengths"])
out = []
outv = []
imnum = 0
imind = np.random.choice(np.arange(len(img_names)),3,replace = False)
for o in np.linspace(0,np.pi,4)[:-1]:
print("ORI: {}".format(o))
#make the gratings
if use_grat:
grats = stim.make_grating(1.,o,par["wavelengths"][0],fullsize/2,fullsize)
print("Getting Coefficients")
#get filters
else:
grats = proc.load_grayscale_image(img_names[imind[imnum]])#.make_grating(1.,o,par["wavelengths"][0],fullsize/2,fullsize)
imnum += 1
print("Getting Coefficients")
#get filters
coeffs = make_data.get_filter_maps(grats,data["kernels"])
print(np.array(coeffs).shape)
#extract the right ones
out.append([])
outv.append([])
for n in range(len(snr)):
out[-1].append([])
outv[-1].append([])
print("SNR: {}".format(snr[n]))
path = [coeffs.shape[-1]/2 for k in range(maxframe)]
print(path)
rundat = make_data.sample_path(coeffs,path,indices,positions)/np.array([data["fac"]])
print(rundat.shape)
#add a few zeroes for before stimulus onset
Z = np.zeros([int(2)] + list(rundat.shape[1:]))
rundat = np.concatenate([Z,rundat],axis = 0)
#get all filters
ind = [k for k in range(len(indices))]
feps = [[linalg.logm(m)/data["params"]["walk_dt"] for m in f] for f in data["F"]]
FF = [[np.float32(linalg.expm(dt * data["params"]["walk_dt"] * m)) for m in f] for f in feps]
#print(FF[0][0])
QQ = [[inference.Q_self_con(data["C"][k][m],FF[k][m]) for m in range(len(data["F"][k]))] for k in range(len(data["F"]))]
NC = [[m/((snr[n])**2) for m in f] for f in data["C"]]
#add noise
noise = np.random.multivariate_normal(np.zeros(rundat.shape[1]),NC[0][0],[n_samples,rundat.shape[0]])
runf = noise + np.expand_dims(rundat,0)
#run the response analysis
for k in [1,len(Z)] + range(len(Z)+1,runf.shape[1],2) + [runf.shape[1]]:
print("{}\t{}".format(k,runf[:,:k].shape))
responses = inference.general_MGSM_g_att(runf[:,:k],data["segs"],data["C"],NC,QQ,FF,data["P"],ind,stable = True,op = False)
vresponses = inference.general_MGSM_p_att(runf[:,:k],data["segs"],data["C"],NC,QQ,FF,data["P"],ind,stable = True,op = False)
out[-1][-1].append(responses)
outv[-1][-1].append(vresponses)
out[-1][-1] = np.array(out[-1][-1])
outv[-1][-1] = np.array(outv[-1][-1])
return out,outv
def run_WDIFF_plots():
#what it is that we want to do here? We want to look at COS
resp1 = []
resp2 = []
resp3 = []
nresp1 = []
nresp2 = []
nresp3 = []
NUM_CURVE = 10
NUM_CON = 50
NUM_W = 20
CMAX = 1.
CMIN = -2.
N = 8
I1 = inp(N,.6,1,1,-1 + N/2)
I2 = inp(N,.6,1,0,-1 + N/2)
I3 = inp(N,.6,0,1,-1 + N/2)
WW = np.logspace(.1,10,NUM_CURVE)
con = np.logspace(CMIN,CMAX,NUM_CON)
for w in [0,1,-1]:
resp1.append([])
resp2.append([])
resp3.append([])
nresp1.append([])
nresp2.append([])
nresp3.append([])
for k in WW:
W = w
CC = cov(N,.6)
if W == 0 or W == -1:
NC = cov_special(N,W)
else:
NC = CC.copy()
NC = NC*np.linalg.norm(CC)/np.linalg.norm(NC)
np.savetxt("./covariance_test.csv",CC)
stim1 = np.array([cc * I1 for cc in np.logspace(CMIN,CMAX,NUM_CON)])
stim2 = np.array([cc * I2 for cc in np.logspace(CMIN,CMAX,NUM_CON)])
stim3 = np.array([cc * I3 for cc in np.logspace(CMIN,CMAX,NUM_CON)])
resp1[-1].append(MGSM.gexp(0,stim1,CC,NC/(k*k)))
resp2[-1].append(MGSM.gexp(0,stim2,CC,NC/(k*k)))
resp3[-1].append(MGSM.gexp(0,stim3,CC,NC/(k*k)))
nresp1[-1].append(MGSM.gnn(stim1,CC)[:,0])
nresp2[-1].append(MGSM.gnn(stim2,CC)[:,0])
nresp3[-1].append(MGSM.gnn(stim3,CC)[:,0])
resp1 = np.array(resp1)
resp2 = np.array(resp2)
resp3 = np.array(resp3)
nresp1 = np.array(nresp1)
nresp2 = np.array(nresp2)
nresp3 = np.array(nresp3)
AI = resp1/(resp2+resp3)
nAI = nresp1/(nresp2+nresp3)
for w in range(len(AI)):
fig = plt.figure()
ax = fig.add_subplot(111)
ax.set_prop_cycle("color",[cm(1.*i/NUM_CURVE) for i in range(NUM_CURVE)])
K = np.linspace(.5,5,NUM_CURVE)
cc = 100*np.logspace(CMIN,CMAX,NUM_CON)/(10**CMAX)
for k in range(0,len(AI[w]),len(AI[w])/5):
ax.plot(100*con/np.max(con),AI[w,k],label=str(K[k]))
ax.plot(100*con/np.max(con),nAI[w,k],'--')
handles, labels = ax.get_legend_handles_labels()
# ax.legend(bbox_to_anchor=(0., 1.02, 1., .102), loc=3,
# ncol=2, mode="expand", borderaxespad=0.)
ax.plot(cc,[1 for k in range(len(cc))],'k--')
plt.xlim([0,20])
plt.xlabel("Contrast")
plt.ylabel("A.I.")
plt.tight_layout()
fig.savefig("./paramfig_K_{}_by_W.pdf".format(w))
