def do_estep(self, Wf, Wt, Z, H):
WZH = self.reconstruct(Wf, Wt, Z, H, circular=self.circular)
logprob = self.compute_logprob(Wf, Wt, Z, H, WZH)
WfZ = Wf * Z[newaxis,:]
WtZ = Wt * Z[:,newaxis]
VdivWZH = (self.V / (WZH + EPS))[:,:,newaxis]
self.VRWf[:] = 0
self.VRWt[:] = 0
self.VRH[:] = 0
for r in xrange(self.winF):
WfZshifted = shift(WfZ, r, 0, self.circularF)
for tau in xrange(self.winT):
Hshifted = shift(H[:,r,:], tau, 1, self.circularT)
## Hshifted : (rank, T)
## tmp : (F, T, rank)
## Previously:
# tmp = ((WZshifted[:,:,tau][:,:,newaxis]
# * Hshifted[newaxis,:,:]).transpose((0,2,1))
# * VdivWZH)
WZtau = WfZshifted * WtZ[:,tau][newaxis,:] # (F, rank)
tmp = Basilica.w_product(WZtau, Hshifted).transpose((0,2,1)) * VdivWZH
self.VRWf += shift(tmp.sum(1), -r, 0, self.circularF)
self.VRWt[:,tau] += tmp.sum(1).sum(0)
self.VRH[:,:,r] += shift(tmp.sum(0), -tau, 0, self.circularT)
return logprob, WZH
def shift_key_to_zero(W, Z, H): newW = np.zeros(W.shape) newH = np.zeros(H.shape) for k in xrange(len(Z)): key_profile = H[k].sum(1) main_key = np.argmax(key_profile) newW[:,k] = plca.shift(W[:,k], main_key, axis=0, circular=True) newH[k] = plca.shift(H[k], -main_key, axis=0, circular=True) return newW, Z, newH
def shift_key_to_zero(W, Z, H):
newW = np.zeros(W.shape)
newH = np.zeros(H.shape)
for k in xrange(len(Z)):
key_profile = H[k].sum(1)
main_key = np.argmax(key_profile)
newW[:, k] = plca.shift(W[:, k], main_key, axis=0, circular=True)
newH[k] = plca.shift(H[k], -main_key, axis=0, circular=True)
return newW, Z, newH
def find_downbeat(V, W):
newW = W.copy()
for k in xrange(W.shape[1]):
Wlen = compute_effective_pattern_length(W[:,k,:])
Wk = W[:,k,:Wlen]
corr = np.array([_compute_summary_correlation(plca.shift(Wk, r, 1), V)
for r in xrange(Wlen)])
bestshift = corr[:,Wlen:-Wlen].sum(1).argmin()
print k, Wlen, bestshift
newW[:,k,:Wlen] = plca.shift(Wk, bestshift, 1)
return newW
def find_downbeat(V, W):
newW = W.copy()
for k in xrange(W.shape[1]):
Wlen = compute_effective_pattern_length(W[:, k, :])
Wk = W[:, k, :Wlen]
corr = np.array([
_compute_summary_correlation(plca.shift(Wk, r, 1), V)
for r in xrange(Wlen)
])
bestshift = corr[:, Wlen:-Wlen].sum(1).argmin()
print k, Wlen, bestshift
newW[:, k, :Wlen] = plca.shift(Wk, bestshift, 1)
return newW
def find_downbeat_slow(V, W, Z, H, **kwargs):
bopt = np.zeros(len(Z), dtype=int)
nW = np.zeros(W.shape)
nH = np.zeros(H.shape)
for k in np.argsort(Z)[::-1]:
Wlen = compute_effective_pattern_length(W[:, k, :])
params = []
logprobs = []
for b in xrange(Wlen):
initW = W
initW[:, k, :Wlen] = plca.shift(W[:, k, :Wlen], b, axis=1)
W, Z, H, norm, recon, logprob = plca.FactoredSIPLCA2.analyze(
V,
rank=len(Z),
win=[H.shape[1], W.shape[-1]],
niter=50,
circular=[True, False],
initW=initW,
initH=np.ones(H.shape),
initZ=Z,
**kwargs)
params.append((W, Z, H))
logprobs.append(logprob)
print b, logprobs[-1]
bopt[k] = np.argmax(logprobs)
W[:, k] = params[bopt[k]][0][:, k]
nH[k] = params[bopt[k]][2][k]
return bopt, W, Z, nH
def find_downbeat_slow(V, W, Z, H, **kwargs):
bopt = np.zeros(len(Z), dtype=int)
nW = np.zeros(W.shape)
nH = np.zeros(H.shape)
for k in np.argsort(Z)[::-1]:
Wlen = compute_effective_pattern_length(W[:,k,:])
params = []
logprobs = []
for b in xrange(Wlen):
initW = W
initW[:,k,:Wlen] = plca.shift(W[:,k,:Wlen], b, axis=1)
W,Z,H,norm,recon,logprob = plca.FactoredSIPLCA2.analyze(
V, rank=len(Z), win=[H.shape[1], W.shape[-1]],
niter=50, circular=[True,False],
initW=initW, initH=np.ones(H.shape), initZ=Z,
**kwargs)
params.append((W,Z,H))
logprobs.append(logprob)
print b, logprobs[-1]
bopt[k] = np.argmax(logprobs)
W[:,k] = params[bopt[k]][0][:,k]
nH[k] = params[bopt[k]][2][k]
return bopt, W, Z, nH
