def grad(self, V):
T, v = V.shape
assert (v == self.v)
h = self.h
VH, HH, b_init = self
G = 0 * self
d_VH, d_HH, d_b_init = G
H = zeros((T, h))
B = zeros((T, h))
H[0] = sigmoid(VH * V[[0]] + b_init[newaxis, :])
for t in range(1, T):
B[[t]] = HH * H[[t - 1]]
H[[t]] = sigmoid(VH * V[[t]] + B[[t]])
dB = zeros((T, h))
dBL = zeros((T, h))
F_t = zeros(h)
loss = 0
VH_t = 1 * VH
for t in reversed(range(T)):
dB[t] = H[t] * (1 - H[t]) * F_t
VH_t[2] = B[t] + VH[2]
if self.CD_n > 0:
dVH_t, dict_loss = rbm.rbm_grad_cd(VH_t, V[[t]], self.CD_n,
self.vis_gauss)
else:
dVH_t, dict_loss = rbm.rbm_grad_exact(VH_t, V[[t]],
self.vis_gauss)
loss += dict_loss['loss']
dBL[t] = dVH_t[2]
d_VH += dVH_t
HH.direction = up
F_t[:] = HH.T() * (dB[[t]] + dBL[[t]])
HH.direction = None
d_b_init += dB[0]
HH.direction = up
VH.direction = up
for t in range(1, T):
d_HH += HH.outp_up(H[[t - 1]], dB[[t]] + dBL[[t]])
d_VH += VH.outp_up(V[[t]], dB[[t]])
d_VH += VH.outp_up(V[[0]], dB[[0]])
HH.direction = None
VH.direction = None
return G, dict(loss=loss)
def grad(self, V):
T, v = V.shape
assert(v == self.v)
h = self.h
VH, HH, b_init = self
G = 0 * self
d_VH, d_HH, d_b_init = G
H = zeros((T, h))
B = zeros((T, h))
H[0] = sigmoid(VH * V[[0]]+ b_init[newaxis, :])
for t in range(1, T):
B[[t]] = HH*H[[t-1]]
H[[t]] = sigmoid(VH*V[[t]] + B[[t]])
dB = zeros((T, h))
dBL = zeros((T, h))
F_t = zeros(h)
loss =0
VH_t = 1 * VH
for t in reversed(range(T)):
VH_t[2] = B[t] + VH[2]
if self.CD_n > 0:
dVH_t, dict_loss = rbm.rbm_grad_cd (VH_t, V[[t]], self.CD_n, self.vis_gauss)
else:
dVH_t, dict_loss = rbm.rbm_grad_exact(VH_t, V[[t]], self.vis_gauss)
loss += dict_loss['loss']
d_VH += dVH_t
if t>0:
HH.direction = up
d_HH += HH.outp(H[[t-1]], dVH_t[2][newaxis,:])
HH.direction = None
else:
d_b_init += dVH_t[2]
return G, dict(loss=loss)
def loss(W, x):
return rbm.rbm_grad_cd(W, x, CD_n)[1]
def grad(W, x):
return rbm.rbm_grad_cd(W, x, CD_n)
def loss(W, x):
return rbm.rbm_grad_cd(W, x, CD_n)[1]
def grad(W, x):
return rbm.rbm_grad_cd(W, x, CD_n)
def loss(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)[1]
def grad(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)
def loss(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)[1]
def grad(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)