def reconstruction(self, params, inputs, **kwargs):
"""
"""
x, y = inputs
n, _ = x.shape
weights_xf = params[:self.xf_sz].reshape(self.xfshape)
weights_yf = params[self.xf_sz:self._cum_xy].reshape(self.yfshape)
weights_fh = params[self._cum_xy:self._cum_xyh].reshape(self.fhshape)
bias_h = params[self._cum_xyh:self.size]
bias_x = params[self.size:-self.shape[0][1]]
bias_y = params[-self.shape[0][1]:]
factors_x = gdot(x, weights_xf)
factors_y = gdot(y, weights_yf)
factors = factors_x * factors_y
h, h_sampled = bernoulli(factors, wm=weights_fh, bias=bias_h, sampling=True)
rho_hat = h.sum()
factors_h = gdot(h, weights_fh.T)
way = np.random.rand() > 0.5
if way:
# reconstruct y (output) first.
tmp = factors_x * factors_h
y1, _ = self.V(tmp, wm=weights_yf.T, bias=bias_y)
factors_y[:] = gdot(y1, weights_yf)
# then reconstruct x (input).
tmp = factors_y * factors_h
x1, _ = self.V(tmp, wm=weights_xf.T, bias=bias_x)
else:
# reconstruct x (input) first.
tmp = factors_y * factors_h
x1, _ = self.V(tmp, wm=weights_xf.T, bias=bias_x)
factors_x[:] = gdot(x1, weights_xf)
# then reconstruct y (output).
tmp = factors_x * factors_h
y1, _ = self.V(tmp, wm=weights_yf.T, bias=bias_y)
xrec = gsum((x - x1)**2)
yrec = gsum((y - y1)**2)
return np.array([xrec, yrec, self.lmbd*rho_hat, self.avg_nxyf, self.avg_nfh])
def cd1_3way_grad(self, params, inputs, **kwargs):
"""
"""
g = gzeros(params.shape)
x, y = inputs
n, _ = x.shape
#print self.avg_nxyf, self.avg_nfh
weights_xf = params[:self.xf_sz].reshape(self.xfshape)
weights_yf = params[self.xf_sz:self._cum_xy].reshape(self.yfshape)
weights_fh = params[self._cum_xy:self._cum_xyh].reshape(self.fhshape)
bias_h = params[self._cum_xyh:self.size]
bias_x = params[self.size:-self.shape[0][1]]
bias_y = params[-self.shape[0][1]:]
# normalize weights
sq_xf = weights_xf * weights_xf
norm_xf = gpu.sqrt(sq_xf.sum(axis=0)) + SMALL
sq_yf = weights_yf * weights_yf
norm_yf = gpu.sqrt(sq_yf.sum(axis=0)) + SMALL
norm_xyf = (norm_xf.mean() + norm_yf.mean())/2.
self.avg_nxyf *= 0.95
self.avg_nxyf += (0.05 * norm_xyf)
weights_xf *= (self.avg_nxyf / norm_xf)
weights_yf *= (self.avg_nxyf / norm_yf)
sq_fh = weights_fh*weights_fh
norm_fh = gpu.sqrt(sq_fh.sum(axis=1)) + SMALL
self.avg_nfh *= 0.95
self.avg_nfh += (0.05 * norm_fh.mean())
weights_fh *= (self.avg_nfh / norm_fh[:, gpu.newaxis])
# normalization done
factors_x = gdot(x, weights_xf)
factors_y = gdot(y, weights_yf)
factors = factors_x * factors_y
h, h_sampled = bernoulli(factors, wm=weights_fh, bias=bias_h, sampling=True)
factors_h = gdot(h_sampled, weights_fh.T)
g[:self.xf_sz] = -gdot(x.T, factors_y*factors_h).ravel()
g[self.xf_sz:self._cum_xy] = -gdot(y.T, factors_x*factors_h).ravel()
g[self._cum_xy:self._cum_xyh] = -gdot(factors.T, h_sampled).ravel()
g[self._cum_xyh:self.size] = -h.sum(axis=0)
g[self.size:-self.shape[0][1]] = -x.sum(axis=0)
g[-self.shape[0][1]:] = -y.sum(axis=0)
# 3way cd
way = np.random.rand() > 0.5
if way:
# reconstruct y (output) first.
tmp = factors_x * factors_h
y1, _ = self.V(tmp, wm=weights_yf.T, bias=bias_y)
factors_y[:] = gdot(y1, weights_yf)
# then reconstruct x (input).
tmp = factors_y * factors_h
x1, _ = self.V(tmp, wm=weights_xf.T, bias=bias_x)
factors_x[:] = gdot(x1, weights_xf)
else:
# reconstruct x (input) first.
tmp = factors_y * factors_h
x1, _ = self.V(tmp, wm=weights_xf.T, bias=bias_x)
factors_x[:] = gdot(x1, weights_xf)
# then reconstruct y (output).
tmp = factors_x * factors_h
y1, _ = self.V(tmp, wm=weights_yf.T, bias=bias_y)
factors_y[:] = gdot(y1, weights_yf)
factors[:] = factors_x * factors_y
h1, _ = bernoulli(factors, wm=weights_fh, bias=bias_h)
factors_h[:] = gdot(h1, weights_fh.T)
g[:self.xf_sz] += gdot(x1.T, factors_y*factors_h).ravel()
g[:self.xf_sz] *= 1./n
g[self.xf_sz:self._cum_xy] += gdot(y1.T, factors_x*factors_h).ravel()
g[self.xf_sz:self._cum_xy] *= 1./n
g[self._cum_xy:self._cum_xyh] += gdot(factors.T, h1).ravel()
g[self._cum_xy:self._cum_xyh] *= 1./n
g[self._cum_xyh:self.size] += h1.sum(axis=0)
g[self._cum_xyh:self.size] *= 1./n
g[self.size:-self.shape[0][1]] += x1.sum(axis=0)
g[self.size:-self.shape[0][1]] *= 1./n
g[-self.shape[0][1]:] += y1.sum(axis=0)
g[-self.shape[0][1]:] *= 1./n
return g