def _get_norms(xz):
x = swyft.get_x(xz)
z = swyft.get_z(xz)
x_mean = sum(x) / len(x)
z_mean = sum(z) / len(z)
x_var = sum([(x[i] - x_mean)**2 for i in range(len(x))]) / len(x)
z_var = sum([(z[i] - z_mean)**2 for i in range(len(z))]) / len(z)
return x_mean, x_var**0.5, z_mean, z_var**0.5
# Generate 2-dim posteriors
network_2d = swyft.MLP_2d(x_dim, n_hidden, xz_init=xz)
losses += swyft.train(network_2d,
xz,
n_steps=n_steps,
lr=1e-3,
n_particles=n_particles)
losses += swyft.train(network_2d,
xz,
n_steps=n_steps,
lr=1e-4,
n_particles=n_particles)
# Plot results 1-dim
z_lnL = swyft.estimate_lnL(network, x0, swyft.get_z(xz))
plt.clf()
plt.plot(z_lnL[0]['z'], np.exp(z_lnL[0]['lnL']))
plt.axvline(0.1)
plt.axvline(0.9)
plt.savefig("figs/testrun_09a.png")
# plot 2-dim results
z_lnL = swyft.estimate_lnL_2d(network_2d, x0, swyft.get_z(xz))
lnL = z_lnL['lnL']
z_ij = z_lnL['z']
plt.clf()
plt.tricontour([zzz[0] for zzz in z_ij], [zzz[1] for zzz in z_ij],
lnL * 2,
levels=[-36, -25, -16, -9, -4, -1, 0])
t = np.linspace(0, 6.5, 1000)
# Generate test z0 and x0
z0 = np.array([0.10, 0.50])
x0 = model(z0, sigma=0.)
x_dim = len(x0)
z_dim = len(z0)
# Initialize loss list
losses = []
# And the first run
xz1 = swyft.init_xz(model, n_sims=n_sims, n_dim=z_dim)
network1 = Network(x_dim, z_dim, n_hidden, xz_init=xz1)
losses += swyft.train(network1,
xz1,
n_steps=n_steps,
lr=1e-3,
n_particles=n_particles)
losses += swyft.train(network1,
xz1,
n_steps=n_steps,
lr=1e-4,
n_particles=n_particles)
# Plot results
z_lnL = swyft.estimate_lnL(network1, x0, swyft.get_z(xz1))
plt.plot(z_lnL[0]['z'], np.exp(z_lnL[0]['lnL']))
plt.axvline(0.1)
plt.axvline(0.9)
plt.savefig("figs/testrun_05.png")
xz2,
n_steps=n_steps,
lr=1e-4,
n_particles=n_particles)
xz3 = swyft.update_xz(xz2,
network2,
x0,
model,
n_sims=n_sims,
lnL_th=np.log(1e-5),
append=False)
network3 = swyft.MLP(x_dim, z_dim, n_hidden, xz_init=xz3)
losses += swyft.train(network3,
xz3,
n_steps=n_steps,
lr=1e-3,
n_particles=n_particles)
losses += swyft.train(network3,
xz3,
n_steps=n_steps,
lr=1e-4,
n_particles=n_particles)
# Plot results
z_lnL = swyft.estimate_lnL(network3, x0, swyft.get_z(xz3))
plt.plot(z_lnL[0]['z'], np.exp(z_lnL[0]['lnL']))
plt.axvline(0.1)
plt.axvline(0.9)
plt.savefig("figs/testrun_02.png")