def produce_mu_and_sd(n_samples,
hWidths,
xtrain,
ytrain,
xtest,
ytest,
precisions,
vy,
burnin=0,
seed=12345):
train_err, test_err, samples, train_op_samples = sampler_on_BayesNN(
burnin=0,
n_samples=n_samples,
precisions=precisions,
vy=vy,
X_train=xtrain,
y_train=ytrain,
hWidths=hWidths,
stepsize=0.001,
n_steps=30,
seed=seed)
# print 'sampling worked'
ntrain = xtrain.shape[0]
test_pred, test_sd = analyse_samples(samples,
xtrain,
ytrain,
hWidths=hWidths,
burnin=burnin,
display=False,
title='ntrain {}'.format(ntrain),
X_test=xtest,
y_test=ytest)
return test_pred, test_sd
def mixing(sf, vy, show_fit=False, showPost=False):
'''
:param sf: scale factor for precisions
:param vy: precision of noise
:return:
'''
ntrain = 20
noise_var = 0.01
X_train = np.random.uniform(low=-1.0, high=1.0,
size=ntrain).reshape(ntrain, 1)
# print X_train.shape
y_train = objective(X_train) + np.random.randn(ntrain, 1) * sqrt(noise_var)
# precisions = [0.6125875773048164, 0.03713439386866191, 14.22759780450891, 5.72501724650353]
# vy = 4.631095917555727
precisions = [1, 1]
precisions = [sf * x for x in precisions]
hWidths = [100]
train_err, test_err, samples, train_op_samples = sampler_on_BayesNN(
burnin=0,
n_samples=5000,
precisions=precisions,
vy=vy,
X_train=X_train,
y_train=y_train,
hWidths=hWidths,
target_acceptance_rate=0.9)
w1 = samples[:, 1]
w2 = samples[:, 3]
w3 = samples[:, 9]
# w4 = samples[:, 200]
plt.figure()
plt.plot(w1, label='w1')
plt.plot(w2, label='w2')
plt.plot(w3, label='w3')
plt.title('weight prec {}, noise prec {}'.format(sf, vy))
plt.legend()
plt.xlabel('Num Iterations')
plt.ylabel('Value')
# plt.savefig('logs/BNN_logs/mixingWeightsPrec10L', dpi=300)
print samples.shape
analyse_samples(samples,
X_train,
y_train,
hWidths=hWidths,
burnin=200,
display=True)
if (showPost):
samples = samples[200:, :] # burning in
w1 = samples[:, 1]
w2 = samples[:, 3]
w3 = samples[:, 9]
plt.figure()
N = samples.shape[0]
n = N / 10
plt.hist(w1, bins=n) # bin it into n = N/10 bins
plt.figure()
plt.hist(w2, bins=n) # bin it into n = N/10 bins
plt.figure()
plt.hist(w3, bins=n) # bin it into n = N/10 bins
def mixing(sf, vy, show_fit=False, show_post=False):
'''
:param sf: scale factor for precisions
:param vy: precision of noise
:return:
'''
ntrain = 20
noise_var = 0.01
X_train = np.random.uniform(low=-1.0, high=1.0, size=ntrain).reshape(ntrain, 1)
# print X_train.shape
y_train = objective(X_train) + np.random.randn(ntrain, 1) * sqrt(noise_var)
ntest = 1000
X_test = np.linspace(-1., 1., ntest)
y_test = objective(X_test)
X_test = X_test.reshape(ntest, 1)
y_test = y_test.reshape(ntest, 1)
# precisions = [0.6125875773048164, 0.03713439386866191, 14.22759780450891, 5.72501724650353]
# vy = 4.631095917555727
precisions = [1, 1, 1, 1]
# precisions = [1, 1]
precisions = [sf * x for x in precisions]
# hWidths = [50, 50, 50]
hWidths = [50, 50, 50]
# a, b, init_MAP = mlp_synthetic(X_train, X_test, y_train, y_test, precision=precisions[0], vy=vy, hWidths=hWidths,
# display=True, epochs=4000)
# # plt.show()
# print 'finished MAP'
# analyse_samples(init_MAP,X_train, y_train, hWidths=hWidths, burnin=0, display=True,title='MAP')
# plt.show()
train_err, test_err, samples, train_op_samples = sampler_on_BayesNN(burnin=0, n_samples=5000, precisions=precisions,
vy=vy,
X_train=X_train, y_train=y_train,
hWidths=hWidths, target_acceptance_rate=0.9,
stepsize=0.001,
n_steps=30)
# print RCodaTools.ess_coda_vec(np.transpose(samples))
# , init_theta=init_MAP
# print 'effective sample sizes'
# a = RCodaTools.ess_coda_vec(samples)
# print np.mean(a)
# print np.min(a)
# w1 = samples[:, 1]
# w2 = samples[:, 5200]
# w3 = samples[:, 1200]
# w4 = samples[:, 200]
theta_indices = [1, 200, 2501]
w1 = samples[:, theta_indices[0]]
w2 = samples[:, theta_indices[1]]
w3 = samples[:, theta_indices[2]]
# w4 = samples[:, 200]
plt.figure()
plt.plot(w1, label='theta {}'.format(theta_indices[0]))
plt.plot(w2, label='theta {}'.format(theta_indices[1]))
plt.plot(w3, label='theta {}'.format(theta_indices[2]))
# plt.title('weight prec {}, noise prec {}'.format(sf, vy))
plt.legend()
plt.xlabel('Sample number')
plt.ylabel('Value')
plt.savefig('report_images/trace.png',dpi=300,bbox_inches='tight')
# plt.savefig('logs/BNN_logs/mixingWeightsPrec10L', dpi=300)
print samples.shape
analyse_samples(samples, X_train, y_train,X_test,y_test, hWidths=hWidths, burnin=200, display=True)
analyse_mult_samples(samples, X_train, y_train,X_test,y_test, hWidths=hWidths, indices=[900,2200,2900])
# analyse_samples((samples[1750,:]).reshape(1,-1),X_train, y_train, hWidths=hWidths, burnin=0, display=True,title='sample=1750')
# analyse_samples((samples[240, :]).reshape(1, -1), X_train, y_train, hWidths=hWidths, burnin=0, display=True,
# title='sample=240')
# analyse_samples((samples[4000, :]).reshape(1,-1), X_train, y_train, hWidths=hWidths, burnin=0, display=True,title='sample=4000')
if (show_post):
samples = samples[200:, :] # burning in
w1 = samples[:, theta_indices[0]]
w2 = samples[:, theta_indices[1]]
w3 = samples[:, theta_indices[2]]
N = samples.shape[0]
n = N / 100
plt.figure()
plt.hist(w1, bins=n, normed=False) # bin it into n = N/10 bins
plt.xlabel('Value')
plt.ylabel('Occurences')
plt.savefig('report_images/posteriorW1.png', dpi=300,bbox_inches='tight')
plt.figure()
plt.hist(w2, bins=n, normed=False) # bin it into n = N/10 bins
plt.xlabel('Value')
plt.ylabel('Occurences')
plt.savefig('report_images/posteriorW200.png', dpi=300,bbox_inches='tight')
plt.figure()
plt.hist(w3, bins=n, normed=False) # bin it into n = N/10 bins
plt.xlabel('Value')
plt.ylabel('Occurences')
plt.savefig('report_images/posteriorW251.png', dpi=300,bbox_inches='tight')
def runSampler(kwargs):
return bayesNN_HMCv2.sampler_on_BayesNN(
kwargs['burnin'], kwargs['n_samples'], kwargs['precisions'],
kwargs['vy'], kwargs['hWidths'], kwargs['X_train'], kwargs['y_train'])
def runSampler1(args):
return bayesNN_HMCv2.sampler_on_BayesNN(*args)