def generate_synth_data(V=10, D=10, T=5, N=100, alpha_beta=10., alpha_theta=10., plot=False, train_frac=0.5):
# true_lda = StandardLDA(T, V, alpha_beta=alpha_beta, alpha_theta=alpha_theta)
true_lda = LogisticNormalCorrelatedLDA(T, V, alpha_beta=alpha_beta)
print("Sigma: ", true_lda.Sigma)
# true_lda = StickbreakingCorrelatedLDA(T, V, alpha_beta=alpha_beta)
data = np.zeros((D,V),dtype=int)
for d in xrange(D):
doc = true_lda.generate(N=N, keep=True)
data[d,:] = doc.w
if plot:
plt.figure()
plt.imshow(data, interpolation="none")
plt.xlabel("Vocabulary")
plt.ylabel("Documents")
plt.colorbar()
plt.show()
# Split each document into two
train_data = np.zeros_like(data)
test_data = np.zeros_like(data)
for d,w in enumerate(data):
# Get vector where i is repeated w[i] times
wcnt = ibincount(w)
# Subsample wcnt
train_inds = np.random.rand(wcnt.size) < train_frac
train_data[d] = np.bincount(wcnt[train_inds], minlength=V)
test_data[d] = np.bincount(wcnt[~train_inds], minlength=V)
assert np.allclose(train_data[d] + test_data[d], w)
return true_lda, train_data, test_data
std_results = \
train(std_model, thetas=true_lda.thetas if init_to_true else None)
std_collapsed_model = StandardLDA(T,V,alpha_beta,alpha_theta)
std_collapsed_model.beta = true_lda.beta if init_to_true else std_collapsed_model.beta
std_collapsed_results = \
train(std_collapsed_model,
method='resample_model_collapsed',
thetas=true_lda.thetas if init_to_true else None)
sb_model = StickbreakingCorrelatedLDA(T, V, alpha_beta)
sb_model.beta = true_lda.beta if init_to_true else sb_model.beta
sb_results = \
train(sb_model, thetas=true_lda.thetas if init_to_true else None)
ln_model = LogisticNormalCorrelatedLDA(T, V, alpha_beta)
ln_model.beta = true_lda.beta if init_to_true else ln_model.beta
ln_results = \
train(ln_model, thetas=true_lda.thetas if init_to_true else None)
all_results = [sb_results, ln_results, std_results, std_collapsed_results]
all_labels = ["SB Corr. LDA", "LN Corr. LDA", "Std. LDA", "Collapsed LDA"]
# all_results = [std_results, std_collapsed_results]
# all_labels = ["Std. LDA", "Collapsed LDA"]
# all_results = [ln_results]
# all_labels = ["LN Corr. LDA"]
plt.figure()
# Plot log likelihood vs iteration
plt.subplot(121)
for ind, (results, label) in enumerate(zip(all_results, all_labels)):