Create example_py3.py

example/example_py3.py

@@ -0,0 +1,144 @@
+"""
+Example script demonstrating usage of pSSLDA (compatible with python3)
+"""
+import pdb
+
+import numpy as NP
+import numpy.random as NPR
+
+from pSSLDA import infer
+import FastLDA 
+
+# First, we create some synthetic data
+#
+# D = number of documents
+# N = corpus length
+# T = number of topics
+# W = vocabulary size
+
+# Synthetic topics
+phi = NP.array([[0.45, 0.45, 0.05, 0.05],
+               [0.05, 0.05, 0.85, 0.05],
+                [0.05, 0.05, 0.05, 0.85]])
+(T, W) = phi.shape
+(D, doclen) = (1000, 100)
+(wordvec, docvec, zvec) = ([], [], [])
+# The theta for each doc will favor one topic
+theta = NP.array([0.8, 0.1, 0.1])
+# Generate each doc
+for di in range(D):
+    # Shuffle the topic weights for this doc
+    theta_di = NPR.permutation(theta)
+    # Sample for each word in the doc
+    for wi in range(doclen):
+        # Sample latent topic z
+        topic = NPR.multinomial(1,theta_di).nonzero()[0][0]
+        # Sample word from phi_z
+        word = NPR.multinomial(1,phi[topic,:]).nonzero()[0][0]
+        # Record sampled values
+        docvec.append(di)
+        wordvec.append(word)
+        zvec.append(topic)
+
+# pSSLDA input types
+#
+# w = Length N NumPy int array of word indices [0 <= w < W]
+# d = Length N NumPy int array of doc indices [0 <= d < D]
+# alpha = 1 x T NumPy float array of doc-topic Dirichlet hyperparameters
+# beta = T x W NumPy float array of topic-word Dirichlet hyperparameters
+#
+# zlabels = Length N Python List (one entry per index in corpus)
+#           If None
+#              no z-label, treat this z normally (as in standard LDA)
+#           Else
+#              Length T NumPy float array (one entry per topic)
+#              The probability of selecting each topic will
+#              be multiplied by exp(this value)
+#                
+
+# Python Lists must be converted to NumPy arrays
+(w, d) = (NP.array(wordvec, dtype = NP.int),
+          NP.array(docvec, dtype = NP.int))
+
+#TEST USING PRE-DEFINED DOCUMENTS (compare with py2 code)
+#(w, d) = (NP.load("wordvec.npy"),
+#          NP.load("docvec.npy"))
+
+# Create parameters
+alpha = NP.ones((1,T)) * 1
+beta = NP.ones((T,W)) * 0.01
+
+# How many parallel samplers do we wish to use?
+P = 2
+
+# Random number seed 
+randseed = 194582
+
+# Number of samples to take
+numsamp = 500
+
+# Do parallel inference
+finalz = infer(w, d, alpha, beta, numsamp, randseed, P)
+
+# Estimate phi and theta
+(nw, nd) = FastLDA.countMatrices(w, W, d, D, finalz, T)
+(estphi,esttheta) = FastLDA.estPhiTheta(nw, nd, alpha, beta)
+
+print('')
+print('True topics')
+print(str(phi))
+print('Estimated topics')
+print('\n'.join(['['+', '.join(['%.2f' % val for val in row]) + ']'
+                 for row in estphi]))
+print('')
+
+print('Estimated topics should be similar to ground truth')
+print('(up to a permutation of the rows)')
+print('enter \'c\' to continue...')
+print('')
+pdb.set_trace()
+
+#
+# Now, we add z-labels to *force* words 0 and 1 into separate topics
+# (note that this is different than ground truth)
+#
+labelweight = 5.0
+
+label0 = NP.zeros((T,), dtype=NP.float)
+label0[0] = labelweight
+
+label1 = NP.zeros((T,), dtype=NP.float)
+label1[1] = labelweight
+
+zlabels = []
+for wi in w:
+    if(wi == 0):
+        zlabels.append(label0)
+    elif(wi == 1):
+        zlabels.append(label1)
+    else:
+        zlabels.append(None)
+
+#print('z labels appended, doing inference')
+
+# Now inference will find topics with 0 and 1 in separate topics
+finalz = infer(w, d, alpha, beta, numsamp, randseed, P,
+               zlabels = zlabels)
+
+# Re-estimate phi and theta
+(nw, nd) = FastLDA.countMatrices(w, W, d, D, finalz, T)
+(estphi,esttheta) = FastLDA.estPhiTheta(nw, nd, alpha, beta)
+
+print( '')
+print( 'z-label word 0 to topic 0')
+print( 'z-label word 1 to topic 1')
+print( '')
+print( 'True topics')
+print( str(phi))
+print( 'z-label estimated topics')
+print( '\n'.join(['['+', '.join(['%.2f' % val for val in row]) + ']'
+                 for row in estphi]))
+print( '')
+print( 'Note that learned topics now obey z-labels')
+print( '(even though that \"disagrees\" with data)')
+print( '')