made gibbs_sample.py look pretty

test_codes/pymc_compare/gibbs_sample.py

+"""
+Perform a Gibbs sampling using my function of the test data.
+
+Also perform maximum likelihood estimation of the same weights.
+"""
 import numpy as np
 import pyhsmm.basic.distributions as distributions
 from scipy.optimize import minimize
 import pickle
 
-n_samples = 100000
+# Data Params
 T = 16
 n_inputs = 3
 step_size = 0.2
 
+# Sampling params
+n_samples = 100000
+
+# Setup
 Q = np.tile(np.eye(n_inputs), (T, 1, 1))
 sample = pickle.load(open('test_data', 'rb'))
 learn = distributions.Dynamic_GLM(n_inputs=n_inputs, T=T, P_0=4 * np.eye(n_inputs), Q=Q * step_size, prior_mean=np.zeros(n_inputs))
 
-
-def wrapper(w, t):
-    learn.weights = np.tile(w, (T, 1))
-    return - np.sum(learn.log_likelihood(sample[t], t))
-
-
+# Draw samples
 samples = []
-pseudo_samples = []
 for _ in range(n_samples):
     if _ % 1000 == 0:
         print(_)
     learn.resample(sample)
     samples.append(learn.weights.copy())
 
+
+def wrapper(w, t):
+    """Reshape weight vector w into the correct shape, then compute the max ll estimate for the desired time t."""
+    learn.weights = np.tile(w, (T, 1))
+    return - np.sum(learn.log_likelihood(sample[t], t))
+
+
+# Compute max ll estimates
 LL_weights = np.zeros((T, n_inputs))
 for t in range(T):
-    f = lambda w: wrapper(w, t)
-    LL_weights[t] = minimize(f, np.zeros(n_inputs)).x
+    LL_weights[t] = minimize(lambda w: wrapper(w, t), np.zeros(n_inputs)).x
 
+# save everything
 pickle.dump((samples, LL_weights), open('gibbs_posterior', 'wb'))