def jac(train_data, engine, penalty=1e-2):
w = engine.scorer.parameters
g = np.zeros_like(w)
for X, y in train_data:
jt = engine.run(X)
jt_cond = engine.observe_target(jt, y)
fe = loglin.feature_expectations(jt)
fe_cond = loglin.feature_expectations(jt_cond)
g += fe_cond - fe
return g
np.round(weights.collapsed()[:10], 3) cm = loglin.ConditionalModel(models) w = cm.initial_weights(0) np.round(np.exp(cm.marg_log_proba(train_data[0][0], w)), 3) import imp imp.reload(loglin) weights = loglin.ModelWeights(num_subtypes) scorer = loglin.Scorer(models, weights) engine = loglin.InferenceEngine(scorer) s, p = engine.run(train_data[0][0]) np.round(s[0], 3) weights = loglin.ModelWeights(num_subtypes) scorer = loglin.Scorer(models, weights) engine = loglin.InferenceEngine(scorer) junction_tree = engine.run(train_data[0][0]) np.round(loglin.feature_expectations(junction_tree)[:20], 3) import imp imp.reload(loglin) weights = loglin.ModelWeights(num_subtypes) scorer = loglin.Scorer(models, weights) engine = loglin.InferenceEngine(scorer) junction_tree = engine.run(train_data[0][0]) np.round(loglin.feature_expectations(junction_tree)[:20], 3) weights = loglin.ModelWeights(num_subtypes) scorer = loglin.Scorer(models, weights) engine = loglin.InferenceEngine(scorer) junction_tree = engine.run(train_data[0][0]) np.round(loglin.feature_expectations(junction_tree)[:20], 3) import imp imp.reload(loglin) weights = loglin.ModelWeights(num_subtypes)
