def run_PMF(X,M_training,M_test,K):
pmf = VariationalPMF(X,M_training,K)
pmf.initialize()
pmf.run(iterations,updates)
X_pred = pmf.predicted_X
# Calculate MSE of predictions.
MSE = statistics.MSE(X,X_pred,M_test)
print "Performance on test set: MSE=%s." % MSE
return MSE
def recover(M,source,target,iterations,calc_predictions=False,M_inv=[]): (I,J,K,X,U,V) = load_X_U_V(source) PMF = VariationalPMF(X,M,K) PMF.run(iterations=iterations,updates=10,calc_predictions=calc_predictions,M_inv=M_inv) predicted_U = PMF.U predicted_V = PMF.V predicted_X = PMF.predicted_X # Write predicted_X, U, V to output file store_X_U_V(target,predicted_X,predicted_U,predicted_V) return
if __name__ == "__main__":
X = read_data("./../../data/gi50_no_missing.txt")
(I,J) = X.shape
fraction_unknown = 0.1
M = generate_M(I,J,fraction_unknown)
M_inv = calc_inverse_M(M)
K = 3
outputfile = "recovered_matrices.txt"
iterations = 10
PMF = VariationalPMF(X,M,K)
PMF.run(iterations=iterations,updates=1,calc_predictions=True,M_inv=M_inv)
predicted_U = PMF.U
predicted_V = PMF.V
predicted_X = PMF.predicted_X
# Store the predicted matrix X with U and V
store_X_U_V(outputfile,predicted_X,predicted_U,predicted_V)
# Now we plot the predictions vs the true values
actual_vs_predicted = recover_predictions(M,X,predicted_X)
(actual,predicted) = zip(*actual_vs_predicted)
RMSE_predictions = compute_RMSE(actual_vs_predicted)
RMSE_training = PMF.RMSE
def test_updates(): # We create a simple test case by overwriting the random initialization, # and then test each of the updates. X = numpy.array([[-1,2],[3,4]]) M = numpy.array([[0,1],[1,1]]) K = 2 # Reset values r_alpha = numpy.array([0.1,0.2]) r_beta = numpy.array([0.3,0.4]) r_tau = 0.5 r_U = numpy.array([[0.6,0.7],[0.8,0.9]]) r_S_U = numpy.array([[1.0,1.1],[1.2,1.3]]) r_V = numpy.array([[1.4,1.5],[1.6,1.7]]) r_S_V = numpy.array([[1.8,1.9],[2.0,2.1]]) r_R = numpy.array([[0, (2-(0.6*1.5)-(0.8*1.7))], [(3-(0.7*1.4)-(0.9*1.6)),(4-(0.7*1.5)-(0.9*1.7))]]) PMF = VariationalPMF(X,M,K) PMF.run(0) def reset(): # Overwrite everything for the test case PMF.alpha[:] = r_alpha PMF.beta[:] = r_beta PMF.tau = r_tau PMF.U[:] = r_U PMF.S_U[:] = r_S_U PMF.V[:] = r_V PMF.S_V[:] = r_S_V PMF.R[:] = r_R # Test updating the hyperparameters reset() PMF.update_hyperparameters() expected_alpha = numpy.array([ 2.0 / ((0.6**2+1.0)+(0.7**2+1.1)), 2.0 / ((0.8**2+1.2)+(0.9**2+1.3)) ]) expected_beta = numpy.array([ 2.0 / ((1.4**2+1.8)+(1.5**2+1.9)), 2.0 / ((1.6**2+2.0)+(1.7**2+2.1)) ]) # tau = E_21 + E_12 + E_22 (because M_11=0) E_12 = (2-(0.6*1.5+0.8*1.7))**2 + (0.6**2*1.9+1.5**2*1.0+1.0*1.9) + (0.8**2*2.1+1.7**2*1.2+2.1*1.2) E_21 = (3-(0.7*1.4+0.9*1.6))**2 + (0.7**2*1.8+1.4**2*1.1+1.1*1.8) + (0.9**2*2.0+1.6**2*1.3+2.0*1.3) E_22 = (4-(0.7*1.5+0.9*1.7))**2 + (0.7**2*1.9+1.5**2*1.1+1.1*1.9) + (0.9**2*2.1+1.7**2*1.3+2.1*1.3) expected_tau = 3.0 / (E_21 + E_12 + E_22) assert numpy.array_equal(PMF.alpha,expected_alpha) assert numpy.array_equal(PMF.beta,expected_beta) assert PMF.tau == expected_tau # Test updating U. Testing U itself is hard because the values of R_ij change, # so to verify you would have to run the program itself anyways. reset() PMF.update_U() expected_S_U = numpy.array([ [ 1.0/(0.1 + 0.5 * (1.5**2+1.9)), 1.0/(0.1 + 0.5 * (1.4**2+1.8 + 1.5**2+1.9)) ], [ 1.0/(0.2 + 0.5 * (1.7**2+2.1)), 1.0/(0.2 + 0.5 * (1.6**2+2.0 + 1.7**2+2.1)) ] ]) assert numpy.array_equal(PMF.S_U,expected_S_U) assert True