def __init__(self, y):
# Pre-cache a sparse LU decomposition of the FL matrix
from pygfl.utils import get_1d_penalty_matrix
from scipy.sparse.linalg import factorized
from scipy.sparse import csc_matrix
D = get_1d_penalty_matrix(y.shape[0])
D = np.vstack([D, np.zeros(y.shape[0])])
D[-1,-1] = 1e-6 # Nugget for full rank matrix
D = csc_matrix(D)
self.invD = factorized(D)
# Setup the fast GFL solver
from pygfl.solver import TrailSolver
from pygfl.trails import decompose_graph
from pygfl.utils import hypercube_edges, chains_to_trails
from networkx import Graph
edges = hypercube_edges(y.shape)
g = Graph()
g.add_edges_from(edges)
chains = decompose_graph(g, heuristic='greedy')
ntrails, trails, breakpoints, edges = chains_to_trails(chains)
self.solver = TrailSolver()
self.solver.set_data(y, edges, ntrails, trails, breakpoints)
from pygfl.easy import solve_gfl
self.beta = solve_gfl(y)
def train_gtv(X, y, q, minlam=0.2, maxlam=10., numlam=30, verbose=1, tf_k=0, penalty='gfl', **kwargs):
if isinstance(q, int):
q = (q, q)
grid = generate_grid(X, q)
# Divide the space into q^2 bins
data = np.zeros(q[0]*q[1])
weights = np.zeros(q[0]*q[1])
i = 0
for x1_left, x1_right in zip(grid[0][:-1], grid[0][1:]):
for x2_left, x2_right in zip(grid[1][:-1], grid[1][1:]):
vals = np.where((X[:,0] >= x1_left) * (X[:,0] < x1_right) * (X[:,1] >= x2_left) * (X[:,1] < x2_right))[0]
weights[i] = len(vals)
data[i] = y[vals].mean() if len(vals) > 0 else 0
i += 1
# Get the edges for a 2d grid
edges = hypercube_edges(q)
########### Setup the graph
if penalty == 'gfl':
g = Graph()
g.add_edges_from(edges)
chains = decompose_graph(g, heuristic='greedy')
ntrails, trails, breakpoints, edges = chains_to_trails(chains)
elif penalty == 'dp' or penalty == 'gamlasso':
trails = np.array(edges, dtype='int32').flatten()
breakpoints = np.array(range(2, len(trails)+1, 2), dtype='int32')
ntrails = len(breakpoints)
print '\tSetting up trail solver'
solver = TrailSolver(maxsteps=30000, penalty=penalty)
# Set the data and pre-cache any necessary structures
solver.set_data(data, edges, ntrails, trails, breakpoints, weights=weights)
print '\tSolving'
# Grid search to find the best lambda
results = solver.solution_path(minlam, maxlam, numlam, verbose=verbose)
results['grid'] = grid
return results
def smooth_fdr(data,
fdr_level,
edges=None,
initial_values=None,
verbose=0,
null_dist=None,
signal_dist=None,
num_sweeps=10,
missing_val=None):
flat_data = data.flatten()
nonmissing_flat_data = flat_data
if edges is None:
if verbose:
print(
'Using default edge set of a grid of same shape as the data: {0}'
.format(data.shape))
edges = hypercube_edges(data.shape)
if missing_val is not None:
if verbose:
print(
'Removing all data points whose data value is {0}'.format(
missing_val))
edges = [(e1, e2) for (e1, e2) in edges
if flat_data[e1] != missing_val
and flat_data[e2] != missing_val]
nonmissing_flat_data = flat_data[flat_data != missing_val]
# Decompose the graph into trails
g = Graph()
g.add_edges_from(edges)
chains = decompose_graph(g, heuristic='greedy')
ntrails, trails, breakpoints, edges = chains_to_trails(chains)
if null_dist is None:
# empirical null estimation
mu0, sigma0 = empirical_null(nonmissing_flat_data,
verbose=max(0, verbose - 1))
elif isinstance(null_dist, GaussianKnown):
mu0, sigma0 = null_dist.mean, null_dist.stdev
else:
mu0, sigma0 = null_dist
null_dist = GaussianKnown(mu0, sigma0)
if verbose:
print('Empirical null: {0}'.format(null_dist))
# signal distribution estimation
if verbose:
print('Running predictive recursion for {0} sweeps'.format(num_sweeps))
if signal_dist is None:
grid_x = np.linspace(min(-20,
nonmissing_flat_data.min() - 1),
max(nonmissing_flat_data.max() + 1, 20), 220)
pr_results = predictive_recursion(nonmissing_flat_data,
num_sweeps,
grid_x,
mu0=mu0,
sig0=sigma0)
signal_dist = GridDistribution(pr_results['grid_x'],
pr_results['y_signal'])
if verbose:
print('Smoothing priors via solution path algorithm')
solver = TrailSolver()
solver.set_data(flat_data, edges, ntrails, trails, breakpoints)
results = solution_path_smooth_fdr(flat_data,
solver,
null_dist,
signal_dist,
verbose=max(0, verbose - 1))
results['discoveries'] = calc_fdr(results['posteriors'], fdr_level)
results['null_dist'] = null_dist
results['signal_dist'] = signal_dist
# Reshape everything back to the original data shape
results['betas'] = results['betas'].reshape(data.shape)
results['priors'] = results['priors'].reshape(data.shape)
results['posteriors'] = results['posteriors'].reshape(data.shape)
results['discoveries'] = results['discoveries'].reshape(data.shape)
results['beta_iters'] = np.array(
[x.reshape(data.shape) for x in results['beta_iters']])
results['prior_iters'] = np.array(
[x.reshape(data.shape) for x in results['prior_iters']])
results['posterior_iters'] = np.array(
[x.reshape(data.shape) for x in results['posterior_iters']])
return results
def smooth_fdr_known_dists(data,
fdr_level,
null_dist,
signal_dist,
edges=None,
initial_values=None,
verbose=0,
missing_val=None):
'''FDR smoothing where the null and alternative distributions are known
(and not necessarily Gaussian). Both must define the function pdf.'''
flat_data = data.flatten()
nonmissing_flat_data = flat_data
if edges is None:
if verbose:
print(
'Using default edge set of a grid of same shape as the data: {0}'
.format(data.shape))
edges = hypercube_edges(data.shape)
if missing_val is not None:
if verbose:
print(
'Removing all data points whose data value is {0}'.format(
missing_val))
edges = [(e1, e2) for (e1, e2) in edges
if flat_data[e1] != missing_val
and flat_data[e2] != missing_val]
nonmissing_flat_data = flat_data[flat_data != missing_val]
# Decompose the graph into trails
g = Graph()
g.add_edges_from(edges)
chains = decompose_graph(g, heuristic='greedy')
ntrails, trails, breakpoints, edges = chains_to_trails(chains)
if verbose:
print('Smoothing priors via solution path algorithm')
solver = TrailSolver()
solver.set_data(flat_data, edges, ntrails, trails, breakpoints)
results = solution_path_smooth_fdr(flat_data,
solver,
null_dist,
signal_dist,
verbose=max(0, verbose - 1))
results['discoveries'] = calc_fdr(results['posteriors'], fdr_level)
results['null_dist'] = null_dist
results['signal_dist'] = signal_dist
# Reshape everything back to the original data shape
results['betas'] = results['betas'].reshape(data.shape)
results['priors'] = results['priors'].reshape(data.shape)
results['posteriors'] = results['posteriors'].reshape(data.shape)
results['discoveries'] = results['discoveries'].reshape(data.shape)
results['beta_iters'] = np.array(
[x.reshape(data.shape) for x in results['beta_iters']])
results['prior_iters'] = np.array(
[x.reshape(data.shape) for x in results['prior_iters']])
results['posterior_iters'] = np.array(
[x.reshape(data.shape) for x in results['posterior_iters']])
return results
def count_plateaus(data):
shape = (100, 100)
edges = edge_map_from_edge_list(hypercube_edges(shape))
return len(calc_plateaus(data, edges))
print dataset
for i, (model, skips) in enumerate(zip(models, skiprows)):
print '\t{0}'.format(model)
for trial in xrange(numtrials):
print '\t\t{0}'.format(trial)
results[i, j, :2, trial] = np.loadtxt(
'data/uci/{0}/results/{1}/{2}.csv'.format(
dataset, model, trial),
delimiter=',',
skiprows=skips)[:2]
sweep = np.loadtxt('data/uci/{0}/sweeps/{1}/{2}.csv'.format(
dataset, model, trial),
delimiter=',',
skiprows=skips)
shape = (1000, 1000)
edges = edge_map_from_edge_list(hypercube_edges(shape))
results[i, j, 2, trial] = len(calc_plateaus(sweep, edges))
results[
i, j, 3,
trial] = -0.5 * n * results[i, j, 0, trial]**2 + results[
i, j, 2, trial] * (np.log(n) - np.log(2 * np.pi))
agg_results = results.mean(axis=3)
agg_std = results.std(axis=3)
dargs = {}
for j, (dataset, n) in enumerate(zip(datasets, N)):
dargs['dataset'] = dataset
dargs['N'] = n
print '''\multicolumn{1}{l}{}{dataset} & \multicolumn{3}{c}{''' + '{dataset} (N = {N})'.format(
**dargs) + '''} \\'''
from pygfl.utils import calc_plateaus, hypercube_edges, edge_map_from_edge_list
if __name__ == '__main__':
cities = [('austin', 'austin2014', 100), ('chicago', 'chicago2015', 200)]
for city, _, q in cities:
models = ['cart', 'crisp', 'gapcrisp', 'gfl']
names = ['CART', 'CRISP', 'GapCRISP', 'GapTV']
skiprows = [1, 1, 1, 0]
numtrials = 20
data = np.loadtxt('data/crime/{0}/all.csv'.format(city), delimiter=',')
shape = (q, q)
N = shape[0] * shape[1]
# Get all the nodes that exist in the grid
nodeset = set([i * shape[0] + j for i, j, _ in data])
edges = hypercube_edges(shape)
edges = edge_map_from_edge_list(edges)
results = np.zeros((len(models), 6, numtrials))
for i, (model, skips) in enumerate(zip(models, skiprows)):
print '\t{0}'.format(model)
for trial in xrange(numtrials):
results[i, :2, trial] = np.loadtxt(
'data/crime/{0}/results/{1}/{2}.csv'.format(
city, model, trial),
delimiter=',',
skiprows=skips)[:2]
sweep = np.loadtxt('data/crime/{0}/sweeps/{1}/{2}.csv'.format(
city, model, trial),
delimiter=',',
skiprows=skips)
