record the number of iterations and convergence status

also log a warning if GLM.fit() doesn't converge

pyglmnet/pyglmnet.py

@@ -1,6 +1,7 @@
 """Python implementation of elastic-net regularized GLMs."""
 
 from copy import deepcopy
+from math import inf
 
 import numpy as np
 from scipy.special import expit
@@ -478,6 +479,8 @@ def __init__(self, distr='poisson', alpha=0.5,
  self.solver = solver
  self.learning_rate = learning_rate
  self.max_iter = max_iter
+ self.n_iter = 0
+ self.converged = False
  self.beta0_ = None
  self.beta_ = None
  self.ynull_ = None
@@ -717,14 +720,21 @@ def fit(self, X, y):
 
  # Iterative updates
  for t in range(0, self.max_iter):
+ logger.info("t: %i" % t)
+ convergence_metric = inf
  if self.solver == 'batch-gradient':
  grad = _grad_L2loss(self.distr,
  alpha, self.Tau,
  reg_lambda, X, y, self.eta,
  beta)
  # Converged if the norm(gradient) < tol
- if (t > 1) and (np.linalg.norm(grad) < tol):
- msg = ('\tConverged in {0:d} iterations'.format(t))
+ convergence_metric = np.linalg.norm(grad)
+ logger.info("convergence_metric: %f" % convergence_metric)
+ if (t > 1) and (convergence_metric < tol):
+ self.converged = True
+ self.n_iter += t
+ msg = ('\tConverged in {0:d} iterations'
+ .format(self.n_iter))
  logger.info(msg)
  break
  beta = beta - self.learning_rate * grad
@@ -734,8 +744,13 @@ def fit(self, X, y):
  beta, z = \
  self._cdfast(X, y, z, ActiveSet, beta, reg_lambda)
  # Converged if the norm(update) < tol
- if (t > 1) and (np.linalg.norm(beta - beta_old) < tol):
- msg = ('\tConverged in {0:d} iterations'.format(t))
+ convergence_metric = np.linalg.norm(beta - beta_old)
+ logger.info("convergence_metric: %f" % convergence_metric)
+ if (t > 1) and (convergence_metric < tol):
+ self.converged = True
+ self.n_iter += t
+ msg = ('\tConverged in {0:d} iterations'
+ .format(self.n_iter))
  logger.info(msg)
  break
  # Apply proximal operator
@@ -749,6 +764,16 @@ def fit(self, X, y):
  # Compute and save loss if callbacks are requested
  if callable(self.callback):
  self.callback(beta)
+ else:
+ # Warn if it hit max_iter without converging
+ self.converged = False
+ self.n_iter += t + 1
+ msg = ('\t'
+ 'Failed to converge after {0:d} iterations.'
+ ' Last convergence metric was {1:f}'
+ ' and convergence threshold {2:f}.'
+ ).format(self.n_iter, convergence_metric, tol)
+ logger.warn(msg)
 
  # Update the estimated variables
  self.beta0_ = beta[0]