lemur_util.py

import scipy.stats

import numpy as np
import pylab as plt


_IMG_SHAPE = (96, 96)

def distort(A, scale=.1):
  if scale == 0:
    return A
  return np.clip(A + np.random.normal(scale=scale, size=A.shape), 0, 1)

def dot(*Ms):
  n = len(Ms)
  shapes = map(lambda M: M.shape[0], Ms) + [Ms[-1].shape[1]]

  best = np.inf * np.ones((n, n))
  best_cut = np.zeros((n, n), int)

  for fr in range(n):
    best[fr][0] = 0

  for span in range(1, n):
    for fr in range(n - span):
      to = fr + span  # [fr, to)
      for cut in range(span):
        mi = fr + cut + 1  # [fr, mi) u [mi, to)
        candidate = best[fr][cut] + best[mi][span-cut-1] \
                    + shapes[fr] * shapes[mi] * shapes[to]
        if best[fr][span] > candidate:
          best[fr][span] = candidate
          best_cut[fr][span] = cut
  
  def go(fr, span):
    if span == 0:
      return Ms[fr]
    cut = best_cut[fr][span]
    mi = fr + cut + 1
    return go(fr, cut).dot(go(mi, span-cut-1))

  return go(0, n-1)

def L2(x):
  return np.sqrt(np.sum(x**2))

def L2_squared(x):
  return np.sum(x**2)

def plot_many(c, *faces):
  faces = np.array(faces)
  faces = faces.reshape(-1, *_IMG_SHAPE)
  n = faces.shape[0]
  r = (n / c) + (n % c != 0)

  f = plt.figure()
  for i, face in enumerate(faces):
    rr = i % r
    cc = i / r
    f.add_subplot(r, c, rr * c + cc + 1)
    plt.imshow(-face, cmap='Greys')
  plt.show()

def evaluate(X_training, Y_training, X_test, Y_test, method):
  model = method.fit(X_training, Y_training)
  predictions = model.predict(X_test)
  return predictions, L2_squared(predictions - Y_test)

def normalize(imgs):
  imgs = imgs.T
  imgs = imgs - np.min(imgs, axis=0)
  imgs = imgs / np.max(imgs, axis=0)
  return imgs.T