Change PseudoData according to new teacher API

codebase/datasets.py

@@ -8,7 +8,22 @@
 from itertools import izip
 from utils import u2t, s2t
 
-PATH = '/mnt/ilcompf5d0/user/rshu/data'
+PATH = '/home/ruishu/data'
+
+def get_info(domain_id, domain):
+ train, test = domain.train, domain.test
+ print '{} info'.format(domain_id)
+ Y_shape = None if train.labels is None else train.labels.shape
+ print 'Train X/Y shapes: {}, {}'.format(train.images.shape, Y_shape)
+ print 'Train X min/max/cast: {}, {}, {}'.format(
+ train.images.min(),
+ train.images.max(),
+ train.cast)
+ print 'Test shapes: {}, {}'.format(test.images.shape, test.labels.shape)
+ print 'Test X min/max/cast: {}, {}, {}\n'.format(
+ test.images.min(),
+ test.images.max(),
+ test.cast)
 
 class Data(object):
  def __init__(self, images, labels=None, labeler=None, cast=False):
@@ -38,51 +53,38 @@ def next_batch(self, bs):
  y = self.labeler(x) if self.labels is None else self.labels[idx]
  return x, y
 
-
 class Mnist(object):
  def __init__(self, shape=(32, 32, 3)):
  """MNIST domain train/test data
 
  shape - (3,) HWC info
  """
+ raise NotImplementedError('Double check max/min')
  print "Loading MNIST"
- data = np.load(os.path.join(PATH, 'mnist_784.npz'))
+ train = loadmat(os.path.join(PATH, 'mnist32_train.mat'))
+ test = loadmat(os.path.join(PATH, 'mnist32_test.mat'))
 
- trainx = data['x_train']
- trainy = data['y_train']
+ trainx = train['X']
+ trainy = train['y']
  trainy = np.eye(10)[trainy].astype('float32')
 
- testx = data['x_test']
- testy = data['y_test'].astype('int')
+ testx = test['X']
+ testy = test['y'].astype('int')
  testy = np.eye(10)[testy].astype('float32')
 
- trainx = self.resize_cast(trainx, shape)
- testx = self.resize_cast(testx, shape)
+ trainx = trainx.reshape(-1, 32, 32, 1).astype('float32')
+ testx = testx.reshape(-1, 32, 32, 1).astype('float32')
 
  self.train = Data(trainx, trainy)
  self.test = Data(testx, testy)
 
- @staticmethod
- def resize_cast(x, shape):
- H, W, C = shape
- x = x.reshape(-1, 28, 28)
-
- resized_x = np.empty((len(x), H, W), dtype='float32')
- for i, img in enumerate(x):
- # imresize returns uint8
- resized_x[i] = u2t(scipy.misc.imresize(img, (H, W)))
-
- # Retile to make RGB
- resized_x = resized_x.reshape(-1, H, W, 1)
- resized_x = np.tile(resized_x, (1, 1, 1, C))
- return resized_x
-
 class Mnistm(object):
  def __init__(self, shape=(28, 28, 3), seed=0, npc=None):
  """Mnist-M domain train/test data
 
  shape - (3,) HWC info
  """
+ raise NotImplementedError('Did not change mnistm yet')
  print "Loading MNIST-M"
  data = pkl.load(open(os.path.join(PATH, 'mnistm_data.pkl')))
  labels = pkl.load(open(os.path.join(PATH, 'mnistm_labels.pkl')))
@@ -263,17 +265,17 @@ def __init__(self):
  self.test = Data(testx, testy, cast=True)
 
 class PseudoData(object):
- def __init__(self, domain_id, domain, M):
+ def __init__(self, domain_id, domain, teacher):
  """Variable domain with psuedolabeler
 
  domain_id - (str) {Mnist,Mnistm,Svhn,etc}
  domain - (obj) {Mnist,Mnistm,Svhn,etc}
- M - (TensorDict) Model used for pseudolabeling
+ teacher - (fn) Teacher model used for pseudolabeling
  """
  print "Constructing pseudodata"
  cast = 'mnist' not in domain_id
  print "{} uses casting: {}".format(domain_id, cast)
- labeler = tb.function(M.sess, [M.test_x], M.back_y)
+ labeler = teacher
 
  self.train = Data(domain.train.images, labeler=labeler, cast=cast)
  self.test = Data(domain.test.images, labeler=labeler, cast=cast)