markdet.py 10.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
#!/usr/bin/env python
# vim: set fileencoding=utf-8 :


"""Trains a new MLP to perform pre-watershed marker detection

Usage: %(prog)s [-v...] [--samples=N] [--model=PATH] [--points=N] [--hidden=N]
                [--batch=N] [--iterations=N] <database> <protocol> <group>
       %(prog)s --help
       %(prog)s --version


Arguments:

  <database>  Name of the database to use for creating the model (options are:
André Anjos's avatar
André Anjos committed
16
              "fv3d" or "verafinger")
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
  <protocol>  Name of the protocol to use for creating the model (options
              depend on the database chosen)
  <group>     Name of the group to use on the database/protocol with the
              samples to use for training the model (options are: "train",
              "dev" or "eval")

Options:

  -h, --help             Shows this help message and exits
  -V, --version          Prints the version and exits
  -v, --verbose          Increases the output verbosity level. Using "-vv"
                         allows the program to output informational messages as
                         it goes along.
  -m PATH, --model=PATH  Path to the generated model file [default: model.hdf5]
  -s N, --samples=N      Maximum number of samples to use for training. If not
                         set, use all samples
  -p N, --points=N       Maximum number of samples to use for plotting
                         ground-truth and classification errors. The more
                         points, the less responsive the plot becomes
                         [default: 1000]
  -H N, --hidden=N       Number of neurons on the hidden layer of the
                         multi-layer perceptron [default: 5]
  -b N, --batch=N        Number of samples to use for every batch [default: 1]
  -i N, --iterations=N   Number of iterations to train the neural net for
                         [default: 2000]


Examples:

  Trains on the 3D Fingervein database:

     $ %(prog)s -vv fv3d central dev

  Saves the model to a different file, use only 100 samples:

    $ %(prog)s -vv -s 100 --model=/path/to/saved-model.hdf5 fv3d central dev

"""


import os
import sys
import schema
import docopt
import numpy
import skimage


def validate(args):
  '''Validates command-line arguments, returns parsed values

  This function uses :py:mod:`schema` for validating :py:mod:`docopt`
  arguments. Logging level is not checked by this procedure (actually, it is
  ignored) and must be previously setup as some of the elements here may use
  logging for outputing information.


  Parameters:

    args (dict): Dictionary of arguments as defined by the help message and
      returned by :py:mod:`docopt`


  Returns

    dict: Validate dictionary with the same keys as the input and with values
      possibly transformed by the validation procedure


  Raises:

    schema.SchemaError: in case one of the checked options does not validate.

  '''

André Anjos's avatar
André Anjos committed
92 93
  from .validate import check_model_does_not_exist, validate_protocol, \
      validate_group
94 95 96 97 98 99 100 101

  sch = schema.Schema({
    '--model': check_model_does_not_exist,
    '--samples': schema.Or(schema.Use(int), None),
    '--points': schema.Use(int),
    '--hidden': schema.Use(int),
    '--batch': schema.Use(int),
    '--iterations': schema.Use(int),
André Anjos's avatar
André Anjos committed
102 103 104
    '<database>': lambda n: n in ('fv3d', 'verafinger'),
    '<protocol>': validate_protocol(args['<database>']),
    '<group>': validate_group(args['<database>']),
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
    str: object, #ignores strings we don't care about
    }, ignore_extra_keys=True)

  return sch.validate(args)


def main(user_input=None):

  if user_input is not None:
    argv = user_input
  else:
    argv = sys.argv[1:]

  import pkg_resources

  completions = dict(
      prog=os.path.basename(sys.argv[0]),
      version=pkg_resources.require('bob.bio.vein')[0].version
      )

  args = docopt.docopt(
      __doc__ % completions,
      argv=argv,
      version=completions['version'],
      )

  try:
    from .validate import setup_logger
    logger = setup_logger('bob.bio.vein', args['--verbose'])
    args = validate(args)
  except schema.SchemaError as e:
    sys.exit(e)

André Anjos's avatar
André Anjos committed
138 139 140 141 142 143 144 145
  if args['<database>'] == 'fv3d':
    from ..configurations.fv3d import database as db
  elif args['<database>'] == 'verafinger':
    from ..configurations.verafinger import database as db
  else:
    raise schema.SchemaError('Database %s is not supported' % \
        args['<database>'])

146 147 148
  database_replacement = "%s/.bob_bio_databases.txt" % os.environ["HOME"]
  db.replace_directories(database_replacement)
  objects = db.objects(protocol=args['<protocol>'], groups=args['<group>'])
André Anjos's avatar
André Anjos committed
149 150
  if args['--samples'] is None:
    args['--samples'] = len(objects)
151 152 153 154

  from ..preprocessor.utils import poly_to_mask
  features = None
  target = None
155
  loaded = 0
156 157
  for k, sample in enumerate(objects):

158 159
    if args['--samples'] is not None and loaded >= args['--samples']:
      break
160 161
    path = sample.make_path(directory=db.original_directory,
        extension=db.original_extension)
162
    logger.info('Loading sample %d/%d (%s)...', loaded, len(objects), path)
163 164 165 166 167
    image = sample.load(directory=db.original_directory,
        extension=db.original_extension)
    if not (hasattr(image, 'metadata') and 'roi' in image.metadata):
      logger.info('Skipping sample (no ROI)')
      continue
168
    loaded += 1
169 170 171 172 173 174 175 176 177 178 179 180 181

    # copy() required by skimage.util.shape.view_as_windows()
    image = image.copy().astype('float64') / 255.
    windows = skimage.util.shape.view_as_windows(image, (3,3))

    if features is None and target is None:
      features = numpy.zeros(
          (args['--samples']*windows.shape[0]*windows.shape[1],
            windows.shape[2]*windows.shape[3]+2), dtype='float64')
      target = numpy.zeros(args['--samples']*windows.shape[0]*windows.shape[1],
          dtype='bool')

    mask = poly_to_mask(image.shape, image.metadata['roi'])
André Anjos's avatar
André Anjos committed
182

183 184 185
    mask = mask[1:-1, 1:-1]
    for y in range(windows.shape[0]):
      for x in range(windows.shape[1]):
186 187
        idx = ((loaded-1)*windows.shape[0]*windows.shape[1]) + \
            (y*windows.shape[1]) + x
188 189 190 191 192
        features[idx,:-2] = windows[y,x].flatten()
        features[idx,-2] = y+1
        features[idx,-1] = x+1
        target[idx] = mask[y,x]

193 194 195 196
  # if number of loaded samples is smaller than expected, clip features array
  features = features[:loaded*windows.shape[0]*windows.shape[1]]
  target = target[:loaded*windows.shape[0]*windows.shape[1]]

197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
  # normalize w.r.t. dimensions
  features[:,-2] /= image.shape[0]
  features[:,-1] /= image.shape[1]

  target_float = target.astype('float64')
  target_float[~target] = -1.0
  target_float = target_float.reshape(len(target), 1)
  positives = features[target]
  negatives = features[~target]
  logger.info('There are %d samples on input dataset', len(target))
  logger.info('  %d are negatives', len(negatives))
  logger.info('  %d are positives', len(positives))

  import bob.learn.mlp

  # by default, machine uses hyperbolic tangent output
  machine = bob.learn.mlp.Machine((features.shape[1], args['--hidden'], 1))
  machine.randomize() #initialize weights randomly
  loss = bob.learn.mlp.SquareError(machine.output_activation)
  train_biases = True
  trainer = bob.learn.mlp.RProp(args['--batch'], loss, machine, train_biases)
  trainer.reset()
  shuffler = bob.learn.mlp.DataShuffler([negatives, positives],
      [[-1.0], [+1.0]])

  # start cost
  output = machine(features)
  cost = loss.f(output, target_float)
  logger.info('[initial] MSE = %g', cost.mean())

  # trains the network until the error is near zero
  for i in range(args['--iterations']):
    try:
      _feats, _tgts = shuffler.draw(args['--batch'])
      trainer.train(machine, _feats, _tgts)
      logger.info('[%d] MSE = %g', i, trainer.cost(_tgts))
    except KeyboardInterrupt:
      print() #avoids the ^C line
      logger.info('Gracefully stopping training before limit (%d iterations)',
236
          args['--batch'])
237 238 239 240 241 242 243 244 245
      break

  # describe errors
  neg_output = machine(negatives)
  pos_output = machine(positives)
  neg_errors = neg_output >= 0
  pos_errors = pos_output < 0
  hter_train = ((sum(neg_errors) / float(len(negatives))) + \
      (sum(pos_errors)) / float(len(positives))) / 2.0
André Anjos's avatar
André Anjos committed
246
  logger.info('Training set HTER: %.2f%%', 100*hter_train)
247 248 249 250 251 252 253 254
  logger.info('  Errors on negatives: %d / %d', sum(neg_errors), len(negatives))
  logger.info('  Errors on positives: %d / %d', sum(pos_errors), len(positives))

  threshold = 0.8
  neg_errors = neg_output >= threshold
  pos_errors = pos_output < -threshold
  hter_train = ((sum(neg_errors) / float(len(negatives))) + \
      (sum(pos_errors)) / float(len(positives))) / 2.0
André Anjos's avatar
André Anjos committed
255 256
  logger.info('Training set HTER (threshold=%g): %.2f%%', threshold,
      100*hter_train)
257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311
  logger.info('  Errors on negatives: %d / %d', sum(neg_errors), len(negatives))
  logger.info('  Errors on positives: %d / %d', sum(pos_errors), len(positives))
  # plot separation threshold
  import matplotlib.pyplot as plt
  from mpl_toolkits.mplot3d import Axes3D

  # only plot N random samples otherwise it makes it too slow
  N = numpy.random.randint(min(len(negatives), len(positives)),
      size=min(len(negatives), len(positives), args['--points']))

  fig = plt.figure()

  ax = fig.add_subplot(211, projection='3d')
  ax.scatter(image.shape[1]*negatives[N,-1], image.shape[0]*negatives[N,-2],
      255*negatives[N,4], label='negatives', color='blue', marker='.')
  ax.scatter(image.shape[1]*positives[N,-1], image.shape[0]*positives[N,-2],
      255*positives[N,4], label='positives', color='red', marker='.')
  ax.set_xlabel('Width')
  ax.set_xlim(0, image.shape[1])
  ax.set_ylabel('Height')
  ax.set_ylim(0, image.shape[0])
  ax.set_zlabel('Intensity')
  ax.set_zlim(0, 255)
  ax.legend()
  ax.grid()
  ax.set_title('Ground Truth')
  plt.tight_layout()

  ax = fig.add_subplot(212, projection='3d')
  neg_plot = negatives[neg_output[:,0]>=threshold]
  pos_plot = positives[pos_output[:,0]<-threshold]
  N = numpy.random.randint(min(len(neg_plot), len(pos_plot)),
      size=min(len(neg_plot), len(pos_plot), args['--points']))
  ax.scatter(image.shape[1]*neg_plot[N,-1], image.shape[0]*neg_plot[N,-2],
      255*neg_plot[N,4], label='negatives', color='red', marker='.')
  ax.scatter(image.shape[1]*pos_plot[N,-1], image.shape[0]*pos_plot[N,-2],
      255*pos_plot[N,4], label='positives', color='blue', marker='.')
  ax.set_xlabel('Width')
  ax.set_xlim(0, image.shape[1])
  ax.set_ylabel('Height')
  ax.set_ylim(0, image.shape[0])
  ax.set_zlabel('Intensity')
  ax.set_zlim(0, 255)
  ax.legend()
  ax.grid()
  ax.set_title('Classifier Errors')
  plt.tight_layout()

  print('Close plot window to save model and end program...')
  plt.show()
  import bob.io.base
  h5f = bob.io.base.HDF5File(args['--model'], 'w')
  machine.save(h5f)
  del h5f
  logger.info('Saved MLP model to %s', args['--model'])