This script illustrates the application of a trained image classifier to a new set of images in Glimpse.

apply-trained-classifier.py

# This script illustrates how to apply a trained image classifier to a new set
# of images.
# Author: Mick Thomure
# Date: 12/29/2013  

from glimpse.experiment import *
from glimpse.pools import *

def PredictImageClasses(exp, images, pool):
  """Apply existing model to a new set of images.

  The experiment object "exp" should have a model and a trained classifier.

  """
  ev = exp.evaluation[0]
  model = exp.extractor.model
  # Compute model activity for each image.
  images = map(model.MakeState, images)
  builder = Callback(BuildLayer, model, ev.layers, save_all=False)
  states = pool.map(builder, images)
  # Create feature vectors from model activity.
  ftrs = ExtractFeatures(ev.layers, states)
  # Use the (existing) classifier to predict the class of each image.
  lbls = ev.results.classifier.predict(ftrs)
  # Convert the numeric label to a class name.
  classes = exp.corpus.class_names[lbls]
  return classes

def main(corpus, images):
  # Number of imprinted prototypes.
  num_prototypes = 10
  # Worker pool used for imprinting and for computing feature vectors. Make it
  # a MulticorePool() if you want.
  pool = MulticorePool()
  # Create the experiment, and train the model.
  print "Training on %s" % corpus
  exp = ExperimentData()
  SetModel(exp)
  SetCorpus(exp, corpus)
  MakePrototypes(exp, num_prototypes, 'imprint')
  ComputeActivation(exp, Layer.C2, pool)
  TrainAndTestClassifier(exp, Layer.C2)
  # Apply the model to non-training images.
  classes = PredictImageClasses(exp, images, pool)
  print "Predicting class for previously unseen images:"
  print "\n".join("%s %s" % x for x in zip(images,classes))

# Labeled images for training/testing.
corpus = 'glimpse/corpora/data/easy'
# Another set of images to which the model is applied.
from glob import glob
images = glob('glimpse/corpora/data/moderate/circle/*')
main(corpus, images)