def holdout_split(I, Xtr, ytr):
Xtr_t = utils.allbut(Xtr, I)
ytr_t = mulzip(Xtr_t, utils.allbut(ytr, I))
Xtr_t = np.vstack(Xtr_t)
clf = NBNN()
clf.fit(Xtr_t, ytr_t)
return [clf.predict(Xtr[i]).argmax() for i in I]
def train(self, patch_size, stride, n_jobs):
'''Train the model on all images in the train/ directory of the dataset'''
self._patch_size = patch_size
self._stride = stride
# Scan for all image files in dataset training dir.
self._dataset = Dataset(self._datadir + '/train')
# Extract masked features from all training images.
(Xtr, ytr) = extract_all_class_features(self._dataset,
n_jobs=n_jobs,
patch_size=patch_size,
stride=stride)
# Stack image features and labels into arrays (they were lists).
Xtr = np.vstack(Xtr)
ytr = np.hstack(ytr)
# Make a Naive Bayes NN classifier and train it.
self._clf = NBNN()
self._clf.fit(Xtr, ytr)
self._trained = True
class CNRecognizer:
'''
Simple class implementing a recognizer based on local colorname histograms.
This recognizer uses a Naive Bayes Nearest Neighbor classifier
over local colorname patches. It trains over all image/mask pairs
in the train/ directory of the dataset given as input.
Instantiate it like this:
clf = CNRecognizer(datadir, modelname)
where datasetdir is the directory of the dataset, and modelname is
the name of the serlialized model saved to the datasetdir/models
directory when the serialize() method is called.
NOTES:
------
The serialization code is hacky and untested. Treat it like
running with scissors.
'''
def __init__(self, datadir, modelname='CNRecognizer.pkl'):
self._datadir = datadir
self._modelname = modelname
self._trained = False
if os.path.exists(datadir + '/models/' + modelname):
inst = unpickle(datadir + '/models/' + modelname)
for att in dir(inst):
setattr(self, att, getattr(inst, att))
def train(self, patch_size, stride, n_jobs):
'''Train the model on all images in the train/ directory of the dataset'''
self._patch_size = patch_size
self._stride = stride
# Scan for all image files in dataset training dir.
self._dataset = Dataset(self._datadir + '/train')
# Extract masked features from all training images.
(Xtr, ytr) = extract_all_class_features(self._dataset,
n_jobs=n_jobs,
patch_size=patch_size,
stride=stride)
# Stack image features and labels into arrays (they were lists).
Xtr = np.vstack(Xtr)
ytr = np.hstack(ytr)
# Make a Naive Bayes NN classifier and train it.
self._clf = NBNN()
self._clf.fit(Xtr, ytr)
self._trained = True
def predict(self, image, mask):
'''
Method to classify an image based on masked colorname
patches. Handles feature extraction and masking.
IMPORTANT: image channels MUST range between 0 and 255 (as
opposed to 0.0 and 1.0)
'''
cns = extract_masked_cns(image, mask, self._stride, self._patch_size)
return self._clf.predict(cns)
def serialize(self):
'''Pickle the trained recognizer to the models/ directory of the dataset.'''
if not os.path.exists(self._datadir + '/models/'):
os.mkdir(self._datadir + '/models/')
pickle(self, self._datadir + '/models/' + self._modelname)
