def run(self):
n_true = self.batch_size - self.n_pseudo
if self.n_pseudo:
test_idx = np.arange(len(self.pseudo_images))
while True:
# send images through queue in batches
for i in range(0, len(self.images), n_true):
# read real train images
batch = list(map(self._transform, self.images[i:i + n_true]))
if self.n_pseudo:
pseudo_idx = self.rng.choice(test_idx, self.n_pseudo)
batch.extend(
list(
map(self._transform,
self.pseudo_images.iloc[pseudo_idx])))
labels = np.vstack(
(to_ordinal(self.labels[i:i + n_true].values),
self.pseudo_labels[pseudo_idx].astype(
theano.config.floatX)))
else:
labels = to_ordinal(self.labels[i:i + n_true].values)
batch = np.vstack(batch)
self.outqueue.put((np.rollaxis(batch, 3, 1), labels))
#self.outqueue.put((batch.reshape(len(batch), 1 , 128, 128), labels))
# shuffle images at epoch end do this for trainig set only
if self.augment:
shuffle_idx = self.rng.permutation(len(self.images))
self.images = self.images.iloc[shuffle_idx]
self.labels = self.labels.iloc[shuffle_idx]
# signal end of epoch
self.outqueue.put(None)
def crossover_ordinal_1p(parent_a, parent_b):
point = randint(1, len(parent_a) - 1)
parent_a = to_ordinal(parent_a)
parent_b = to_ordinal(parent_b)
child_a = parent_a[:point]
child_a += parent_b[point:]
child_b = parent_b[:point]
child_b += parent_a[point:]
return from_ordinal(child_a), from_ordinal(child_b)
def _image_iterator(self, image_list, labels=None, transform=False):
# allocate an array for images
images = np.zeros(
(self.batch_size, self.image_size, self.image_size, 3),
dtype=theano.config.floatX)
n_images = len(image_list)
n_chunks = int(np.ceil(n_images * 1. / self.batch_size))
for chunk in range(n_chunks):
# prepare a slice of images to read during this pass
chunk_end = (chunk + 1) * self.batch_size
chunk_slice = slice(chunk * self.batch_size, chunk_end)
# read a chunk of images
for i, image in enumerate(image_list[chunk_slice]):
if transform:
images[i, ...] = self._transform(image)
else:
images[i, ...] = imread(image) / 255.
if self.norm:
images = self.normalize(images)
# change axis order (see comments in valid_gen function) and yield images with labels
if labels is not None:
# transform labels to a collumn, but first we need to add a new axis
#print labels[chunk_slice].values.astype(np.float32).reshape(chunk_slice, 1)
yield np.rollaxis(images, 3,
1), to_ordinal(labels[chunk_slice].values)
else:
yield np.rollaxis(images, 3, 1)