def sample(self, nb_classes, nb_samples_per_class,
nb_samples_per_class_test):
sampled_characters = random.sample(self.data.keys(),
nb_classes) # list of keys
labels_and_images_support = []
labels_and_images_test = []
for (k, char) in enumerate(sampled_characters):
deg = random.sample(range(4), 1)[0]
_imgs = self.data[char]
_ind = random.sample(
range(len(_imgs)),
nb_samples_per_class + nb_samples_per_class_test)
_ind_tr = _ind[:nb_samples_per_class]
labels_and_images_support.extend([(k, rotate_right(_imgs[i],
deg).flatten())
for i in _ind_tr])
_ind_test = _ind[nb_samples_per_class:]
labels_and_images_test.extend([(k, rotate_right(_imgs[i],
deg).flatten())
for i in _ind_test])
random.shuffle(labels_and_images_support)
labels_tr, images_tr = zip(*labels_and_images_support)
random.shuffle(labels_and_images_test)
labels_test, images_test = zip(*labels_and_images_test)
return images_tr, labels_tr, images_test, labels_test
def sample(self, nb_classes, nb_samples_per_class):
sampled_characters = random.sample(self.data.keys(), nb_classes) # list of keys
labels_and_images = []
for (k, char) in enumerate(sampled_characters):
deg = random.sample(range(4), 1)[0]
_imgs = self.data[char]
_ind = random.sample(range(len(_imgs)), nb_samples_per_class)
labels_and_images.extend([(k, rotate_right(_imgs[i], deg).flatten()) for i in _ind])
random.shuffle(labels_and_images)
sequence_length = len(labels_and_images)
labels, images = zip(*labels_and_images)
return images, labels
def sample(self, nb_classes, nb_samples_per_class):
sampled_characters = random.sample(self.data.keys(),
nb_classes) # list of keys
labels_and_images = []
for (k, char) in enumerate(sampled_characters):
deg = random.sample(range(4), 1)[0]
_imgs = self.data[char]
_ind = random.sample(range(len(_imgs)), nb_samples_per_class)
labels_and_images.extend([(k, rotate_right(_imgs[i],
deg).flatten())
for i in _ind])
random.shuffle(labels_and_images)
sequence_length = len(labels_and_images)
labels, images = zip(*labels_and_images)
return images, labels
def next(self):
if (self.max_iter is None) or (self.num_iter < self.max_iter):
labels_and_images_support = []
labels_and_images_test = []
self.num_iter += 1
sampled_characters = random.sample(self.data.keys(),
self.nb_classes) # list of keys
for _ in range(self.batchsize):
support_set = []
for k, char in enumerate(sampled_characters):
deg = random.sample(range(4), 1)[0]
_imgs = self.data[char]
_ind = random.sample(
range(len(_imgs)), self.nb_samples_per_class +
self.nb_samples_per_class_test)
_ind_tr = _ind[:self.nb_samples_per_class]
if self.augment:
support_set.extend([(k,
self.xp.array(
rotate_right(_imgs[i],
deg).flatten()))
for i in _ind_tr])
else:
support_set.extend([(k,
self.xp.array(_imgs[i].flatten()))
for i in _ind_tr])
_ind_test = _ind[self.nb_samples_per_class:]
if self.augment:
labels_and_images_test.extend([
(k,
self.xp.array(
rotate_right(_imgs[i], deg).flatten()))
for i in _ind_test
])
else:
labels_and_images_test.extend([
(k, self.xp.array(_imgs[i].flatten()))
for i in _ind_test
])
random.shuffle(support_set)
labels_tr, images_tr = zip(*support_set)
labels_and_images_support.append((images_tr, labels_tr))
_images, _labels = zip(*labels_and_images_support)
# images_tr = [
# self.xp.concatenate(map(lambda x: x.reshape((1,-1)), _img), axis=0)
# for _img in zip(*_images)]
labels_tr = [_lbl for _lbl in zip(*_labels)]
random.shuffle(labels_and_images_test)
labels_test, images_test = zip(*labels_and_images_test)
images_test = self.xp.concatenate(
[img.reshape((1, -1)) for img in images_test], axis=0)
return (self.num_iter - 1), (images_tr, labels_tr, images_test,
labels_test)
else:
raise StopIteration()
def next(self):
if (self.max_iter is None) or (self.num_iter < self.max_iter):
labels_and_images_support = []
labels_and_images_test = []
self.num_iter += 1
sampled_characters = random.sample(self.data.keys(),
self.nb_classes) # list of keys
for _ in xrange(self.batchsize):
support_set = []
for k, char in enumerate(sampled_characters):
deg = random.sample(range(4), 1)[0]
_imgs = self.data[char]
_ind = random.sample(
range(len(_imgs)), self.nb_samples_per_class +
self.nb_samples_per_class_test)
_ind_tr = _ind[:self.nb_samples_per_class]
if self.augment:
support_set.extend([
(k, self.xp.array(rotate_right(_imgs[i], deg)))
for i in _ind_tr
])
else:
support_set.extend([(k, self.xp.array(_imgs[i]))
for i in _ind_tr])
_ind_test = _ind[self.nb_samples_per_class:]
if self.augment:
labels_and_images_test.extend([
(k, self.xp.array(rotate_right(_imgs[i], deg)))
for i in _ind_test
])
else:
labels_and_images_test.extend([
(k, self.xp.array(_imgs[i])) for i in _ind_test
])
random.shuffle(support_set)
labels_tr, images_tr = zip(*support_set)
labels_and_images_support.append((images_tr, labels_tr))
_images, _labels = zip(*labels_and_images_support)
images_tr = [
# NOTE: be careful here!
# I had to permute the dimensions due the output image format from preporcessing
# please make sure your img dimensions are correct
self.xp.concatenate(map(
lambda x: np.transpose(x, (2, 0, 1)).reshape(
(1, 3, 84, 84)), _img),
axis=0) for _img in zip(*_images)
]
labels_tr = [_lbl for _lbl in zip(*_labels)]
random.shuffle(labels_and_images_test)
labels_test, images_test = zip(*labels_and_images_test)
# NOTE: be careful here!
# I had to permute the dimensions due the output image format from preporcessing
# please make sure your img dimensions are correct
images_test = self.xp.concatenate([
np.transpose(img, (2, 0, 1)).reshape((1, 3, 84, 84))
for img in images_test
],
axis=0)
return (self.num_iter - 1), (images_tr, labels_tr, images_test,
labels_test)
else:
raise StopIteration()