def __init__(self,
X,
types_softmax,
probs_softmax,
magnitudes_softmax,
argmax=False):
# Ekin Dogus says he sampled the softmaxes, and has not used argmax
# We might still want to use argmax=True for the last predictions, to ensure
# the best solutions are chosen and make it deterministic.
transformations = get_transformations(X)
if argmax:
self.type = types_softmax.argmax()
t = transformations[self.type]
self.prob = probs_softmax.argmax() / (OP_PROBS - 1)
m = magnitudes_softmax.argmax() / (OP_MAGNITUDES - 1)
self.magnitude = m * (t[2] - t[1]) + t[1]
else:
self.type = np.random.choice(OP_TYPES, p=types_softmax)
t = transformations[self.type]
self.prob = np.random.choice(np.linspace(0, 1, OP_PROBS),
p=probs_softmax)
self.magnitude = np.random.choice(np.linspace(
t[1], t[2], OP_MAGNITUDES),
p=magnitudes_softmax)
self.transformation = t[0]
def __init__(self,
test=False,
batch_size=200,
one_hot=True,
shuffle=True,
sampling=1.0,
policy=None):
(x_train, y_train), (x_test, y_test) = load_data()
if test:
self._data = x_test
self._label = y_test
else:
self._data = x_train
self._label = y_train
if one_hot:
one_hot_labels = np.zeros((len(self._label), 10))
one_hot_labels[np.arange(len(self._label)),
self._label.flatten()] = 1
self._label = one_hot_labels
self._batch_size = batch_size
# shuffle before reduce the dataset
self._num_samples = len(self._data)
self._num_left = self._num_samples
self.shuffle = shuffle
if (shuffle):
self.shuffle_data()
# reduce dataset
assert sampling > 0, "Sampling must be greater than 0"
assert sampling <= 1, "Sampling must be less than 0"
self._data = self._data[:int(len(self._data) * sampling)]
self._label = self._label[:int(len(self._label) * sampling)]
self._num_samples = len(self._data)
self._num_left = self._num_samples
self._batch_pointer = 0
if policy == None:
#raise Exception("Policy not given. Need to provide policy")
print("No policy given. Using dataset without data augmentation.")
else:
self._policy = policy
# Note that for pairing sample, the img2 is chose randomly from the entire dataset
self._transformations = get_transformations(x_train)
self._apply_policy()
(Xtr, ytr), (Xts, yts) = datasets.cifar10.load_data()
elif dataset == 'cifar100':
(Xtr, ytr), (Xts, yts) = datasets.cifar100.load_data()
else:
raise Exception('Unknown dataset %s' % dataset)
if reduced:
ix = np.random.choice(len(Xtr), 4000, False)
Xtr = Xtr[ix]
ytr = ytr[ix]
ytr = utils.to_categorical(ytr)
yts = utils.to_categorical(yts)
return (Xtr, ytr), (Xts, yts)
(Xtr, ytr), (Xts, yts) = get_dataset('cifar10', False)
transformations = get_transformations(Xtr)
# Experiment parameters
LSTM_UNITS = 100
SUBPOLICIES = 5
SUBPOLICY_OPS = 2
OP_TYPES = 16
OP_PROBS = 11
OP_MAGNITUDES = 10
CHILD_BATCH_SIZE = 128
CHILD_BATCHES = len(Xtr) // CHILD_BATCH_SIZE
CHILD_EPOCHS = 10