def __init__(self, X, Y, training=True, global_step=None):
self.shapeX = X.get_shape().as_list()
self.shapeY = Y.get_shape().as_list()
# if data dype is not float32, we assume that there is no preprocess
if X.dtype != tf.float32: # Here go CIFAR10 and SVHN (Iban)
X = tf.cast(X, tf.float32)
print('Input data dype is not float32, perform simple preprocess [0,255]->[-1,1]')
X = X / 127.5 - 1
else:
print('Input data type is float32, we assume they are preprocessed already') # MNIST X is float32, no additional preprocessing needed (Iban)
# quantize inputs
self.H = [X]
self._QA(X)
self.Y = Y
self.lossFunc = Option.lossFunc
self.L2 = Option.L2
self.initializer = myInitializer.variance_scaling_initializer(
factor=1.0, mode='FAN_IN', uniform=True)
self.is_training = training
self.GPU = Option.GPU
self.W = []
self.W_q = []
self.W_clip_op = []
self.W_q_op = []
def __init__(self, X, Y, training=True, global_step=None):
self.shapeX = X.get_shape().as_list()
self.shapeY = Y.get_shape().as_list()
# if data dype is not float32, we assume that there is no preprocess
if X.dtype != tf.float32:
X = tf.cast(X, tf.float32)
# add simple preprocess
if Option.dataSet == 'MNIST':
X = X / 256.
else:
X = X / 128. - 1
self.H = [X]
# quantize inputs
self._QA(X)
self.Y = Y
self.use_batch_norm = Option.use_batch_norm
self.lossFunc = Option.lossFunc
self.L2 = Option.L2
self.initializer = myInitializer.variance_scaling_initializer(
factor=1.0, mode='FAN_IN', uniform=True)
self.is_training = training
self.GPU = Option.GPU
self.W = []
self.W_q = []
self.W_clip_op = []
self.W_q_op = []
