def initialize():
# Different losses need different method to create batches
if FLAGS.LossType == "Contrastive_Loss":
method = "pair"
elif FLAGS.LossType == "NpairLoss" or FLAGS.LossType == "AngularLoss" or FLAGS.LossType == "NCA_loss":
method = "n_pairs_mc"
elif FLAGS.LossType == "Triplet":
method = 'triplet'
else:
method = "clustering"
print("method: " + method)
# Create the stream of datas from dataset
streams = data_provider.get_streams(FLAGS.batch_size,
FLAGS.dataSet,
method,
crop_size=FLAGS.default_image_size)
regularizer = layers.l2_regularizer(FLAGS.Regular_factor)
if FLAGS.SaveVal:
nn_Ops.create_path(_time)
summary_writer = tf.summary.FileWriter(LOGDIR)
return streams, summary_writer
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.compat.v1.Session(config=config)
sess.as_default()
# Creating Streams from the dataset
streams = data_provider.get_streams(BATCH_SIZE, DATASET, crop_size=IMAGE_SIZE)
stream_train, stream_train_eval, stream_test = streams
LEN_TRAIN = stream_train.data_stream.dataset.num_examples
MAX_ITER = int(LEN_TRAIN / BATCH_SIZE)
# check system time
_time = time.strftime('%m-%d-%H-%M', time.localtime(time.time()))
LOGDIR = './tensorboard_log/' + DATASET + '/' + _time + '/'
nn_Ops.create_path(_time)
# tfd = tfp.distributions
# prior = tfd.Independent(tfd.Normal(loc=tf.zeros(EMBEDDING_SIZE), scale=1),reinterpreted_batch_ndims=1)
def samplingGaussian(z_mean, z_log_var):
"""Reparameterization trick by sampling from an isotropic unit Gaussian.
# Arguments
args (tensor): mean and log of variance of Q(z|X)
# Returns
z (tensor): sampled latent vector
"""
batch = K.shape(z_mean)[0]
dim = K.int_shape(z_mean)[1]
image_mean = np.array([123, 117, 104], dtype=np.float32) # RGB
# To shape the array image_mean to (1, 1, 1, 3) => three channels
image_mean = image_mean[None, None, None, [2, 1, 0]]
neighbours = [1, 2, 4, 8, 16, 32]
products_neighbours = [1, 10, 1000]
############## DATASET GENERATOR #############################
streams = data_provider.get_streams(BATCH_SIZE,
DATASET,
"n_pairs_mc",
crop_size=IMAGE_SIZE)
stream_train, stream_train_eval, stream_test = streams
LOGDIR = './tensorboard_log/' + DATASET + '/' + time.strftime(
'%m-%d-%H-%M', time.localtime(time.time())) + '/'
nn_Ops.create_path(time.strftime('%m-%d-%H-%M', time.localtime(time.time())))
tfd = tfp.distributions
prior = tfd.Independent(tfd.Normal(loc=tf.zeros(EMBEDDING_SIZE), scale=1),
reinterpreted_batch_ndims=1)
def mysampling(z_mean, z_log_var):
"""Reparameterization trick by sampling from an isotropic unit Gaussian.
# Arguments
args (tensor): mean and log of variance of Q(z|X)
# Returns
z (tensor): sampled latent vector
"""
batch = K.shape(z_mean)[0]
