#files protoset est une liste de 1000 listes, chaque liste contient les exemplaire (paths) d'une classe
files_from_cl = files_train[itera][:]
for i in range(itera * nb_cl): # pour toutes les classes anterieures
# nb_protos_cl = int( np.ceil(nb_proto * nb_groups * 1. / itera))
#pour batch courant:
nb_protos_cl = int(math.ceil(nb_proto / (nb_cl * itera)))
tmp_var = files_protoset[
i] #recuperer la listes des exemplaires de cette classe
files_from_cl += tmp_var[0:min(
len(tmp_var), nb_protos_cl
)] #ajouter les exemplaires (paths) disponibles a la liste courante de train
#files from cl now contains paths of new+old data (imbalanced)
## Import the data reader ##
image_train, label_train = utils_data.read_data(
files_from_cl=files_from_cl)
image_batch, label_batch_0 = tf.train.batch([image_train, label_train],
batch_size=batch_size,
num_threads=8)
label_batch = tf.one_hot(label_batch_0, nb_groups * nb_cl)
## Define the objective for the neural network ##
if itera == 0:
# No distillation
variables_graph, variables_graph2, scores, scores_stored = utils_icarl.prepare_networks(
images_mean, device, image_batch, nb_cl, nb_groups)
# Define the objective for the neural network: 1 vs all cross_entropy
with tf.device(device):
scores = tf.concat(scores, 0)
l2_reg = wght_decay * tf.reduce_sum(
print('Batch of classes number {0} arrives ...'.format(itera + 1))
# Adding the stored exemplars to the training set
if itera == 0:
files_from_cl = files_train[itera]
else:
files_from_cl = files_train[itera][:]
for i in range(itera * nb_cl):
nb_protos_cl = int(np.ceil(
nb_proto * nb_groups * 1. /
itera)) # Reducing number of exemplars of the previous classes
tmp_var = files_protoset[i]
files_from_cl += tmp_var[0:min(len(tmp_var), nb_protos_cl)]
## Import the data reader ##
image_train, label_train = utils_data.read_data(
train_path, labels_dic, mixing, files_from_cl=files_from_cl)
image_batch, label_batch_0 = tf.train.batch([image_train, label_train],
batch_size=batch_size,
num_threads=8)
label_batch = tf.one_hot(label_batch_0, nb_groups * nb_cl)
## Define the objective for the neural network ##
if itera == 0:
# No distillation
variables_graph, variables_graph2, scores, scores_stored = utils_icarl.prepare_networks(
gpu, image_batch, nb_cl, nb_groups)
# Define the objective for the neural network: 1 vs all cross_entropy
with tf.device('/gpu:' + gpu):
scores = tf.concat(scores, 0)
l2_reg = wght_decay * tf.reduce_sum(
for itera in range(nb_groups):
# Files to load : training samples + protoset
print('Batch of classes number {0} arrives ...'.format(itera+1))
# Adding the stored exemplars to the training set
if itera == 0:
files_from_cl = files_train[itera]
else:
files_from_cl = files_train[itera][:]
for i in range(itera*nb_cl):
nb_protos_cl = int(np.ceil(nb_proto*nb_groups*1./itera)) # Reducing number of exemplars of the previous classes
tmp_var = files_protoset[i]
files_from_cl += tmp_var[0:min(len(tmp_var),nb_protos_cl)]
## Import the data reader ##
image_train, label_train = utils_data.read_data(train_path, labels_dic, mixing, files_from_cl=files_from_cl)
image_batch, label_batch_0 = tf.train.batch([image_train, label_train], batch_size=batch_size, num_threads=8)
label_batch = tf.one_hot(label_batch_0,nb_groups*nb_cl)
## Define the objective for the neural network ##
if itera == 0:
# No distillation
variables_graph,variables_graph2,scores,scores_stored = utils_icarl.prepare_networks(gpu,image_batch, nb_cl, nb_groups)
# Define the objective for the neural network: 1 vs all cross_entropy
with tf.device('/gpu:0'):
scores = tf.concat(scores,0)
l2_reg = wght_decay * tf.reduce_sum(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES, scope='ResNet18'))
loss_class = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=label_batch, logits=scores))
loss = loss_class + l2_reg
learning_rate = tf.placeholder(tf.float32, shape=[])