for itera in range(nb_groups):
print("Processing network after {} increments\t".format(itera))
# Evaluation on cumul of classes or original classes
if is_cumul == 'cumul':
eval_groups = np.array(range(itera + 1))
else:
eval_groups = [0]
print("Evaluation on batches {} \t".format(eval_groups))
# Load the evaluation files
files_from_cl = []
for i in eval_groups:
files_from_cl.extend(files_valid[i])
inits, scores, label_batch, loss_class, file_string_batch, op_feature_map = utils_icarl.reading_data_and_preparing_network(
files_from_cl, gpu, itera, batch_size, train_path, labels_dic, mixing,
nb_groups, nb_cl, save_path)
with tf.Session(config=config) as sess:
# Launch the prefetch system
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
sess.run(inits)
# Evaluation routine
stat_hb1 = []
stat_icarl = []
stat_ncm = []
for i in range(int(np.ceil(len(files_from_cl) / batch_size))):
sc, l, loss, files_tmp, feat_map_tmp = sess.run([
scope='ResNet18',
sess=sess)
# Reset the graph
tf.reset_default_graph()
## Exemplars management part ##
#calculer le nombre exige d'exemplaires par classe
# nb_protos_cl = int(np.ceil(nb_proto * nb_groups * 1. / (itera + 1))) # Reducing number of exemplars for the previous classes
# nb_protos_cl = int(np.ceil(nb_proto * nb_groups * 1. / (itera + 1))) # Reducing number of exemplars for the previous classes
nb_protos_cl = int(math.ceil(nb_proto / (nb_cl * (itera + 1))))
#files from cl contains NEW DATA ONLY
files_from_cl = files_train[itera]
inits, scores, label_batch, loss_class, file_string_batch, op_feature_map = utils_icarl.reading_data_and_preparing_network(
images_mean, files_from_cl, device, itera, batch_size, nb_groups,
nb_cl, save_path)
with tf.Session(config=config) as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
void3 = sess.run(inits)
# Load the training samples of the current batch of classes in the feature space to apply the herding algorithm
Dtot, processed_files, label_dico = utils_icarl.load_class_in_feature_space(
files_from_cl, batch_size, scores, label_batch, loss_class,
file_string_batch, op_feature_map, sess)
processed_files = np.array([x.decode() for x in processed_files])
# Herding procedure : ranking of the potential exemplars
print('Exemplars selection starting ...')
# Loading the labels
labels_dic, label_names, validation_ground_truth = utils_data.parse_devkit_meta(devkit_path)
# Initialization
acc_list = np.zeros((nb_groups,3))
# Load the evaluation files
print("Processing network after {} increments\t".format(itera))
print("Evaluation on batches {} \t".format(eval_groups))
files_from_cl = []
for i in eval_groups:
files_from_cl.extend(files_valid[i])
inits,scores,label_batch,loss_class,file_string_batch,op_feature_map = utils_icarl.reading_data_and_preparing_network(files_from_cl, gpu, itera, batch_size, train_path, labels_dic, mixing, nb_groups, nb_cl, save_path)
with tf.Session(config=config) as sess:
# Launch the prefetch system
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
sess.run(inits)
# Evaluation routine
stat_hb1 = []
stat_icarl = []
stat_ncm = []
for i in range(int(np.ceil(len(files_from_cl)/batch_size))):
sc, l , loss,files_tmp,feat_map_tmp = sess.run([scores, label_batch,loss_class,file_string_batch,op_feature_map])
mapped_prototypes = feat_map_tmp[:,0,0,:]
#get data:
if is_cumul == 'cumul':
eval_groups = np.array(range(itera + 1))
else:
eval_groups = [0]
# Load the evaluation files
extract_files = []
for i in eval_groups:
extract_files.extend(files_valid[i])
destination_path = os.path.join(destination_dir,
'b' + str(itera + 1) + '_weight_bias.tf')
print('Saving stats in: ' + destination_path)
inits, scores, label_batch, loss_class, file_string_batch, op_feature_map = utils_icarl.reading_data_and_preparing_network(
images_mean, extract_files, device, itera, batch_size, S, P, save_path)
with tf.Session(config=config) as sess:
# Launch the prefetch system
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
sess.run(inits)
weights = [
v for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
if v.name.endswith('ResNet18/fc/W:0')
]
bias = [
v for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
if v.name.endswith('ResNet18/fc/b:0')
]
