1、先构建网络，定义一些变量
2、构建损失函数
3、构建循环网络
4、筛选保留集样本
5、先实现残差网络 再实现增量学习
6、实现简单的残差网络
"""
# Create neural network model
print('Run {0} starting ...'.format(itera))
print("Building model and compiling functions...")
image_train, label_train, image_test, label_test = utils_data.load_data(
Cifar_train_file, Cifar_test_file)
#next batch
image_batch, label_batch_0, file_protoset_batch = utils_data.Prepare_train_data_batch(
image_train, label_train, files_protoset, itera, order, nb_cl,
batch_size)
label_batch = tf.one_hot(label_batch_0, 100)
#初次训练
if itera == 0:
#不需要蒸馏
variables_graph, variables_graph2, scores, scores_stored = utils_cifar.prepareNetwork(
gpu, image_batch, itera)
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='ResNet34'))
loss_class = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=label_batch,
logits=scores))
5、先实现残差网络 再实现增量学习
6、实现简单的残差网络
"""
# Select the order for the class learning
order = np.arange(100)
np.random.shuffle(order)
np.save('order', order) #### 存储样本顺序的序列
# Create neural network model
print('Run {0} starting ...'.format(itera))
print("Building model and compiling functions...")
image_train, label_train, image_test, label_test = utils_data.load_data(
Cifar_train_file, Cifar_test_file)
#next batch
image_batch, label_batch_0 = utils_data.Prepare_train_data_batch(
image_train, label_train, itera, order, nb_cl, batch_size)
label_batch = tf.one_hot(label_batch_0, 100)
variables_graph, variables_graph2, scores, scores_stored = utils_cifar.prepareNetwork(
gpu, image_batch)
op_assign = [(variables_graph2[i]).assign(variables_graph[i])
for i in range(len(variables_graph))]
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='ResNet34'))
loss_class = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=label_batch,
logits=scores))
loss = loss_class + l2_reg