}
# Create Tensorboard Filewriter & Save Results
[fw, log_path_dt] = utils.create_tensorboard(sess, log_path)
# Start Sequential Learning
acc_pre = []
acc_curr = []
for task in range(5):
# Reinitialize optimizers
sess.run(
tf.variables_initializer(model.opt_disc.variables() +
model.opt_recon.variables() +
model.opt_fool.variables()))
# Load data for training
data = datasets.split_mnist([2 * task], [2 * task + 1])
#data = datasets.split_fashion_mnist([2 * task], [2 * task + 1])
[train_data, train_labels] = data.get_train_samples()
train_data = train_data / 255.0
sess.run(iterator.initializer,
feed_dict={
data_ph: train_data,
labels_ph: train_labels,
batch_size_ph: batch_size,
shufflebuffer_ph: train_data.shape[0],
epochs_ph: epochs
})
# Train model
i = 0
while True:
try:
"learning_rate": learning_rate,
"num_classes": num_classes,
"N_plot": N_plot,
"log_path": log_path}
# Create Tensorboard Filewriter & Save Results
[fw,log_path_dt]=utils.create_tensorboard(sess,log_path)
# Start Sequential Learning
acc_pre = []
acc_curr = []
for task in range(5):
# Reinitialize optimizers
sess.run(tf.variables_initializer(model.opt_disc.variables() + model.opt_recon.variables() + model.opt_fool.variables()))
# Load data for training
data = datasets.split_mnist(np.arange(2 * (task + 1)), [])
#data = datasets.split_mnist(list(range(0, 2*(task+1))),[])
#data = datasets.split_fashion_mnist(list(range(0, 2*(task+1))),[])
[train_data, train_labels] = data.get_train_samples()
train_data = train_data / 255.0
sess.run(iterator.initializer,feed_dict={data_ph: train_data, labels_ph: train_labels, batch_size_ph: batch_size,shufflebuffer_ph: train_data.shape[0],epochs_ph: epochs})
# Train model
i=0
while True:
try:
[_, _, _, loss, summaries] = sess.run([model.update_disc, model.update_fool, model.update_recon, model.loss, model.summaries],
feed_dict={model.batch_size: batch_size, model.learning_rate: learning_rate,model.b_replay: False,model.repl_batch_size: batch_size})
i += 1
fw.add_summary(summaries, i)
if (i%100 == 0):
print("Task{}\tIteration: {}\tloss: {:.5}".format(task,i,loss))