# data preprocessing
x, y, x_, y_, xyc_info = preprocess.XycPackage()
start = time.time()
with tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True))) as sess:
if drop_rate < 0:
keep_prob_info = [0.8, 0.5, 0.5]
else:
keep_prob_info = [
1.0 - drop_rate / 2, 1.0 - drop_rate, 1.0 - drop_rate
]
mlp = imm.TransferNN([784, 800, 800, 10], (optimizer, learning_rate),
keep_prob_info=keep_prob_info)
sess.run(tf.global_variables_initializer())
L_copy = []
FM = []
for i in range(no_of_task):
print("")
print("================= Train task #%d (%s) ================" %
(i + 1, optimizer))
if i > 0:
model_utils.CopyLayers(
sess, mlp.Layers,
mlp.Layers_dropbase) # Dropout from weight of pre-task
model_utils.ZeroLayers(sess, mlp.Layers)
learning_rate = FLAGS.learning_rate
epoch = int(FLAGS.epoch)
batch_size = FLAGS.batch_size
no_of_task = 3
no_of_node = [784,800,800,10]
keep_prob_info = [0.8, 0.5, 0.5]
# data preprocessing
x, y, x_, y_, xyc_info = preprocess.XycPackage()
start = time.time()
with tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))) as sess:
mlp = imm.TransferNN(no_of_node, (optimizer, learning_rate), keep_prob_info=keep_prob_info)
sess.run(tf.global_variables_initializer())
L_copy = []
FM = []
for i in range(no_of_task):
print("")
print("================= Train task #%d (%s) ================" % (i+1, optimizer))
mlp.Train(sess, x[i], y[i], x_[i], y_[i], epoch, mb=batch_size)
mlp.Test(sess, [[x[i],y[i]," train"], [x_[i],y_[i]," test"]])
if mean_imm or mode_imm:
L_copy.append(model_utils.CopyLayerValues(sess, mlp.Layers))
if mode_imm:
