def evaluate(test_data, model_file):
with tf.Graph().as_default() as g:
with tf.device('cpu:0'):
x = tf.placeholder(tf.float32, [
test_data.num_examples, DTN.IMAGE_SIZE, DTN.IMAGE_SIZE,
DTN.NUM_CHANNELS
],
name='x-input')
y_ = tf.placeholder(tf.float32, [None, DTN.OUTPUT_NODE],
name='y-input')
_, y = DTN.inference(x, False, None, reuse=False, trainable=False)
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# accuracy_sum = tf.reduce_sum(tf.cast(correct_prediction, tf.float32))
# load saved model
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_file)
accuracy_score = sess.run(
accuracy,
feed_dict={
x:
np.reshape(test_data.images,
(test_data.num_examples, DTN.IMAGE_SIZE,
DTN.IMAGE_SIZE, DTN.NUM_CHANNELS)),
y_:
test_data.labels
})
print(model_file, "test accuracy = %f" % accuracy_score)
def pre_train(source_domain_data):
xs = tf.placeholder(
tf.float32,
[PRE_BATCH_SIZE, DTN.IMAGE_SIZE, DTN.IMAGE_SIZE, DTN.NUM_CHANNELS],
name='source-x-input')
ys_ = tf.placeholder(tf.float32, [None, DTN.OUTPUT_NODE],
name='source-y-input')
regularizer = tf.contrib.layers.l2_regularizer(RAGULARZTION_RATE)
_, ys = DTN.inference(xs, True, regularizer, reuse=False, trainable=True)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=ys, labels=tf.argmax(ys_, 1))
cross_entropy_mean = tf.reduce_mean(cross_entropy)
loss = cross_entropy_mean + tf.add_n(tf.get_collection('losses'))
train = tf.train.AdamOptimizer(LEARNING_RATE).minimize(loss)
saver = tf.train.Saver()
with tf.Session() as sess:
tf.global_variables_initializer().run()
for i in range(PRE_TRAINING_STEPS):
x, y = source_domain_data.next_batch(PRE_BATCH_SIZE)
reshaped_xs = np.reshape(x, (PRE_BATCH_SIZE, DTN.IMAGE_SIZE,
DTN.IMAGE_SIZE, DTN.NUM_CHANNELS))
loss_value, _ = sess.run([loss, train],
feed_dict={
xs: reshaped_xs,
ys_: y
})
print("After %d training steps, loss on training batch is %f" %
(i, loss_value))
saver.save(sess, os.path.join(MODEL_SAVE_PATH, PRE_MODEL_NAME))
xs = tf.placeholder(
tf.float32,
[BATCH_SIZE, DTN.IMAGE_SIZE, DTN.IMAGE_SIZE, DTN.NUM_CHANNELS],
name='source-x-input')
xt = tf.placeholder(
tf.float32,
[BATCH_SIZE, DTN.IMAGE_SIZE, DTN.IMAGE_SIZE, DTN.NUM_CHANNELS],
name='target-x-input')
ys_ = tf.placeholder(tf.float32, [None, DTN.OUTPUT_NODE],
name='source-y-input')
# trainable=False fix lower layer param
_, _, source_fc, ys = DTN.inference(
xs, False, new_regularizer, reuse=False,
trainable=False) # False: without dropout
_, _, target_fc, yt = DTN.inference(
xt, False, new_regularizer, reuse=True,
trainable=False) # pesudo logit: yt
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=ys, labels=tf.argmax(ys, 1))
# 加上target y中有标签的部分
cross_entropy_mean = tf.reduce_mean(cross_entropy)
mmd_mar, mmd_con = con_MMD(source_fc, ys_, target_fc, yt)
loss = cross_entropy_mean + theta1 * mmd_mar + theta2 * mmd_con + tf.add_n(
tf.get_collection('losses'))
DTN.IMAGE_SIZE,
DTN.NUM_CHANNELS],
name='source-x-input')
xt = tf.placeholder(tf.float32, [target_domain_data.num_examples,
DTN.IMAGE_SIZE,
DTN.IMAGE_SIZE,
DTN.NUM_CHANNELS],
name='target-x-input')
ys_ = tf.placeholder(tf.float32, [None, DTN.OUTPUT_NODE], name='source-y-input')
yt_ = tf.placeholder(tf.float32, [None, DTN.OUTPUT_NODE], name='source-y-input')
# trainable=False fix lower layer param
s_pool, s_fc1, s_fc2, ys = DTN.inference(xs, False, None, reuse=False, trainable=False) # False: without dropout
t_pool, t_fc1, t_fc2, yt = DTN.inference(xt, False, None, reuse=True, trainable=False) # pesudo logit: yt
# TODO ys_: ground truth label, yt: predicted label
pool_mmd_mar, pool_mmd_con = con_MMD(s_pool, ys_, t_pool, yt)
fc1_mmd_mar, fc1_mmd_con = con_MMD(s_fc1, ys_, t_fc1, yt)
fc2_mmd_mar, fc2_mmd_con = con_MMD(s_fc2, ys_, t_fc2, yt)
# TODO learning rate decay
variables = slim.get_variables_to_restore()
variables_to_restore = [v for v in variables if v.name.split('_')[0] != 'step']
saver = tf.train.Saver(variables_to_restore)
with tf.Session() as sess:
# print(variables_to_restore)
