def __init__(self, dataset, batch_size):
# Get cifar data
print(
'-----------------------------data_layer--------------------------------'
)
self.dataset = dataset
self.batch_size = batch_size
if dataset == 'cifar10':
print('--------loading cifar10 dataset------------------')
self.train_images, self.train_labels, self.valid_images, self.valid_labels = cifar_tools.get_dataset_cifar10(
'training')
print(self.train_images.shape, self.train_labels.shape)
print(self.valid_images.shape, self.valid_labels.shape)
self.test_images, self.test_labels = cifar_tools.get_dataset_cifar10(
'test')
print(self.test_images.shape, self.test_labels.shape)
self.sup_by_label = backend.sample_by_label(
self.train_images, self.train_labels, -1, 10, None)
elif dataset == 'cifar100':
print('--------loading cifar100 dataset------------------')
self.train_images, self.train_labels, self.valid_images, self.valid_labels = cifar_tools.get_dataset_cifar100(
'training')
print(self.train_images.shape, self.train_labels.shape)
print(self.valid_images.shape, self.valid_labels.shape)
self.test_images, self.test_labels = cifar_tools.get_dataset_cifar100(
'test')
print(self.test_images.shape, self.test_labels.shape)
self.sup_by_label = None
elif dataset == 'mnist':
print('--------loading mnist dataset------------------')
self.train_images, self.train_labels = mnist_tools.get_data(
'train')
print(self.train_images.shape, self.train_labels.shape)
self.test_images, self.test_labels = mnist_tools.get_data('test')
print(self.test_images.shape, self.test_labels.shape)
self.sup_by_label = None
elif dataset == 'stl10':
print('--------loading stl10 dataset------------------')
self.trainval_images, self.trainval_labels = stl10_tools.get_data(
'training')
self.train_images, self.train_labels = self.trainval_images[:
4500], self.trainval_labels[:
4500]
self.valid_images, self.valid_labels = self.trainval_images[
4500:], self.trainval_labels[4500:]
print(self.train_images.shape, self.train_labels.shape)
print(self.valid_images.shape, self.valid_labels.shape)
self.test_images, self.test_labels = stl10_tools.get_data('test')
print(self.test_images.shape, self.test_labels.shape)
self.sup_by_label = None
self.it = [0, 0]
self.perm = [[], []]
print(
'-----------------------------data_layer--------------------------------'
)
def init_data():
train_images, train_labels = mnist_tools.get_data('train')
test_images, test_labels = mnist_tools.get_data('test')
# Sample labeled training subset.
seed = FLAGS.sup_seed if FLAGS.sup_seed != -1 else np.random.randint(
0, 1000)
print('Seed:', seed)
sup_by_label = semisup.sample_by_label(train_images, train_labels,
FLAGS.sup_per_class, NUM_LABELS,
seed)
return sup_by_label, train_images, train_labels, test_images, test_labels, seed
def main(_):
train_images, train_labels = mnist_tools.get_data('train')
test_images, test_labels = mnist_tools.get_data('test')
# Sample labeled training subset.
seed = FLAGS.sup_seed if FLAGS.sup_seed != -1 else np.random.randint(
0, 1000)
print('Seed:', seed)
sup_by_label = semisup.sample_by_label(train_images, train_labels,
FLAGS.sup_per_class, NUM_LABELS,
seed)
# add twos, fours, and sixes
if FLAGS.testadd:
num_to_add = 10
for i in range(10):
items = np.where(train_labels == i)[0]
inds = np.random.choice(len(items), num_to_add, replace=False)
sup_by_label[i] = np.vstack(
[sup_by_label[i], train_images[items[inds]]])
add_random_samples = 0
if add_random_samples > 0:
rng = np.random.RandomState()
indices = rng.choice(len(train_images), add_random_samples, False)
for i in indices:
l = train_labels[i]
sup_by_label[l] = np.vstack([sup_by_label[l], [train_images[i]]])
print(l)
graph = tf.Graph()
with graph.as_default():
model = semisup.SemisupModel(semisup.architectures.mnist_model,
NUM_LABELS,
IMAGE_SHAPE,
dropout_keep_prob=0.8)
# Set up inputs.
if FLAGS.random_batches:
sup_lbls = np.asarray(
np.hstack([
np.ones(len(i)) * ind for ind, i in enumerate(sup_by_label)
]), np.int)
sup_images = np.vstack(sup_by_label)
t_sup_images, t_sup_labels = semisup.create_input(
sup_images, sup_lbls, FLAGS.sup_per_batch * NUM_LABELS)
else:
t_sup_images, t_sup_labels = semisup.create_per_class_inputs(
sup_by_label, FLAGS.sup_per_batch)
# Compute embeddings and logits.
t_sup_emb = model.image_to_embedding(t_sup_images)
t_sup_logit = model.embedding_to_logit(t_sup_emb)
# Add losses.
if FLAGS.semisup:
if FLAGS.equal_cls_unsup:
allimgs_bylabel = semisup.sample_by_label(
train_images, train_labels, 5000, NUM_LABELS, seed)
t_unsup_images, _ = semisup.create_per_class_inputs(
allimgs_bylabel, FLAGS.sup_per_batch)
else:
t_unsup_images, _ = semisup.create_input(
train_images, train_labels, FLAGS.unsup_batch_size)
t_unsup_emb = model.image_to_embedding(t_unsup_images)
model.add_semisup_loss(t_sup_emb,
t_unsup_emb,
t_sup_labels,
walker_weight=FLAGS.walker_weight,
visit_weight=FLAGS.visit_weight,
proximity_weight=FLAGS.proximity_weight)
model.add_logit_loss(t_sup_logit, t_sup_labels)
t_learning_rate = tf.train.exponential_decay(FLAGS.learning_rate,
model.step,
FLAGS.decay_steps,
FLAGS.decay_factor,
staircase=True)
train_op = model.create_train_op(t_learning_rate)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(FLAGS.logdir, graph)
saver = tf.train.Saver()
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for step in range(FLAGS.max_steps):
_, summaries = sess.run([train_op, summary_op])
if (step + 1) % FLAGS.eval_interval == 0 or step == 99:
print('Step: %d' % step)
test_pred = model.classify(test_images, sess).argmax(-1)
conf_mtx = semisup.confusion_matrix(test_labels, test_pred,
NUM_LABELS)
test_err = (test_labels != test_pred).mean() * 100
print(conf_mtx)
print('Test error: %.2f %%' % test_err)
print()
test_summary = tf.Summary(value=[
tf.Summary.Value(tag='Test Err', simple_value=test_err)
])
summary_writer.add_summary(summaries, step)
summary_writer.add_summary(test_summary, step)
saver.save(sess, FLAGS.logdir, model.step)
coord.request_stop()
coord.join(threads)
def main(_):
FLAGS.emb_size = 128
optimizer = 'adam'
train_images, train_labels = mnist_tools.get_data('train')
test_images, test_labels = mnist_tools.get_data('test')
# Sample labeled training subset.
seed = FLAGS.sup_seed if FLAGS.sup_seed != -1 else np.random.randint(0, 1000)
print('Seed:', seed)
sup_by_label = semisup.sample_by_label(train_images, train_labels,
FLAGS.sup_per_class, NUM_LABELS, seed)
graph = tf.Graph()
with graph.as_default():
model_func = semisup.architectures.mnist_model
if FLAGS.dropout_keep_prob < 1:
model_func = semisup.architectures.mnist_model_dropout
model = semisup.SemisupModel(model_func, NUM_LABELS,
IMAGE_SHAPE, optimizer=optimizer, emb_size=FLAGS.emb_size,
dropout_keep_prob=FLAGS.dropout_keep_prob)
# Set up inputs.
t_sup_images, t_sup_labels = semisup.create_per_class_inputs(
sup_by_label, FLAGS.sup_per_batch)
# Compute embeddings and logits.
t_sup_emb = model.image_to_embedding(t_sup_images)
t_sup_logit = model.embedding_to_logit(t_sup_emb)
t_unsup_images = tf.placeholder("float", shape=[None] + IMAGE_SHAPE)
proximity_weight = tf.placeholder("float", shape=[])
visit_weight = tf.placeholder("float", shape=[])
walker_weight = tf.placeholder("float", shape=[])
t_logit_weight = tf.placeholder("float", shape=[])
t_l1_weight = tf.placeholder("float", shape=[])
t_learning_rate = tf.placeholder("float", shape=[])
t_unsup_emb = model.image_to_embedding(t_unsup_images)
model.add_semisup_loss(
t_sup_emb, t_unsup_emb, t_sup_labels,
walker_weight=walker_weight, visit_weight=visit_weight, proximity_weight=proximity_weight)
model.add_logit_loss(t_sup_logit, t_sup_labels, weight=t_logit_weight)
model.add_emb_regularization(t_sup_emb, weight=t_l1_weight)
model.add_emb_regularization(t_unsup_emb, weight=t_l1_weight)
train_op = model.create_train_op(t_learning_rate)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(FLAGS.logdir, graph)
saver = tf.train.Saver()
sess = tf.InteractiveSession(graph=graph)
unsup_images_iterator = semisup.create_input(train_images, train_labels,
FLAGS.unsup_batch_size)
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
use_new_visit_loss = False
for step in range(FLAGS.max_steps):
unsup_images, _ = sess.run(unsup_images_iterator)
if use_new_visit_loss:
_, summaries, train_loss = sess.run([train_op, summary_op, model.train_loss], {
t_unsup_images: unsup_images,
walker_weight: FLAGS.walker_weight,
proximity_weight: 0.3 + apply_envelope("lin", step, 0.7, FLAGS.warmup_steps, 0)
- apply_envelope("lin", step, FLAGS.visit_weight, 2000, FLAGS.warmup_steps),
t_logit_weight: FLAGS.logit_weight,
t_l1_weight: FLAGS.l1_weight,
visit_weight: apply_envelope("lin", step, FLAGS.visit_weight, 2000, FLAGS.warmup_steps),
t_learning_rate: 5e-5 + apply_envelope("log", step, FLAGS.learning_rate, FLAGS.warmup_steps, 0)
})
else:
_, summaries, train_loss = sess.run([train_op, summary_op, model.train_loss], {
t_unsup_images: unsup_images,
walker_weight: FLAGS.walker_weight,
visit_weight: 0.3 + apply_envelope("lin", step, 0.7, FLAGS.warmup_steps, 0),
proximity_weight: 0,
t_logit_weight: FLAGS.logit_weight,
t_l1_weight: FLAGS.l1_weight,
t_learning_rate: 5e-5 + apply_envelope("log", step, FLAGS.learning_rate, FLAGS.warmup_steps, 0)
})
if (step + 1) % FLAGS.eval_interval == 0 or step == 99:
print('Step: %d' % step)
test_pred = model.classify(test_images).argmax(-1)
conf_mtx = semisup.confusion_matrix(test_labels, test_pred, NUM_LABELS)
test_err = (test_labels != test_pred).mean() * 100
print(conf_mtx)
print('Test error: %.2f %%' % test_err)
print('Train loss: %.2f ' % train_loss)
print()
test_summary = tf.Summary(
value=[tf.Summary.Value(
tag='Test Err', simple_value=test_err)])
summary_writer.add_summary(summaries, step)
summary_writer.add_summary(test_summary, step)
def main(_):
train_images, train_labels = mnist_tools.get_data('train')
test_images, test_labels = mnist_tools.get_data('test')
# Sample labeled training subset.
seed = FLAGS.sup_seed if FLAGS.sup_seed != -1 else None
sup_by_label = semisup.sample_by_label(train_images, train_labels,
FLAGS.sup_per_class, NUM_LABELS, seed)
graph = tf.Graph()
with graph.as_default():
model = semisup.SemisupModel(semisup.architectures.mnist_model, NUM_LABELS,
IMAGE_SHAPE)
# Set up inputs.
t_unsup_images, _ = semisup.create_input(train_images, train_labels,
FLAGS.unsup_batch_size)
t_sup_images, t_sup_labels = semisup.create_per_class_inputs(
sup_by_label, FLAGS.sup_per_batch)
# Compute embeddings and logits.
t_sup_emb = model.image_to_embedding(t_sup_images)
t_unsup_emb = model.image_to_embedding(t_unsup_images)
t_sup_logit = model.embedding_to_logit(t_sup_emb)
# Add losses.
model.add_semisup_loss(
t_sup_emb, t_unsup_emb, t_sup_labels, visit_weight=FLAGS.visit_weight)
model.add_logit_loss(t_sup_logit, t_sup_labels)
t_learning_rate = tf.train.exponential_decay(
FLAGS.learning_rate,
model.step,
FLAGS.decay_steps,
FLAGS.decay_factor,
staircase=True)
train_op = model.create_train_op(t_learning_rate)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(FLAGS.logdir, graph)
saver = tf.train.Saver()
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for step in xrange(FLAGS.max_steps):
_, summaries = sess.run([train_op, summary_op])
if (step + 1) % FLAGS.eval_interval == 0 or step == 99:
print('Step: %d' % step)
test_pred = model.classify(test_images).argmax(-1)
conf_mtx = semisup.confusion_matrix(test_labels, test_pred, NUM_LABELS)
test_err = (test_labels != test_pred).mean() * 100
print(conf_mtx)
print('Test error: %.2f %%' % test_err)
print()
test_summary = tf.Summary(
value=[tf.Summary.Value(
tag='Test Err', simple_value=test_err)])
summary_writer.add_summary(summaries, step)
summary_writer.add_summary(test_summary, step)
saver.save(sess, FLAGS.logdir, model.step)
coord.request_stop()
coord.join(threads)
def main(_):
if FLAGS.logdir is not None:
FLAGS.logdir = FLAGS.logdir + '/t_mnist_eval'
try:
shutil.rmtree(FLAGS.logdir)
except OSError as e:
print ("Error: %s - %s." % (e.filename, e.strerror))
train_images, train_labels = mnist_tools.get_data('train')
test_images, test_labels = mnist_tools.get_data('test')
# Sample labeled training subset.
if FLAGS.sup_seed >= 0:
seed = FLAGS.sup_seed
elif FLAGS.sup_seed == -2:
seed = FLAGS.sup_per_class
else:
seed = np.random.randint(0, 1000)
print('Seed:', seed)
sup_by_label = semisup.sample_by_label(train_images, train_labels,
FLAGS.sup_per_class, NUM_LABELS, seed)
graph = tf.Graph()
with graph.as_default():
model = semisup.SemisupModel(semisup.architectures.mnist_model, NUM_LABELS,
IMAGE_SHAPE, dropout_keep_prob=FLAGS.dropout_keep_prob)
# Set up inputs.
t_sup_images, t_sup_labels = semisup.create_per_class_inputs(
sup_by_label, FLAGS.sup_per_batch)
# Compute embeddings and logits.
t_sup_emb = model.image_to_embedding(t_sup_images)
t_sup_logit = model.embedding_to_logit(t_sup_emb)
# Add losses.
if FLAGS.semisup:
t_unsup_images, _ = semisup.create_input(train_images, train_labels,
FLAGS.unsup_batch_size)
t_unsup_emb = model.image_to_embedding(t_unsup_images)
model.add_semisup_loss(
t_sup_emb, t_unsup_emb, t_sup_labels,
walker_weight=FLAGS.walker_weight, visit_weight=FLAGS.visit_weight)
#model.add_emb_regularization(t_unsup_emb, weight=FLAGS.l1_weight)
model.add_logit_loss(t_sup_logit, t_sup_labels, weight=FLAGS.logit_weight)
#model.add_emb_regularization(t_sup_emb, weight=FLAGS.l1_weight)
t_learning_rate = tf.placeholder("float", shape=[])
train_op = model.create_train_op(t_learning_rate)
summary_op = tf.summary.merge_all()
if FLAGS.logdir is not None:
summary_writer = tf.summary.FileWriter(FLAGS.logdir, graph)
saver = tf.train.Saver()
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
learning_rate_ = FLAGS.learning_rate
for step in range(FLAGS.max_steps):
lr = learning_rate_
if step < FLAGS.warmup_steps:
lr = 1e-6 + semisup.apply_envelope("log", step, FLAGS.learning_rate, FLAGS.warmup_steps, 0)
_, summaries = sess.run([train_op, summary_op], {
t_learning_rate: lr
})
sys.stderr.write("\rstep: %d" % step)
sys.stdout.flush()
if (step + 1) % FLAGS.eval_interval == 0 or step == 99:
print('\nStep: %d' % step)
test_pred = model.classify(test_images, sess).argmax(-1)
conf_mtx = semisup.confusion_matrix(test_labels, test_pred, NUM_LABELS)
test_err = (test_labels != test_pred).mean() * 100
print(conf_mtx)
print('Test error: %.2f %%' % test_err)
if FLAGS.logdir is not None:
sum_values = {
'Test error': test_err
}
summary_writer.add_summary(summaries, step)
for key, value in sum_values.items():
summary = tf.Summary(
value=[tf.Summary.Value(tag=key, simple_value=value)])
summary_writer.add_summary(summary, step)
if step % FLAGS.decay_steps == 0 and step > 0:
learning_rate_ = learning_rate_ * FLAGS.decay_factor
coord.request_stop()
coord.join(threads)
print('FINAL RESULTS:')
print('Test error: %.2f %%' % (test_err))
print('final_score', 1 - test_err/100)
print('@@test_error:%.4f' % (test_err/100))
print('@@train_loss:%.4f' % 0)
print('@@reg_loss:%.4f' % 0)
print('@@estimated_error:%.4f' % 0)
print('@@centroid_norm:%.4f' % 0)
print('@@emb_norm:%.4f' % 0)
print('@@k_score:%.4f' % 0)
print('@@svm_score:%.4f' % 0)
def main(_):
train_images, train_labels = mnist_tools.get_data('train')
test_images, test_labels = mnist_tools.get_data('test')
# Sample labeled training subset.
seed = FLAGS.sup_seed if FLAGS.sup_seed != -1 else None
sup_by_label = semisup.sample_by_label(train_images, train_labels,
FLAGS.sup_per_class, NUM_LABELS, seed)
import numpy as np
if 0:
indices = [374, 2507, 9755, 12953, 16507, 16873, 23474,
23909, 30280, 35070, 49603, 50106, 51171, 51726, 51805, 55205, 57251, 57296, 57779, 59154] + \
[16644, 45576, 52886, 42140, 29201, 7767, 24, 134, 8464, 15022, 15715, 15602, 11030, 3898, 10195, 1454,
3290, 5293, 5806, 274]
indices = [374, 2507, 9755, 12953, 16507, 16873, 23474,
23909, 30280, 35070, 49603, 50106, 51171, 51726, 51805, 55205, 57251, 57296, 57779, 59154] + \
[9924, 34058, 53476, 15715, 6428, 33598, 33464, 41753, 21250, 26389, 12950,
12464, 3795, 6761, 5638, 3952, 8300, 5632, 1475, 1875]
sup_by_label = [ [] for i in range(NUM_LABELS)]
for ind in indices:
i = train_labels[ind]
sup_by_label[i] = sup_by_label[i] + [train_images[ind]]
for i in range(10):
sup_by_label[i] = np.asarray(sup_by_label[i])
sup_by_label = np.asarray(sup_by_label)
add_random_samples = 20
if add_random_samples > 0:
rng = np.random.RandomState()
indices = rng.choice(len(train_images), add_random_samples, False)
for i in indices:
l = train_labels[i]
sup_by_label[l] = np.vstack([sup_by_label[l], [train_images[i]]])
graph = tf.Graph()
with graph.as_default():
model = semisup.SemisupModel(semisup.architectures.mnist_model, NUM_LABELS,
IMAGE_SHAPE)
# Set up inputs.
t_unsup_images, _ = semisup.create_input(train_images, train_labels,
FLAGS.unsup_batch_size)
t_sup_images, t_sup_labels = semisup.create_per_class_inputs(
sup_by_label, FLAGS.sup_per_batch)
# Compute embeddings and logits.
t_sup_emb = model.image_to_embedding(t_sup_images)
t_unsup_emb = model.image_to_embedding(t_unsup_images)
t_sup_logit = model.embedding_to_logit(t_sup_emb)
# Add losses.
model.add_semisup_loss(
t_sup_emb, t_unsup_emb, t_sup_labels, visit_weight=FLAGS.visit_weight)
model.add_logit_loss(t_sup_logit, t_sup_labels)
t_learning_rate = tf.train.exponential_decay(
FLAGS.learning_rate,
model.step,
FLAGS.decay_steps,
FLAGS.decay_factor,
staircase=True)
train_op = model.create_train_op(t_learning_rate)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(FLAGS.logdir, graph)
saver = tf.train.Saver()
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for step in range(FLAGS.max_steps):
_, summaries = sess.run([train_op, summary_op])
if (step + 1) % FLAGS.eval_interval == 0 or step == 99:
print('Step: %d' % step)
test_pred = model.classify(test_images).argmax(-1)
conf_mtx = semisup.confusion_matrix(test_labels, test_pred, NUM_LABELS)
test_err = (test_labels != test_pred).mean() * 100
print(conf_mtx)
print('Test error: %.2f %%' % test_err)
print()
test_summary = tf.Summary(
value=[tf.Summary.Value(
tag='Test Err', simple_value=test_err)])
summary_writer.add_summary(summaries, step)
summary_writer.add_summary(test_summary, step)
saver.save(sess, FLAGS.logdir, model.step)
coord.request_stop()
coord.join(threads)