def evaluate():
"""Eval CIFAR-10 for a number of steps."""
with tf.Graph().as_default() as g:
# Get images and labels for CIFAR-10.
eval_data = FLAGS.eval_data == 'test'
data, labels = architecture.inputs(True)
# Make the placeholders
d_pl, l_pl = architecture.placeholder_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = architecture.inference(d_pl)
# Calculate predictions.
top_k_op = tf.nn.in_top_k(logits, l_pl, 1)
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(
architecture.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
summary_writer = tf.train.SummaryWriter(FLAGS.eval_dir, g)
while True:
eval_once(saver, summary_writer, top_k_op, summary_op, data, labels, d_pl, l_pl)
if FLAGS.run_once:
break
time.sleep(FLAGS.eval_interval_secs)
def eval():
with tf.Graph().as_default() as graph:
images = input_.inputs("test", batch_size)
summary_op = tf.merge_all_summaries()
summary_writer = tf.train.SummaryWriter(eval_dir, graph)
logits = architecture.inference(images, "test")
variables_to_restore = tf.all_variables()
saver = tf.train.Saver(variables_to_restore)
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
coord = tf.train.Coordinator()
try:
tf.train.start_queue_runners(sess=sess)
threads = []
s = 0
while not coord.should_stop():
for q in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):
threads.extend(q.create_threads(sess, coord=coord, daemon=True, start=True))
imgs, gen_imgs = sess.run([images, logits])
for im, gen in zip(imgs, gen_imgs):
im = np.uint8(im)
im = cv2.resize(im, (200, 200))
gen = np.uint8(gen)
gen = cv2.resize(gen, (200, 200))
result = np.hstack((im, gen))
#cv2.imshow('result', result)
#cv2.waitKey(0)
#cv2.destroyAllWindows()
print "Step: " + str(s)
cv2.imwrite('../evaluations/images/image-'+str(s)+'.png', result)
s += 1
if s == int(sys.argv[1]):
print "Done"
exit()
except Exception as e:
print "Error"
raise(e)
coord.request_stop(e)
exit()
coord.request_stop()
coord.join(threads, stop_grace_period_secs=10)
def eval(checkpoint_dir, image):
with tf.Graph().as_default() as graph:
input_image = tf.placeholder(tf.float32, shape=(1, 144, 160, 3))
logit = architecture.inference(input_image, 'test')
variables = tf.all_variables()
init = tf.initialize_all_variables()
sess = tf.Session()
saver = tf.train.Saver(variables)
tf.train.start_queue_runners(sess=sess)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
print 'Restoring model...'
try:
saver.restore(sess, ckpt.model_checkpoint_path)
except:
print 'Could not restore model'
raise
exit()
graph_def = sess.graph.as_graph_def(add_shapes=True)
img = cv2.imread(image)
if len(img.shape) != 3:
img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
img = np.expand_dims(img, axis=2)
if img.shape[0] != 160:
img = cv2.resize(img, (160, 144))
img = img.astype('float')
fake = np.zeros((1, 144, 160, 3))
fake[0, :, :, :] = img
gen_img = sess.run([logit], feed_dict={input_image: fake})[0]
#gen_img = gen_img*255
image_name = image.split('.png')[0] + '_output.png'
image_name = image_name.split('/')[-1]
try:
print 'Writing image ', image_name
cv2.imwrite(image_name, gen_img[0, :, :, :])
except:
raise
def test(checkpoint_dir, record_file, image_path):
with tf.Graph().as_default():
input_image = tf.placeholder(tf.float32, shape=(10, 144, 160, 3))
logits = architecture.inference(1, input_image, "train")
# summary for tensorboard graph
summary_op = tf.merge_all_summaries()
variables = tf.all_variables()
init = tf.initialize_all_variables()
sess = tf.Session()
# saver for the model
saver = tf.train.Saver(tf.all_variables())
tf.train.start_queue_runners(sess=sess)
# restore previous model if one
print checkpoint_dir
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
print "Restoring previous model..." + ckpt.model_checkpoint_path
try:
saver.restore(sess, ckpt.model_checkpoint_path)
print "Model restored"
except:
print "Could not restore model"
pass
# Summary op
graph_def = sess.graph.as_graph_def(add_shapes=True)
# run on image
img = cv2.imread(image_path)
print(img)
img = cv2.resize(img, (160, 144)) / 255.0
#img = np.transpose(img, (1,0,2))
fake = np.zeros((10, 144, 160, 3))
fake[0, :, :, :] = img
high_res = sess.run([logits], feed_dict={input_image: fake})[0]
high_res = np.uint8(np.maximum(high_res, 0) * 255)
cv2.imwrite('hd_mario.jpg', high_res[0, :, :, :])
def tower_loss(scope):
"""Calculate the total loss on a single tower.
Args:
scope: unique prefix string identifying the tower, e.g. 'tower_0'
Returns:
Tensor of shape [] containing the total loss for a batch of data
"""
# Get images and labels.
images, labels = architecture.inputs(phase='train')
# Build inference Graph.
logits = architecture.inference(images, train=True)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
_ = architecture.loss(logits, labels)
# Assemble all of the losses for the current tower only.
losses = tf.get_collection('losses', scope)
# Calculate the total loss for the current tower.
total_loss = tf.add_n(losses, name='total_loss')
# Compute the moving average of all individual losses and the total loss.
loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg')
loss_averages_op = loss_averages.apply(losses + [total_loss])
# Attach a scalar summary to all individual losses and the total loss; do the
# same for the averaged version of the losses.
for l in losses + [total_loss]:
# Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training
# session. This helps the clarity of presentation on tensorboard.
loss_name = re.sub('%s_[0-9]*/' % architecture.TOWER_NAME, '',
l.op.name)
# Name each loss as '(raw)' and name the moving average version of the loss
# as the original loss name.
tf.scalar_summary(loss_name + ' (raw)', l)
tf.scalar_summary(loss_name, loss_averages.average(l))
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
return total_loss
def evaluate():
"""Eval for a number of steps."""
with tf.Graph().as_default() as g:
# Get images and labels.
images, labels = architecture.inputs(phase=FLAGS.phase)
# Build a Graph that computes the logits predictions from the
# inference model.
logits = architecture.inference(images, train=False)
# adapt logits
logits = tf.reshape(logits, (-1, NUM_CLASSES))
epsilon = tf.constant(value=1e-4)
logits = logits + epsilon
# predict
predictions = tf.argmax(logits, dimension=1)
labels = tf.cast(tf.reshape(labels, shape=predictions.get_shape()), dtype=tf.int64)
# compute accuracy
correct_predictions = tf.equal(predictions, labels)
accuracy = tf.reduce_mean(tf.cast(correct_predictions, dtype=tf.float32))
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(
architecture.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
summary_writer = tf.train.SummaryWriter(FLAGS.eval_dir, g)
tf.initialize_all_variables()
while True:
eval_once(saver, summary_writer, accuracy, summary_op)
if FLAGS.run_once:
break
time.sleep(FLAGS.eval_interval_secs)
def eval(checkpoint_dir, data_dir):
with tf.Graph().as_default() as graph:
pattern = '*.png'
image_list = []
for d, s, fList in os.walk(data_dir):
for filename in fList:
if fnmatch.fnmatch(filename, pattern):
image_list.append(os.path.join(d,filename))
num_images = len(image_list)
input_image = tf.placeholder(tf.float32, shape=(1,144,160,3))
logit = architecture.inference(input_image, 'test')
# specifies to use the CPU for eval, meaning I can eval while training
config = tf.ConfigProto(device_count={'GPU':0})
variables = tf.all_variables()
init = tf.initialize_all_variables()
sess = tf.Session(config=config)
saver = tf.train.Saver(variables)
tf.train.start_queue_runners(sess=sess)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
print 'Restoring model...'
try:
saver.restore(sess, ckpt.model_checkpoint_path)
except:
print 'Could not restore model'
raise
exit()
graph_def = sess.graph.as_graph_def(add_shapes=True)
for image in image_list:
img = cv2.imread(image)
if len(img.shape) != 3:
img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
img = np.expand_dims(img, axis=2)
if img.shape[0] != 160:
img = cv2.resize(img, (160,144))
img = img.astype('float')
#img = img/255.0
fake = np.zeros((1,144,160,3))
fake[0,:,:,:] = img
gen_img = sess.run([logit], feed_dict={input_image:fake})[0]
#gen_img = gen_img*255
image_name = image.split('.png')[0]+'_output.png'
image_name = image_name.split('/')[-1]
try:
print 'Writing image', image_name
cv2.imwrite('../images/output/'+image_name, gen_img[0,:,:,:])
except:
raise
def train(checkpoint_dir, image_list, batch_size, normalize):
with tf.Graph().as_default():
global_step = tf.Variable(0, name='global_step', trainable=False)
original_images_placeholder = tf.placeholder(tf.float32,
shape=(batch_size, 144,
160, 3))
gray_images_placeholder = tf.placeholder(tf.float32,
shape=(batch_size, 144, 160,
3))
# image summary for tensorboard
tf.image_summary('original_images',
original_images_placeholder,
max_images=100)
tf.image_summary('gray_images',
gray_images_placeholder,
max_images=100)
logits = architecture.inference(gray_images_placeholder, "train")
loss = architecture.loss(original_images_placeholder, logits)
tf.scalar_summary('loss', loss)
train_op = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(
loss, global_step=global_step)
# summary for tensorboard graph
summary_op = tf.merge_all_summaries()
variables = tf.all_variables()
init = tf.initialize_all_variables()
sess = tf.Session()
try:
os.mkdir(checkpoint_dir)
except:
pass
sess.run(init)
print "\nRunning session\n"
# saver for the model
saver = tf.train.Saver(tf.all_variables())
tf.train.start_queue_runners(sess=sess)
# restore previous model if one
ckpt = tf.train.get_checkpoint_state(checkpoint_dir + "training")
if ckpt and ckpt.model_checkpoint_path:
print "Restoring previous model..."
try:
saver.restore(sess, ckpt.model_checkpoint_path)
print "Model restored"
except:
print "Could not restore model"
pass
# Summary op
graph_def = sess.graph.as_graph_def(add_shapes=True)
summary_writer = tf.train.SummaryWriter(checkpoint_dir + "training",
graph_def=graph_def)
# Constants
step = int(sess.run(global_step))
#epoch_num = step/(train_size/batch_size)
while True:
step += 1
feed_dict = get_feed_dict(batch_size, original_images_placeholder,
gray_images_placeholder, image_list,
normalize)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
print " Step: " + str(
sess.run(global_step)) + " Loss: " + str(loss_value)
# save tensorboard stuff
#if step%200 == 0:
# summary_str = sess.run(summary_op)
# summary_writer.add_summary(summary_str, step)
if step % 300 == 0:
print "Saving model"
print
saver.save(sess,
checkpoint_dir + "training/checkpoint",
global_step=global_step)
print
def train(checkpoint_dir, record_file, batch_size, color):
with tf.Graph().as_default():
batch_size = int(batch_size)
global_step = tf.Variable(0, name='global_step', trainable=False)
train_size = 148623
gameboy_images, hd_images = input_.inputs(record_file, batch_size,
color, "train")
# image summary for tensorboard
tf.image_summary('gameboy_images', gameboy_images, max_images=100)
tf.image_summary('hd_images', hd_images, max_images=100)
logits = architecture.inference(batch_size, gameboy_images, "train")
# loss is the l2 norm of my input vector (the image) and the output vector (generated image)
loss = architecture.loss(hd_images, logits)
tf.scalar_summary('loss', loss)
train_op = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(
loss, global_step=global_step)
# summary for tensorboard graph
summary_op = tf.merge_all_summaries()
variables = tf.all_variables()
init = tf.initialize_all_variables()
sess = tf.Session()
try:
os.mkdir(checkpoint_dir)
except:
pass
sess.run(init)
print "\nRunning session\n"
# saver for the model
saver = tf.train.Saver(tf.all_variables())
tf.train.start_queue_runners(sess=sess)
# restore previous model if one
ckpt = tf.train.get_checkpoint_state(checkpoint_dir + "training/")
if ckpt and ckpt.model_checkpoint_path:
print "Restoring previous model..."
try:
saver.restore(sess, ckpt.model_checkpoint_path)
print "Model restored"
except:
print "Could not restore model"
pass
# Summary op
graph_def = sess.graph.as_graph_def(add_shapes=True)
summary_writer = tf.train.SummaryWriter(checkpoint_dir + "training",
graph_def=graph_def)
# Constants
step = int(sess.run(global_step))
epoch_num = step / (train_size / batch_size)
while True:
_, loss_value = sess.run([train_op, loss])
step += 1
print "Epoch: " + str(epoch_num) + " Step: " + str(
sess.run(global_step)) + " Loss: " + str(loss_value)
# save tensorboard stuff
if step % 200 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# save checkpoint
if step % 1000 == 0:
print "Finished epoech " + str(epoch_num) + " ....saving model"
print
#saver.save(sess, checkpoint_dir+"epoch-"+str(epoch_num), global_step=global_step)
saver.save(sess,
checkpoint_dir + "training/checkpoint",
global_step=global_step)
print
#____________________Setup____________________#
print('Beginning program...')
if tf.gfile.Exists(train_dir):
tf.gfile.DeleteRecursively(train_dir)
tf.gfile.MakeDirs(train_dir)
# data_dir, data_type, starting file, ending file, batch_size
data = pvpinput.Input(pvp_dir, 0, 100, batch_size)
data_pl = tf.placeholder(tf.float32, shape=(None, ny, nx, nf))
label_pl = tf.placeholder(tf.int32, shape=(None))
print('Building graph...')
logits = architecture.inference(data_pl, 10, batch_size)
loss = architecture.loss(logits, label_pl)
train_op = architecture.train(loss, global_step)
acc_op = architecture.accuracy(logits, label_pl)
#tf.scalar_summary('Accuracy', acc_op)
#tf.scalar_summary('Loss', loss)
labels_max = tf.reduce_max(label_pl)
print('Starting session...')
saver = tf.train.Saver(tf.all_variables())
summary_op = tf.merge_all_summaries()
init = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init)
tf.train.start_queue_runners(sess=sess)