def run_training(data):
# TODO: read in data
training_data = data[0]
labels = data[1]
# test_data =
with tf.Graph().as_default():
input_pl, labels_pl = network.placeholder_inputs(BATCH_SIZE)
logits = network.feedforward(training_data,
NUM_HIDDEN1,
NUM_HIDDEN2)
loss = network.loss(logits, labels_pl)
train_op = network.training(loss,ETA)
eval_correct = network.evaluation(logits, labels_pl)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
init = tf.initialize_all_variables()
sess.run(init)
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph)
for epoch in xrange(EPOCHS):
start_time = time.time()
training_size = len(training_data)
batches = [training_data[k:k+BATCHSIZE] for k in xrange(0, training_size,BATCH_SIZE)]
for batch in batches:
feed_dict = fill_dict(training_data,
labels,
input_pl,
labels_pl)
_, loss_value = sess.run([train_op, loss], feed_dict = feed_dict)
duration = time.time() - start_time
# Write summarry after each 10 epochs
if epoch % 10 == 0:
print 'Epoch %d: loss = %.2f (%.3f sec)'%(epoch, loss_value, duration)
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, epoch)
summary_writer.flush()
print 'Evaluate with the validation set...'
validate(sess,
eval_correct,
input_pl,
labels_pl,
test_data)
def main(_):
# Read the expert rollouts from disk.
observations, actions = load_data(ARGS.rollouts_file)
print("observations shape = " + str(observations.shape))
print("actions shape = " + str(actions.shape))
observation_length = observations.shape[1]
action_length = actions.shape[1]
assert (observations.shape[0] == actions.shape[0])
assert (observations.shape[0] % ARGS.batch_size == 0)
# Assemble the network.
opl = tf.placeholder(tf.float32,
shape=(None, observation_length),
name="observations")
apl = tf.placeholder(tf.float32,
shape=(None, action_length),
name="actions")
logits = network.inference(opl, observation_length, ARGS.hidden1,
ARGS.hidden2, action_length)
errors, loss = network.loss(logits, apl)
global_step, train_op = network.training(loss, ARGS.learning_rate)
# Initialize the network.
init = tf.global_variables_initializer()
saver = tf.train.Saver(max_to_keep=10)
sess = tf.Session()
if os.path.exists(ARGS.checkpoint_dir):
saver.restore(sess, tf.train.latest_checkpoint(ARGS.checkpoint_dir))
else:
sess.run(init)
# Train the network.
num_batches = observations.shape[0] / ARGS.batch_size
for step in range(ARGS.training_steps):
i = step % num_batches
if i == 0:
p = np.random.permutation(observations.shape[0])
observations = observations[p]
actions = actions[p]
start = int(i * ARGS.batch_size)
stop = int((i + 1) * ARGS.batch_size)
feed_dict = {opl: observations[start:stop], apl: actions[start:stop]}
_, loss_value, step_value = sess.run([train_op, loss, global_step],
feed_dict=feed_dict)
if step % 100 == 0:
basename = os.path.basename(ARGS.checkpoint_dir)
checkpoint_file = os.path.join(ARGS.checkpoint_dir, basename)
saver.save(sess, checkpoint_file, global_step=step_value)
loss_value = sess.run(loss,
feed_dict={
opl: observations,
apl: actions
})
msg = "step {}; loss = {}".format(step_value, loss_value)
print(msg)
def train():
"""Train network for a number of steps."""
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
# Get images and labels for network.
images, labels = network.distorted_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = network.inference(images)
# Calculate loss.
loss = network.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = network.train(loss, global_step)
# Create a saver.
saver = tf.train.Saver(tf.all_variables())
# Build the summary operation based on the TF collection of Summaries.
#summary_op = tf.merge_all_summaries()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Start the queue runners.
tf.train.start_queue_runners(sess=sess)
#summary_writer = tf.train.SummaryWriter(FLAGS.train_dir,
# graph_def=sess.graph_def)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_input_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
def train():
"""Train network for a number of steps."""
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
# Get images and labels for network.
images, labels = network.distorted_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = network.inference(images)
# Calculate loss.
loss = network.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = network.train(loss, global_step)
# Create a saver.
saver = tf.train.Saver(tf.all_variables())
# Build the summary operation based on the TF collection of Summaries.
#summary_op = tf.merge_all_summaries()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Start the queue runners.
tf.train.start_queue_runners(sess=sess)
#summary_writer = tf.train.SummaryWriter(FLAGS.train_dir,
# graph_def=sess.graph_def)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_input_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print (format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
def train():
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
images, labels = network.distorted_inputs()
logits = network.inference(images)
loss = network.loss(logits, labels)
train_op = network.train(loss, global_step)
saver = tf.train.Saver(tf.all_variables())
summary_op = tf.merge_all_summaries()
init = tf.initialize_all_variables()
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_input_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
if step % 100 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
if step % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
def train_all(epochs_count=EPOCHS_COUNT, batch_size=BATCH_SIZE):
model.train()
for i in range(epochs_count):
training_loss = 0
for j in range(batches_count):
network_output = model(get_batch_features_tensor(j).to(device))
optimizer.zero_grad()
loss_value = network.loss(network_output, get_batch_labels_tensor(j))
training_loss += loss_value.item()
loss_value.backward()
optimizer.step()
training_loss /= batches_count
print("Epoch number:", i + 1)
print("Training loss:", training_loss)
validate()
print("---------------------")
def train():
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
# 获得图片数据和对应的标签batch
float_image, label = tfrecord.train_data_read(tfrecord_path=FLAGS.train_data)
images, labels = tfrecord.create_batch(float_image,label,count_num=FLAGS.train_num)
logits = network.inference(images)
# 误差计算
loss = network.loss(logits, labels)
# 模型训练
train_op = network.train(loss, global_step)
# 存储模型
saver = tf.train.Saver(tf.global_variables())
# 存储所有操作
summary_op = tf.summary.merge_all()
# 初始化所有变量.
init = tf.initialize_all_variables()
# 开始计算流图
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# 队列开始
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.summary.FileWriter(FLAGS.train_dir,
graph_def=sess.graph_def)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print (format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
if step % 50 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# 保存模型检查点.
if step % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
parameter_table = [["Initial parameters", parameter_path],
["Ranking loss", ranking_loss], ["SPP", spp], ["Pooling", args.pooling],
['Experiment', experiment_name],
['Embedding dim', embedding_dim], ['Batch size', batch_size_trn],
['Initial LR', initial_lr], ['Momentum', momentum_coeff],
['LR Step size', step_size], ['LR Step factor', step_factor],
['Total Steps', total_steps]]
training_images = inputs(args.training_db, batch_size_trn, None, True, augment_training_data)
test_images = inputs(args.validation_db, batch_size_val, None, False)
net_data = np.load(parameter_path).item()
var_dict= nw.get_variable_dict(net_data)
with tf.variable_scope("ranker") as scope:
feature_vec = nw.build_alexconvnet(training_images, var_dict, embedding_dim, spp, args.pooling)
L, p = nw.loss(feature_vec, nw.build_loss_matrix(batch_size_trn), ranking_loss)
scope.reuse_variables()
val_feature_vec = nw.build_alexconvnet(test_images, var_dict, embedding_dim, spp, args.pooling)
L_val, p_val = nw.loss(val_feature_vec, nw.build_loss_matrix(batch_size_val), ranking_loss)
lr = tf.Variable(initial_lr)
opt = tf.train.AdamOptimizer()
grads = opt.compute_gradients(L)
apply_grad_op = opt.apply_gradients(grads)
init = tf.global_variables_initializer()
saver = tf.train.Saver(tf.global_variables(), max_to_keep=10)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
def train():
"""Train CIFAR-10 for a number of steps."""
with tf.Graph().as_default():
gloabl_step = tf.train.get_or_create_global_step()
# Get images and lables for CIFAR-10
# Force input pipelines to CPU:0 to avoid operations sometimes ending up on GPU and resultign in a slow down.
with tf.device('/cpu:0'):
images, labels = network.distorted_inputs()
# Build a Graph that computes the logits predictions from the inference model.
logits = network.inference(images)
# print(logits.get_shape())
# print(labels.get_shape())
# os.system('pause')
# Calcute loss.
loss = network.loss(logits, labels)
# Buid a Graph that trains the model with one batch of examples and updates the model parameters.
train_op = network.train(loss, gloabl_step)
class _LoggerHook(tf.train.SessionRunHook):
"""Logs loss and runtime"""
def begin(self):
self._step = -1
self._start_time = time.time()
def before_run(self, run_context):
self._step += 1
return tf.train.SessionRunArgs(loss) # Asks for loss value
def after_run(self, run_context, run_values):
if self._step % FLAGS.log_frequency == 0:
current_time = time.time()
duration = current_time - self._start_time
self._start_time = current_time
loss_value = run_values.results
examples_per_sec = FLAGS.log_frequency * FLAGS.batch_size / duration
sec_per_batch = float(duration / FLAGS.log_frequency)
format_str = (
"%s: step %d, loss = %.2f (%.1f exmples/sec: %.3f sec/batch"
)
print(format_str % (datetime.now(), self._step, loss_value,
examples_per_sec, sec_per_batch))
with tf.train.MonitoredTrainingSession(
checkpoint_dir=FLAGS.train_dir,
hooks=[
tf.train.StopAtStepHook(last_step=FLAGS.max_steps),
tf.train.NanTensorHook(loss),
_LoggerHook()
],
config=tf.ConfigProto(log_device_placement=FLAGS.
log_device_placement)) as mon_sess:
print(
'********************************Sussessfully creating session'
)
while not mon_sess.should_stop():
mon_sess.run(train_op)
def main():
print "initial model generator"
with tf.Graph().as_default():
train_sets = dataset.get_datasets(main_path, EPIWidth, disp_precision,
'train')
test_sets = dataset.get_datasets(main_path, EPIWidth, disp_precision,
'test')
global_step = tf.Variable(0, trainable=False)
images_placeholder_v = tf.placeholder(tf.float32,
shape=(None, 9, EPIWidth, 1))
images_placeholder_u = tf.placeholder(tf.float32,
shape=(None, 9, EPIWidth, 1))
labels_placeholder = tf.placeholder(tf.int32, shape=None)
prop_placeholder = tf.placeholder('float')
phase_train = tf.placeholder(tf.bool, name='phase_train')
logits = network.inference_ds(images_placeholder_u,
images_placeholder_v, prop_placeholder,
phase_train, disp_precision)
logits_softmax = network.softmax(logits)
loss = network.loss(logits_softmax, labels_placeholder)
train_op = network.training(loss, 1e-4, global_step)
eval = network.evaluation(logits_softmax)
summary = tf.summary.merge_all()
saver = tf.train.Saver(tf.global_variables())
gpu_option = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_option))
summary_writer = tf.summary.FileWriter(summary_path, sess.graph)
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(checkpoint_path)
if ckpt:
# saver.restore(sess,checkpoint_path+'/model.ckpt')#利用不同平台的训练结果
# saver.restore(sess, ckpt.model_checkpoint_path) # 本地训练的结果
print("restore from checkpoint!")
else:
print("no checkpoint found!")
start_time = time.time()
step = 0
while not train_sets.complete:
feed_dict = fill_feed_dict(train_sets, images_placeholder_u,
images_placeholder_v,
labels_placeholder, prop_placeholder,
phase_train, 'train')
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 1000 == 0:
print('Step:%d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
if step % 25000 == 24999:
print('test Data Eval:')
do_eval_true(sess, eval, logits_softmax, images_placeholder_u,
images_placeholder_v, prop_placeholder,
phase_train, test_sets)
if step % 50000 == 49999:
saver.save(sess,
checkpoint_path + '/model.ckpt',
global_step=step)
LEARNING_RATE = float(os.environ['LEARNING_RATE'] or 0.1)
RESTORE = ((os.environ['RESTORE'] or '') == 'true') or False
learning_rate_value = LEARNING_RATE
session_config = tf.ConfigProto(log_device_placement=True)
session_config.gpu_options.allow_growth = True
# this is required if want to use GPU as device.
# see: https://github.com/tensorflow/tensorflow/issues/2292
session_config.allow_soft_placement = True
if __name__ == "__main__":
with tf.Graph().as_default() as g, tf.device(USE_DEVICE):
# inference()
input, deep_features = network.inference()
labels, logits, cross_entropy = network.loss(deep_features)
centroid_loss, centroids, spread = network.center_loss(
deep_features, labels)
# combine the two losses
_lambda = tf.placeholder(dtype=tf.float32)
total_loss = cross_entropy + _lambda / 2. * centroid_loss
learning_rate, train, global_step = network.training(total_loss)
eval = network.evaluation(logits, labels)
init = tf.initialize_all_variables()
with tf.Session(config=session_config) as sess, \
h5py.File(DUMP_FILE, 'a', libver='latest', swmr=True) as h5_file:
# Merge all the summaries and write them out to /tmp/mnist_logs (by default)
def run_training():
"""Train network for a number of epochs."""
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
with tf.name_scope('input'):
# Input data, pin to CPU because rest of pipeline is CPU-only
with tf.device('/cpu:0'):
input_data = tf.constant(training_data)
input_labels = tf.constant(training_labels)
input, label = tf.train.slice_input_producer(
[input_data, input_labels], num_epochs=FLAGS.num_epochs)
label = tf.cast(label, tf.int32)
input, labels = tf.train.batch([input, label],
batch_size=FLAGS.batch_size)
# Build a Graph that computes predictions from the inference model.
logits = network.inference(input, FLAGS.hidden1, FLAGS.hidden2)
# Add to the Graph the Ops for loss calculation.
loss = network.loss(logits, labels)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = network.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = network.evaluation(logits, labels)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create the op for initializing variables.
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
sess.run(init_op)
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir,
sess.graph)
# Start input enqueue threads.
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# And then after everything is built, start the training loop.
for ep in xrange(FLAGS.num_epochs):
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if loss_value - 0.0 <= 0.00001:
print(
'Loss value: %.4f, done training for %d epochs, %d steps.'
% (loss_value, ep, ep * FLAGS.max_steps + step))
return
if step % 100 == 0:
# Print status to stdout.
print('Epochs %d: loss = %.4f (%.3f sec)' %
(ep, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# Save a checkpoint periodically.
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
print('Saving')
saver.save(sess, FLAGS.train_dir, global_step=step)
def main():
########################################## USER INPUT ##############################################################
# Training parameters:
if len(sys.argv) >= 8:
IMAGE_NAME = sys.argv[1] # IMAGE_NAME = '1'
NETWORK_NAME = sys.argv[2] # 'unet', 'deep_decoder'
LOSS_NAME = sys.argv[
3] # 'mse', 'l1', 'mse_l1', 'mse_with_tv_reg', 'mse_with_edge_reg'
OPTIMIZER_TYPE = sys.argv[4] # 'sgd', 'adam'
LEARNING_RATE = float(sys.argv[5])
NUM_ITERATIONS = int(sys.argv[6])
ITERATIONS_TO_SAVE = int(sys.argv[7])
if len(sys.argv) == 11:
w_h = float(sys.argv[8])
w_v = float(sys.argv[9])
w_mse = float(sys.argv[10])
else:
w_h = None
w_v = None
w_mse = None
else:
print('Not enough input parameters.')
return
####################################################################################################################
# Load images:
RAW_FILENAME = os.path.join('Raw', '{}_Raw Image.tif'.format(IMAGE_NAME))
AVERAGED_FILENAME = os.path.join(
'Averaged', '{}_Averaged Image.tif'.format(IMAGE_NAME))
try:
input_image = hf.get_training_image(RAW_FILENAME)
except:
print("Error loading {}".format(RAW_FILENAME))
return
try:
ground_truth = hf.get_training_image(AVERAGED_FILENAME)
except:
print("Error loading {}".format(AVERAGED_FILENAME))
return
# Validate settings:
VALID_NETWORK_NAMES = ["unet", "deep_decoder"]
VALID_OPTIMIZER_TYPES = ["sgd", "adam"]
VALID_LOSS_NAMES = [
"mse", "l1", "mse_l1", "mse_with_tv_reg", "mse_with_edge_reg"
]
if not (NETWORK_NAME in VALID_NETWORK_NAMES):
print("Error: {} network does not exist.".format(NETWORK_NAME))
return
if not (OPTIMIZER_TYPE in VALID_OPTIMIZER_TYPES):
print("Error: {} optimizer does not exist.".format(OPTIMIZER_TYPE))
return
if not (LOSS_NAME in VALID_LOSS_NAMES):
print("Error: {} loss does not exist.".format(LOSS_NAME))
return
# Create folder to save results:
SAVE_FOLDER = os.path.join('./results', IMAGE_NAME)
count = 0
CHECK_FOLDER = SAVE_FOLDER
while os.path.exists(CHECK_FOLDER):
count += 1
CHECK_FOLDER = '{}({})'.format(SAVE_FOLDER, count)
SAVE_FOLDER = CHECK_FOLDER
os.mkdir(SAVE_FOLDER)
WRITE_FILENAME = os.path.join(SAVE_FOLDER, 'metrics.txt')
with open(WRITE_FILENAME, 'a') as wf:
wf.write(
'PARAMETERS\nNetwork: {}\nLoss: {}\nOptimizer: {}\nLearning rate: {}\nNumber of iterations: {}'
.format(NETWORK_NAME, LOSS_NAME, OPTIMIZER_TYPE, LEARNING_RATE,
NUM_ITERATIONS))
wf.write('\n\nw_h: {}\nw_v: {}\nw_mse: {}'.format(w_h, w_v, w_mse))
wf.write('\n\nIteration\tLoss\tSNR\tCNR\tSSIM')
# Get input noise:
if NETWORK_NAME == "unet":
input_noise = hf.get_noise_matrix(input_image.shape[1],
input_image.shape[2], 32)
elif NETWORK_NAME == "deep_decoder":
input_noise = hf.get_noise_matrix(input_image.shape[1] / (2**4),
input_image.shape[2] / (2**4), 64)
# Save inputs:
save_filename = os.path.join(SAVE_FOLDER, 'input_image.tif')
imsave(save_filename, input_image[0, :, :, 0], cmap='gray')
save_filename = os.path.join(SAVE_FOLDER, 'ground_truth.tif')
imsave(save_filename, ground_truth[0, :, :, 0], cmap='gray')
# Calculate initial metrics:
snr_i = hf.calculate_metrics(ground_truth, input_image, 'snr', IMAGE_NAME)
cnr_i = hf.calculate_metrics(ground_truth, input_image, 'cnr', IMAGE_NAME)
ssim_i = hf.calculate_metrics(ground_truth, input_image, 'ssim',
IMAGE_NAME)
with open(WRITE_FILENAME, 'a') as wf:
wf.write('\ninput_image\tN/A\t{}\t{}\t{}'.format(snr_i, cnr_i, ssim_i))
# Placeholders:
z = tf.placeholder(tf.float32, shape=[1, None, None,
input_noise.shape[3]]) # input noise
x = tf.placeholder(tf.float32, shape=[1, None, None, 1]) # input image
# Network:
y = network.inference(NETWORK_NAME,
z,
height=input_noise.shape[1],
width=input_noise.shape[2],
channels=input_noise.shape[3])
if LOSS_NAME == "mse_with_edge_reg" or LOSS_NAME == "mse_with_tv_reg":
loss, mse, edge_h, edge_v = network.loss(y, x, LOSS_NAME, w_h, w_v,
w_mse)
else:
loss = network.loss(y, x, LOSS_NAME)
# Update moving mean and variance for batch normalization (if required):
if NETWORK_NAME == "deep_decoder":
update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# Create different optimizers here:
if OPTIMIZER_TYPE == "sgd":
train_op = tf.train.GradientDescentOptimizer(
learning_rate=LEARNING_RATE).minimize(loss)
elif OPTIMIZER_TYPE == "adam":
train_op = tf.train.AdamOptimizer(
learning_rate=LEARNING_RATE).minimize(loss)
# Start session:
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# Keep track of metrics:
track_iter = []
track_loss = []
track_snr = []
track_cnr = []
track_ssim = []
for i in range(NUM_ITERATIONS + 1):
if NETWORK_NAME == "unet":
if LOSS_NAME == "mse_with_edge_reg" or LOSS_NAME == "mse_with_tv_reg":
_, output_image, loss_i, mse_i, edge_h_i, edge_v_i = sess.run(
[train_op, y, loss, mse, edge_h, edge_v],
feed_dict={
z: input_noise,
x: input_image
})
else:
_, output_image, loss_i = sess.run([train_op, y, loss],
feed_dict={
z: input_noise,
x: input_image
})
elif NETWORK_NAME == "deep_decoder":
if LOSS_NAME == "mse_with_edge_reg" or LOSS_NAME == "mse_with_tv_reg":
_, _, output_image, loss_i, mse_i, edge_h_i, edge_v_i = sess.run(
[update_op, train_op, y, loss, mse, edge_h, edge_v],
feed_dict={
z: input_noise,
x: input_image
})
else:
_, _, output_image, loss_i = sess.run(
[update_op, train_op, y, loss],
feed_dict={
z: input_noise,
x: input_image
})
if i % ITERATIONS_TO_SAVE == 0:
# Save image:
save_filename = os.path.join(SAVE_FOLDER,
'iteration_{}.tif'.format(i))
imsave(save_filename, output_image[0, :, :, 0], cmap='gray')
# Calculate metrics:
snr_i = hf.calculate_metrics(ground_truth, output_image, 'snr',
IMAGE_NAME)
cnr_i = hf.calculate_metrics(ground_truth, output_image, 'cnr',
IMAGE_NAME)
ssim_i = hf.calculate_metrics(ground_truth, output_image,
'ssim', IMAGE_NAME)
with open(WRITE_FILENAME, 'a') as wf:
wf.write('\n{}\t{}\t{}\t{}\t{}'.format(
i, loss_i, snr_i, cnr_i, ssim_i))
# Display:
if LOSS_NAME == "mse_with_edge_reg" or LOSS_NAME == "mse_with_tv_reg":
print(
'Iteration {}/{}\t| Loss: {}\tSNR: {}\tCNR: {}\tSSIM: {}\tMSE: {}\tEdge_h: {}\tEdge_v: {}'
.format(i, NUM_ITERATIONS, loss_i, snr_i, cnr_i,
ssim_i, mse_i, edge_h_i, edge_v_i))
else:
print(
'Iteration {}/{}\t| Loss: {}\tSNR: {}\tCNR: {}\tSSIM: {}'
.format(i, NUM_ITERATIONS, loss_i, snr_i, cnr_i,
ssim_i))
# Track:
track_iter.append(i)
track_loss.append(loss_i)
track_snr.append(snr_i)
track_cnr.append(cnr_i)
track_ssim.append(ssim_i)
# Plot:
hf.plot_metrics(track_iter, track_loss, 'loss',
os.path.join(SAVE_FOLDER, 'loss.tif'))
hf.plot_metrics(track_iter, track_snr, 'snr',
os.path.join(SAVE_FOLDER, 'snr.tif'))
hf.plot_metrics(track_iter, track_cnr, 'cnr',
os.path.join(SAVE_FOLDER, 'cnr.tif'))
hf.plot_metrics(track_iter, track_ssim, 'ssim',
os.path.join(SAVE_FOLDER, 'ssim.tif'))
print('Completed.')
sys.path.append(os.path.abspath(os.path.join(
os.path.dirname(__file__),
os.path.pardir,
'tracker')))
import network
# Load tensorflow
tf.Graph().as_default()
batchSize = 1
delta = 1
imagePlaceholder = tf.placeholder(tf.float32, shape=(batchSize * delta * 2, 227, 227, 3))
labelsPlaceholder = tf.placeholder(tf.float32, shape=(batchSize * delta, 4))
learningRate = tf.placeholder(tf.float32)
tfOutputs = network.inference(imagePlaceholder, num_unrolls=delta, train=True)
tfLossFull, tfLoss = network.loss(tfOutputs, labelsPlaceholder)
train_op = network.training(tfLossFull, learningRate)
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
summary_writer = tf.summary.FileWriter('logs/train/caffe_copy', sess.graph)
ops = []
with sess.as_default():
sess.run(init)
import caffe
caffe.set_mode_cpu()
# Load caffe net
def main():
args = get_parser().parse_args()
observation_length = 17
action_length = 6
# Read the expert rollouts from disk.
observations, actions = load_data(args.rollouts_file)
print("observations shape = " + str(observations.shape))
print("actions shape = " + str(actions.shape))
# Make sure our files exist!
assert (os.path.exists(os.path.dirname(os.path.abspath(args.stats_file))))
# Load the expert.
print("Loading and building expert policy.")
policy_fn = load_policy.load_policy(args.expert_policy_file)
print("Expert policy loaded and built.")
# Assemble the network.
opl = tf.placeholder(tf.float32,
shape=(None, observation_length),
name="observations")
apl = tf.placeholder(tf.float32,
shape=(None, action_length),
name="actions")
logits = network.inference(opl, observation_length, args.hidden1,
args.hidden2, action_length)
errors, loss = network.loss(logits, apl)
global_step, train_op = network.training(loss, args.learning_rate)
with tf.Session() as sess:
# Initialize the network.
tf_util.initialize()
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(args.checkpoint_dir))
env = gym.make("Walker2d-v1")
max_steps = env.spec.timestep_limit
avg_returns = []
stddev_returns = []
observations = list(observations)
actions = list(actions)
for iteration in range(args.num_iterations):
obs = np.array(observations)
acts = np.array(actions)
assert (obs.shape[0] == acts.shape[0])
# Train the network.
if iteration != 0:
num_batches = int(obs.shape[0] / args.batch_size)
for step in range(args.training_steps):
i = step % num_batches
if i == 0:
p = np.random.permutation(obs.shape[0])
obs = obs[p]
acts = acts[p]
start = int(i * args.batch_size)
stop = int((i + 1) * args.batch_size)
feed_dict = {opl: obs[start:stop], apl: acts[start:stop]}
_, loss_value, step_value = sess.run(
[train_op, loss, global_step], feed_dict=feed_dict)
if step % 100 == 0:
loss_value = sess.run(loss,
feed_dict={
opl: obs,
apl: acts
})
msg = "Iteration {}; step {}; loss = {}".format(
iteration, step_value, loss_value)
print(msg)
# Generate new rollouts.
rewards = []
for i in range(args.num_rollouts):
print("Iteration {}; rollout {}".format(iteration, i))
obs = env.reset()
done = False
steps = 0
totalr = 0
while not done:
expert_action = policy_fn(obs[None, :])
observations.append(obs)
actions.append(expert_action[0])
action = sess.run(logits, feed_dict={opl: obs[None, :]})
obs, r, done, _ = env.step(action)
totalr += r
steps += 1
if steps >= max_steps:
break
rewards.append(totalr)
print("Iteration {}; average return {}".format(
iteration, np.mean(rewards)))
print("Iteration {}; stddev return {}".format(
iteration, np.std(rewards)))
avg_returns.append(np.mean(rewards))
stddev_returns.append(np.std(rewards))
with open(args.stats_file, "w") as f:
stats = {
"mean_return": avg_returns,
"stddev_returns": stddev_returns
}
json.dump(stats, f, indent=4)
['Batch size', batch_size_trn],
['Initial LR',
initial_lr], ['Momentum', momentum_coeff],
['LR Step size', step_size],
['LR Step factor', step_factor],
['Total Steps', total_steps]]
training_images = inputs(args.training_db, batch_size_trn, None, True,
augment_training_data)
test_images = inputs(args.validation_db, batch_size_val, None, False)
net_data = np.load(parameter_path).item()
var_dict = nw.get_variable_dict(net_data)
with tf.variable_scope("ranker") as scope:
feature_vec = nw.build_alexconvnet(training_images, var_dict,
embedding_dim, spp, args.pooling)
L, p = nw.loss(feature_vec, nw.build_loss_matrix(batch_size_trn),
ranking_loss)
scope.reuse_variables()
val_feature_vec = nw.build_alexconvnet(test_images, var_dict,
embedding_dim, spp,
args.pooling)
L_val, p_val = nw.loss(val_feature_vec,
nw.build_loss_matrix(batch_size_val),
ranking_loss)
lr = tf.Variable(initial_lr)
opt = tf.train.AdamOptimizer()
grads = opt.compute_gradients(L)
apply_grad_op = opt.apply_gradients(grads)
init = tf.global_variables_initializer()
LEARNING_RATE = float(os.environ['LEARNING_RATE'] or 0.1)
RESTORE = ((os.environ['RESTORE'] or '') == 'true') or False
learning_rate_value = LEARNING_RATE
session_config = tf.ConfigProto(log_device_placement=True)
session_config.gpu_options.allow_growth = True
# this is required if want to use GPU as device.
# see: https://github.com/tensorflow/tensorflow/issues/2292
session_config.allow_soft_placement = True
if __name__ == "__main__":
with tf.Graph().as_default() as g, tf.device(USE_DEVICE):
# inference()
input, deep_features = network.inference()
labels, logits, cross_entropy = network.loss(deep_features)
centroid_loss, centroids, spread = network.center_loss(deep_features, labels)
# combine the two losses
_lambda = tf.placeholder(dtype=tf.float32)
total_loss = cross_entropy + _lambda / 2. * centroid_loss
learning_rate, train, global_step = network.training(total_loss)
eval = network.evaluation(logits, labels)
init = tf.initialize_all_variables()
with tf.Session(config=session_config) as sess, \
h5py.File(DUMP_FILE, 'a', libver='latest', swmr=True) as h5_file:
# Merge all the summaries and write them out to /tmp/mnist_logs (by default)
# to see the tensor graph, fire up the tensorboard with --logdir="./train"
