def val(data_path, image_upsampler, wm_upsampler, blender, downsampler, extrator, summ):
data_file = os.path.join(data_path, 'test_images.tfr')
wm_file = os.path.join(data_path, 'watermark.mat')
assert os.path.isfile(data_file), "Invalid file name"
assert os.path.isfile(wm_file), "Invalid file name"
input_ = Input(FLAGS.batch_size, [FLAGS.img_height, FLAGS.img_width, FLAGS.num_chans])
images = input_(data_file)
wm = Watermark(wm_file)()
image_upsampled = image_upsampler(images)
wm_upsampled = wm_upsampler(wm)
image_blended = blender(image_upsampled, wm_upsampled)
image_downsampled = downsampler(image_blended)
wm_extracted = extrator(image_downsampled)
image_loss = LossRegression()(image_downsampled, images)
wm_loss = LossRegression()(wm_extracted, wm)
summ.register('val', 'image_loss', image_loss)
summ.register('val', 'wm_loss', wm_loss)
val_summ_op = summ('val')
return val_summ_op
def test():
from config import FLAGS
from input_ import Input, Watermark
dim = [
FLAGS.train_batch_size, FLAGS.img_height, FLAGS.img_width,
FLAGS.num_chans
]
image_upsampler = Upsampler(dim)
wm_upsampler = Upsampler([1] + dim[1:])
downsampler = Downsampler(dim, factor=4)
blender = Blender(dim)
extrator = Extractor(dim)
input_ = Input(FLAGS.batch_size,
[FLAGS.img_height, FLAGS.img_width, FLAGS.num_chans])
images = input_('/data/yuming/watermark-data/train_images.tfr')
wm = Watermark('/data/yuming/watermark-data/watermark.mat')()
image_upsampled = image_upsampler(images)
wm_upsampled = wm_upsampler(wm)
image_blended = blender(image_upsampled, wm_upsampled)
image_downsampled = downsampler(image_blended)
wm_extracted = extrator(image_downsampled)
writer = tf.summary.FileWriter("model-output", tf.get_default_graph())
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
v = sess.run(wm_extracted)
print(v.shape)
writer.close()
def train(data_path, image_upsampler, wm_upsampler, blender, downsampler, extrator, summ):
data_file = os.path.join(data_path, 'train_images.tfr')
wm_file = os.path.join(data_path, 'watermark.mat')
assert os.path.isfile(data_file), "Invalid file name"
assert os.path.isfile(wm_file), "Invalid file name"
input_ = Input(FLAGS.batch_size, [FLAGS.img_height, FLAGS.img_width, FLAGS.num_chans])
images = input_(data_file)
wm = Watermark(wm_file)()
image_upsampled = image_upsampler(images)
wm_upsampled = wm_upsampler(wm)
image_blended = blender(image_upsampled, wm_upsampled)
image_downsampled = downsampler(image_blended, training = True)
wm_extracted = extrator(image_downsampled, training = True)
image_loss = LossRegression()(image_downsampled, images)
wm_loss = LossRegression()(wm_extracted, wm)
opt = Adam(FLAGS.learning_rate, lr_decay = FLAGS.lr_decay, lr_decay_steps = FLAGS.lr_decay_steps,
lr_decay_factor = FLAGS.lr_decay_factor)
train_op = opt(image_loss + wm_loss)
summ.register('train', 'image_loss', image_loss)
summ.register('train', 'wm_loss', wm_loss)
train_summ_op = summ('train')
return image_loss + wm_loss, train_op, train_summ_op
def xval(model, data_path, summ):
input_ = Input(FLAGS.xval_batch_size, [FLAGS.img_height, FLAGS.img_width, FLAGS.num_channels])
images, labels, = input_(data_path)
logits = model(images, training = False)
xval_accu_op = XValClassification(gpu = True)(logits, labels)
summ.register('xval', 'xval_accuracy', xval_accu_op)
xval_summ_op = summ('xval')
return xval_accu_op, xval_summ_op
def val(data_path, model, summ):
input_ = Input(FLAGS.batch_size, [FLAGS.num_chans, FLAGS.num_points])
data, labels = input_(data_path)
'''
print_op = tf.print("In validation procedure -> label shape: ", tf.shape(labels), output_stream = sys.stdout)
with tf.control_dependencies([print_op]):
logits = model(data)
'''
logits = model(data)
loss = LossRegression()(logits, tf.expand_dims(labels, axis=-1))
summ.register('val', 'loss', loss)
val_summ_op = summ('val')
return val_summ_op
def train(model, data_path, cur_iter, summ):
input_ = Input(FLAGS.train_batch_size, [FLAGS.img_height, FLAGS.img_width, FLAGS.num_channels])
images, labels = input_(data_path)
logits = model(images, cur_iter = cur_iter)
# Calculate the loss of the model.
loss = LossClassification(FLAGS.num_classes, gpu = True)(logits, labels)
# Create an optimizer that performs gradient descent.
optimizer = Adam(FLAGS.learning_rate, lr_decay = False,
lr_decay_steps = FLAGS.lr_decay_steps,
lr_decay_factor = FLAGS.lr_decay_factor)
train_op = optimizer(loss)
summ.register('train', 'train_loss', loss)
summ.register('train', 'learning_rate', optimizer.lr)
train_summ_op = summ('train')
return loss, train_op, train_summ_op
def train(data_path, model, summ):
input_ = Input(FLAGS.batch_size, [FLAGS.num_chans, FLAGS.num_points])
data, labels = input_(data_path)
'''
print_op = tf.print("In train procedure -> label shape: ", tf.shape(labels), output_stream = sys.stdout)
with tf.control_dependencies([print_op]):
logits = model(data)
'''
logits = model(data)
loss = LossRegression()(logits, tf.expand_dims(labels, axis=-1))
opt = Adam(FLAGS.learning_rate,
lr_decay=FLAGS.lr_decay,
lr_decay_steps=FLAGS.lr_decay_steps,
lr_decay_factor=FLAGS.lr_decay_factor)
train_op = opt(loss)
summ.register('train', 'loss', loss)
summ.register('train', 'learning_rate', opt.lr)
train_summ_op = summ('train')
return loss, train_op, train_summ_op
def main(unused_argv):
tf.logging.set_verbosity(tf.logging.ERROR)
# tf.logging.set_verbosity(3) # Print INFO log messages.
summaries = []
assert FLAGS.test_size > 0, 'batch size for testing'
if FLAGS.data_dir == '' or not os.path.exists(FLAGS.data_dir):
raise ValueError('invalid data directory {}'.format(FLAGS.data_dir))
test_data_path = os.path.join(FLAGS.data_dir, 'quantum-test.tfr')
if FLAGS.output_dir == '':
raise ValueError('invalid output directory {}'.format(
FLAGS.output_dir))
if FLAGS.data_type == "momentum":
filenames_datafile = '/scratch/quantum-meta/filenames_transition_momentum.dat'
warnings.warn(
"Check the data file containing file names: {}".format(
filenames_datafile), UserWarning)
with open(filenames_datafile, 'rb') as f:
filenames = pickle.load(f)
elif FLAGS.data_type == "position":
filenames_datafile = '/scratch/quantum-meta/filenames_transition_position.dat'
warnings.warn(
"Check the data file containing file names: {}".format(
filenames_datafile), UserWarning)
with open(filenames_datafile, 'rb') as f:
filenames = pickle.load(f)
else:
print("invalid data type {}".format(FLAGS.data_type))
return
test_filenames = filenames['test_filenames']
checkpoint_dir = os.path.join(FLAGS.output_dir, '')
test_input = Input(1,
[FLAGS.img_height, FLAGS.img_width, FLAGS.num_channels])
inputs, labels = test_input(test_data_path)
model = Model(act=FLAGS.activation,
pool=FLAGS.pooling,
with_memory=FLAGS.with_memory,
log=False)
inputs, labels = test_input(test_data_path)
logits = model(inputs, training=False)
logit_indices = tf.argmax(logits, axis=-1)
# Define the metric and update operations
metric_op, metric_update_op = tf.metrics.accuracy(labels, logit_indices)
saver = tf.train.Saver(tf.trainable_variables())
with tf.Session() as sess:
# sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])
sess.run([tf.local_variables_initializer()])
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
# Restores from checkpoint.
saver.restore(sess, ckpt.model_checkpoint_path)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/imagenet_train/model.ckpt-0,
# extract global_step from it.
# global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
# print('Successfully loaded model from %s at step = %s.' %
# (ckpt.model_checkpoint_path, global_step))
print('Successfully loaded model from %s.' %
ckpt.model_checkpoint_path)
else:
print('No checkpoint file found')
return
logits_lst = []
labels_lst = []
indices_lst = []
image_data_placeholder = tf.placeholder(
tf.float32,
[FLAGS.img_height, FLAGS.img_width, FLAGS.num_channels])
image_descr_placeholder = tf.placeholder(tf.string, [],
name="image_descr")
image_data = tf.unstack(image_data_placeholder, axis=-1)
image_data = [(data - tf.reshape(tf.reduce_min(tf.reduce_min(data)), [1, 1])) / \
tf.reshape((tf.reduce_max(tf.reduce_max(data)) -tf.reduce_min(tf.reduce_min(data))), [1, 1]) \
for data in image_data]
image_data = [tf.expand_dims(data, axis=-1) for data in image_data]
image_data = [tf.expand_dims(data, axis=0) for data in image_data]
add_img_op = [
tf.summary.image(
"misclassified image data {}th component".format(i),
image_data[i]) for i in range(FLAGS.num_channels)
]
add_txt_op = [
tf.summary.text(
"misclassified image description {}th component".format(i),
image_descr_placeholder) for i in range(FLAGS.num_channels)
]
add_summ = tf.summary.merge_all()
writer = tf.summary.FileWriter("output-outlier",
tf.get_default_graph())
logit_indices = tf.squeeze(logit_indices)
labels = tf.squeeze(labels)
for i in range(FLAGS.test_size):
_, images_val, logits_val, indices_val, labels_val = sess.run(
[metric_update_op, inputs, logits, logit_indices, labels])
logits_lst.append(logits_val)
labels_lst.append(labels_val)
indices_lst.append(indices_val)
# print("{}: ground-truth label: {}, predicted label: {}".format(i, labels_val, indices_val))
if indices_val != labels_val:
print("{}: ground-truth {}, predicted {}".format(
i, labels_val, indices_val))
filename = test_filenames[i]
filename = filename.split('\\')[-1]
print("'{}'".format(filename))
summ_str = sess.run(
add_summ,
feed_dict={
image_data_placeholder:
np.squeeze(images_val, axis=0),
image_descr_placeholder:
"{}: ground-truth {}, predicted {}".format(
filename, labels_val, indices_val)
})
writer.add_summary(summ_str, i)
accu = sess.run(metric_op)
print("accu -> {}".format(accu))
# io.savemat("logits.mat", {"logits": logits_lst, 'classes': labels_lst, 'indices': indices_lst})
writer.close()
def main(unused_argv):
tf.logging.set_verbosity(tf.logging.ERROR)
# tf.logging.set_verbosity(3) # Print INFO log messages.
summaries = []
assert FLAGS.test_size > 0, 'batch size for testing'
if FLAGS.data_dir == '' or not os.path.exists(FLAGS.data_dir):
raise ValueError('invalid data directory {}'.format(FLAGS.data_dir))
test_data_path = os.path.join(FLAGS.data_dir, 'quantum-test.tfr')
if FLAGS.output_dir == '':
raise ValueError('invalid output directory {}'.format(
FLAGS.output_dir))
checkpoint_dir = os.path.join(FLAGS.output_dir, '')
test_input = Input(1,
[FLAGS.img_height, FLAGS.img_width, FLAGS.num_channels])
inputs, labels = test_input(test_data_path)
model = Model(act=FLAGS.activation,
pool=FLAGS.pooling,
with_memory=FLAGS.with_memory,
log=False)
inputs, labels = test_input(test_data_path)
logits = model(inputs, training=False)
logit_indices = tf.argmax(logits, axis=-1)
# Define the metric and update operations
metric_op, metric_update_op = tf.metrics.accuracy(labels, logit_indices)
saver = tf.train.Saver(tf.trainable_variables())
with tf.Session() as sess:
# sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])
sess.run([tf.local_variables_initializer()])
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
# Restores from checkpoint.
saver.restore(sess, ckpt.model_checkpoint_path)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/imagenet_train/model.ckpt-0,
# extract global_step from it.
# global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
# print('Successfully loaded model from %s at step = %s.' %
# (ckpt.model_checkpoint_path, global_step))
print('Successfully loaded model from %s.' %
ckpt.model_checkpoint_path)
else:
print('No checkpoint file found')
return
logits_lst = []
labels_lst = []
indices_lst = []
for i in range(FLAGS.test_size):
_, logits_val, indices_val, labels_val = sess.run(
[metric_update_op, logits, logit_indices, labels])
logits_lst.append(logits_val)
labels_lst.append(labels_val)
indices_lst.append(indices_val)
accu = sess.run(metric_op)
print("accu -> {}".format(accu))
io.savemat("tsne.mat", {
"logits": logits_lst,
'classes': labels_lst,
'indices': indices_lst
})
eng = matlab.engine.start_matlab("-nodisplay -nosplash")
accus = eng.calc_class_accu(
matlab.int32(np.squeeze(labels_lst).tolist()),
matlab.int32(np.squeeze(indices_lst).tolist()))
eng.quit()
for i, a in enumerate(np.squeeze(np.array(accus)).tolist()):
print("C = {}: {}".format(i - 2, a))