def _test_dataset():
import sflow.py as py
data = dataset_trainA(16)
with tf.feeding() as (sess, coord):
while not coord.should_stop():
py.plt.imshow(data.eval())
if not py.plot_pause():
break
def _test():
import sflow.py as py
data = dataset(16)
with tf.feeding() as (sess, coord):
while not coord.should_stop():
d = sess.run(data)
print(d.shape)
py.plt.imshow(d)
if not py.plt.plot_pause():
break
def train_test():
# model build
data = helen.trainset(batch=16, threads=8)
model = face_parse(data.image, data.label)
optim = tf.train.AdamOptimizer(learning_rate=0.001)
trainop = optim.minimize(model.loss) # train op
tf.summary.scalar('loss', model.loss)
# tf.summary_loss(model.losses)
summary = tf.summary.merge_all()
twriter = tf.summary.FileWriter('/tmp/face_parse/005')
tf.global_variables_initializer().run()
with tf.feeding() as (sess, coord):
i = 0
while True:
s, loss, _ = sess.run([summary, model.loss, trainop])
# loss = sess.run([model.loss, trainop])
print(loss)
twriter.add_summary(s, i)
i += 1
# remove background image
bg = 1. - label[:, :, :, :1]
image = image * bg
return tf.dic(image=image, label=label)
def label_to_rgb(label):
"""
utility for viewing label in a glance
:param label: [b x h x w x 11]
:return: [b x h x w x 3]
"""
pass
if __name__ == '__main__':
import sflow.tf as tf
d = trainset(16)
import matplotlib.pyplot as plt
with tf.feeding() as (sess, coord):
while not coord.should_stop():
out = sess.run(d)
print(out.image.shape)
print(out.label.shape)
plt.imshow(out.image[0])
plt.show()