def train():
is_training_v = tf.placeholder(tf.bool, shape=[])
phv = {'is_training_v': is_training_v}
with tf.Session() as sess:
fa_images = V_graph(sess, phv)
saver = tf.train.Saver(Genc.vars + Gdec.vars)
init = tf.global_variables_initializer()
sess.run(init)
saver.restore(sess, LOG_PATH + '/weight.ckpt')
##################### generate data #####################
v_dict = {phv['is_training_v']: True}
cnt = 0
for i in range(train_iters):
sample = sess.run(fa_images, feed_dict=v_dict)
sample = util.to_uint8(sample)
for s in sample:
cnt += 1
io.imsave('%s/%06d.jpg' % (OUT_PATH, cnt), s, quality=95)
def saveTextures(outData, textures, out_path, loadFromDirectory):
"""
Save the textures to the specified path.
Inputs:
outData - height x width x depth array for the output texture(s)
textures - list of InputTextureFile objects
out_path - path to output
loadFromDirectory - is the output path a directory?
Output:
Saves textures.
"""
assert len(outData.shape) == 3
out_dir = os.path.dirname(out_path)
if (out_dir != "" and not os.path.exists(out_dir)):
os.makedirs(out_dir)
current_depth = 0
for textureFile in textures:
next_depth = current_depth + textureFile.depth
textureData = outData[:, :, current_depth:next_depth]
if (textureData.shape[2] < 2):
textureData = numpy.squeeze(textureData, axis=2)
if (not textureFile.isFloat):
textureData = to_uint8(textureData, normalize=False)
# Save the solved texture
if (loadFromDirectory):
base, ext = os.path.splitext(textureFile.name)
out_texture = os.path.join(out_path, base + "-erased" + ext)
else:
out_texture = out_path
if (textureFile.isDataFile):
weight_data.write_tex_to_path(out_texture, textureData)
else:
texture.save_texture(textureData, out_texture)
current_depth = next_depth
def train():
lr_g = tf.placeholder(tf.float32, shape=[])
lr_d = tf.placeholder(tf.float32, shape=[])
is_training_g = tf.placeholder(tf.bool, shape=[])
is_training_d = tf.placeholder(tf.bool, shape=[])
is_training_v = tf.placeholder(tf.bool, shape=[])
is_training_a = tf.placeholder(tf.bool, shape=[])
phg = {'lr_g': lr_g,
'is_training_g': is_training_g}
phd = {'lr_d': lr_d,
'is_training_d': is_training_d}
phv = {'is_training_v': is_training_v}
pha = {'is_training_a': is_training_a}
with tf.Session() as sess:
G_u, G_loss, G_update = G_graph(sess, phg)
Ga_u, Ga_loss, Ga_acc, Ga_update = G_adv_graph(sess, phg)
D_loss, D_update = D_graph(sess, phd)
tr_images, re_images, fa_images = V_graph(sess, phv)
adv_loss, adv_acc, adv_update, adv_pred = Adv_graph(sess, pha)
saver = tf.train.Saver(Genc.vars+Gdec.vars)
init = tf.global_variables_initializer()
sess.run(init)
# saver.restore(sess,LOG_PATH+'/weight.ckpt')
##################### pre-train Generator & Discriminator #####################
g_dict = {phg['lr_g']: 1e-4,
phg['is_training_g']: True}
d_dict = {phd['lr_d']: 1e-4,
phd['is_training_d']: True}
for epoch in range(N_EPOCH):
for i in range(train_iters):
if i % 6 !=0:
d_loss, _ = sess.run([D_loss, D_update], feed_dict=d_dict)
else:
g_loss, _ = sess.run([G_loss, G_update], feed_dict=g_dict)
if i % 100 == 0 and i != 0:
log_string('====pre_epoch_%d====iter_%d: g_loss=%.4f, d_loss=%.4f' % (epoch, i, g_loss, d_loss))
saver.save(sess,LOG_PATH+'/pretrain.ckpt')
##################### train and fix adversarial classifier #####################
a_dict = {pha['is_training_a']: True}
for epoch in range(N_ADV):
for i in range(train_iters):
a_loss, a_acc, _ = sess.run([adv_loss, adv_acc, adv_update], feed_dict=a_dict)
if i % 100 == 0 and i != 0:
log_string('====adv_epoch_%d====iter_%d: a_loss=%.4f, a_acc=%.4f' % (epoch, i, a_loss, a_acc))
##################### fine-tune Generator & Discriminator #####################
g_dict = {phg['lr_g']: 5e-6,
phg['is_training_g']: True}
d_dict = {phd['lr_d']: 5e-6,
phd['is_training_d']: True}
v_dict = {phv['is_training_v']: True}
for epoch in range(N_EPOCH):
for i in range(train_iters):
if i % 6 !=0:
d_loss, _ = sess.run([D_loss, D_update], feed_dict=d_dict)
else:
g_loss, a_acc, _ = sess.run([Ga_loss, Ga_acc, Ga_update], feed_dict=g_dict)
if i % 100 == 0 and i != 0:
log_string('====epoch_%d====iter_%d: g_loss=%.4f, d_loss=%.4f, a_acc=%.4f' % (epoch, i, g_loss, d_loss, a_acc))
a,b,c = sess.run([tr_images, re_images, fa_images], feed_dict = v_dict)
image_list = [a,b,c]
sample = np.transpose(image_list, (1, 2, 0, 3, 4))
sample = np.reshape(sample, (-1, sample.shape[2] * sample.shape[3], sample.shape[4]))
sample = util.to_uint8(sample)
io.imsave('%s/Epoch-%d.jpg' % (LOG_PATH, epoch), sample, quality=95)
saver.save(sess,LOG_PATH+'/weight.ckpt')