# train
obj_class += model.pretraining_class(sess, batch_figs,
batch_labels)
obj_disc += model.pretraining_disc(sess, batch_figs,
batch_labels, batch_z)
obj_vae += model.pretraining_vae(sess, batch_figs,
batch_labels)
obj_dec += model.pretraining_dec(sess, batch_figs,
batch_labels, batch_z)
if step % 10 == 0:
print(' step {}/{} end'.format(step, num_one_epoch))
sys.stdout.flush()
tmp_z = model.encoding(sess, batch_figs, batch_labels)
tmp_figs = model.gen_fig(sess, batch_labels, tmp_z)
dump_figs(np.asarray(tmp_figs), 'sample_result')
print('epoch:{}, v_obj = {}, dec_obj = {}, disc_obj = {}'.format(
epoch, obj_vae / num_one_epoch, obj_dec / num_one_epoch,
obj_disc / num_one_epoch))
saver.save(sess, './model.dump')
# -- main training ---------------
for epoch in range(epoch_num):
epoch_figs, epoch_labels = shuffle(figs, labels)
print('** epoch {} begin **'.format(epoch))
obj_class, obj_vae, obj_dec, obj_disc = 0.0, 0.0, 0.0, 0.0
for step in range(num_one_epoch):
def __call__(self, z_inputs):
assert(len(z_inputs) == self.batch_size)
return self.model.gen_fig(self.sess, z_inputs)
if __name__ == u'__main__':
# dump file
dump_file = u'./model.dump'
# parameter
batch_size = 1
z_dim = 100
# figure generator
fig_gen = FigGenerator(dump_file, z_dim, batch_size)
# make figure
fig1 = cv2.imread('./input1.jpg')
fig1 = fig1/127.5 - 1.0
fig2 = cv2.imread('./input2.jpg')
fig2 = fig2/127.5 - 1.0
z1 = fig_gen.encoding([fig1])[0]
z2 = fig_gen.encoding([fig2])[0]
diff = (z2 - z1)/50.0
zs = []
for i in range(51):
z_target = z1 + diff * float(i)
zs.append(fig_gen([z_target])[0])
dump_figs(np.asarray(zs), 'morphing_result')