sess = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True)))
sess.run(tf.global_variables_initializer())
np.random.seed(int(time.time()))
test_images = mnist.test.images[0:16]
test_origin = test_images*0.5
test_ref,_ = mnist.test.next_batch(16)
test_input = np.minimum(test_origin + test_ref*0.7 + np.random.uniform(size = (16,784)), 1.0)
test_ref = test_ref*0.85
test_origin = test_origin/0.5*0.9
my_lib.mnist_4by4_save(np.reshape(test_input,(-1,784)),file_name + '/input_noise.png')
my_lib.mnist_4by4_save(np.reshape(test_origin,(-1,784)),file_name + '/ground_true.png')
my_lib.mnist_4by4_save(np.reshape(test_ref,(-1,784)),file_name + '/input_ref.png')
log_txt = open(file_name +'/log.txt','w')
content_error = []
mse_error = []
cross_entropy_error = []
for i in range(1000000) :
train_images,_ = mnist.train.next_batch(100)
scale = np.random.uniform(0.2,1.0)
train_origin = train_images * scale
mse = sess.graph.get_tensor_by_name("mse:0")
content_loss = sess.graph.get_tensor_by_name("content_loss:0")
content_optim = sess.graph.get_operation_by_name("content_optim")
np.random.seed(1)
test_images = mnist.test.images[0:16]
test_origin = test_images * 0.5
test_ref, _ = mnist.test.next_batch(16)
test_input = np.minimum(
test_origin + test_ref * 0.7 + np.random.uniform(size=(16, 784)), 1.0)
np.random.seed(int(time.time()))
my_lib.mnist_4by4_save(np.reshape(test_input, (-1, 784)),
file_name + '/1_input_noise.png')
my_lib.mnist_4by4_save(np.reshape(test_origin, (-1, 784)),
file_name + '/1_ground_true.png')
my_lib.mnist_4by4_save(np.reshape(test_ref, (-1, 784)),
file_name + '/1_input_ref.png')
r, val_e = sess.run(
[G_y, content_loss],
feed_dict={
u: np.reshape(test_input, (-1, 28, 28, 1)),
ref: np.reshape(test_ref, (-1, 28, 28, 1)),
t: np.reshape(test_origin, (-1, 28, 28, 1)),
isTrain: False
})
my_lib.mnist_4by4_save(np.reshape(r, (-1, 784)), file_name + '/1_result.png')