saver_best = tf.train.Saver() # defaults to saving all variables
saver_dict = {}
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
sess.run(
tf.global_variables_initializer()
) #Seems to be necessary for the batch normalization layers for some reason.
# if FLAGS.debug:
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# sess.add_tensor_filter("has_inf_or_nan", tf_debug.has_inf_or_nan)
start_time = time.time()
print("Load Initial Weights ...")
if ResumeTraining or learning_rate != learning_rates[0]:
##Load previous values for the weights and BNs
saver_initvars_name_chckpt = LDAMP.GenDnCNNFilename(
sigma_w_min, sigma_w_max, useSURE=useSURE) + ".ckpt"
for l in range(0, n_DnCNN_layers):
saver_dict.update({"l" + str(l) + "/w": theta_dncnn[0][l]})
for l in range(1, n_DnCNN_layers -
1): #Associate variance, means, and beta
gamma_name = "l" + str(l) + "/BN/gamma:0"
beta_name = "l" + str(l) + "/BN/beta:0"
var_name = "l" + str(l) + "/BN/moving_variance:0"
mean_name = "l" + str(l) + "/BN/moving_mean:0"
gamma = [
v for v in tf.global_variables() if v.name == gamma_name
][0]
beta = [
v for v in tf.global_variables() if v.name == beta_name
][0]
moving_variance = [
beta_name = "Iter" + str(iter) + "/l" + str(l) + "/BN/beta:0"
var_name = "Iter" + str(iter) + "/l" + str(l) + "/BN/moving_variance:0"
mean_name = "Iter" + str(iter) + "/l" + str(l) + "/BN/moving_mean:0"
gamma = [v for v in tf.global_variables() if v.name == gamma_name][0]
beta = [v for v in tf.global_variables() if v.name == beta_name][0]
moving_variance = [v for v in tf.global_variables() if v.name == var_name][0]
moving_mean = [v for v in tf.global_variables() if v.name == mean_name][0]
saver_dict.update({"Iter" + str(iter) + "/l" + str(l) + "/BN/gamma": gamma})
saver_dict.update({"Iter" + str(iter) + "/l" + str(l) + "/BN/beta": beta})
saver_dict.update({"Iter" + str(iter) + "/l" + str(l) + "/BN/moving_variance": moving_variance})
saver_dict.update({"Iter" + str(iter) + "/l" + str(l) + "/BN/moving_mean": moving_mean})
saver_initvars = tf.train.Saver(saver_dict)
saver_initvars.restore(sess, saver_initvars_name_chckpt)
if InitWeightsMethod=='denoiser':
#load initial weights that were trained on a denoising problem
saver_initvars_name_chckpt=LDAMP.GenDnCNNFilename(300./255.,500./255.)+".ckpt"
iter = 0
for l in range(0, n_DnCNN_layers):
saver_dict.update({"l" + str(l) + "/w": theta[iter][0][l]})#, "l" + str(l) + "/b": theta[iter][1][l]})
for l in range(1,n_DnCNN_layers-1):#Associate variance, means, and beta
gamma_name = "Iter"+str(iter)+"/l" + str(l) + "/BN/gamma:0"
beta_name="Iter"+str(iter)+"/l" + str(l) + "/BN/beta:0"
var_name="Iter"+str(iter)+"/l" + str(l) + "/BN/moving_variance:0"
mean_name="Iter"+str(iter)+"/l" + str(l) + "/BN/moving_mean:0"
gamma = [v for v in tf.global_variables() if v.name == gamma_name][0]
beta = [v for v in tf.global_variables() if v.name == beta_name][0]
moving_variance = [v for v in tf.global_variables() if v.name == var_name][0]
moving_mean = [v for v in tf.global_variables() if v.name == mean_name][0]
saver_dict.update({"l" + str(l) + "/BN/gamma": gamma})
saver_dict.update({"l" + str(l) + "/BN/beta": beta})
saver_dict.update({"l" + str(l) + "/BN/moving_variance": moving_variance})
test_images = test_images[:, 0, :, :]
assert (len(test_images) >= n_Test_Images), "Requested too much Test data"
x_test = np.transpose(
np.reshape(test_images, (-1, height_img * width_img * channel_img)))
# with tf.Session() as sess:
# y_test=sess.run(y_measured,feed_dict={x_true: x_test, A_val_tf: A_val})#All the batches will use the same measurement matrix
## Test the Model
saver = tf.train.Saver() # defaults to saving all variables
saver_dict = {}
with tf.Session() as sess:
if tie_weights == 1: # Load weights from pretrained denoiser
save_name = LDAMP.GenDnCNNFilename(80. / 255.) + ".ckpt"
for l in range(0, n_DnCNN_layers):
saver_dict.update({"l" + str(l) + "/w": theta[0][0][l]
}) #, "l" + str(l) + "/b": theta[0][1][l]})
for l in range(1, n_DnCNN_layers -
1): # Associate variance, means, and beta
gamma_name = "Iter" + str(0) + "/l" + str(l) + "/BN/gamma:0"
beta_name = "Iter" + str(0) + "/l" + str(l) + "/BN/beta:0"
var_name = "Iter" + str(0) + "/l" + str(
l) + "/BN/moving_variance:0"
mean_name = "Iter" + str(0) + "/l" + str(l) + "/BN/moving_mean:0"
gamma = [v for v in tf.global_variables()
if v.name == gamma_name][0]
beta = [v for v in tf.global_variables() if v.name == beta_name][0]
moving_variance = [
v for v in tf.global_variables() if v.name == var_name
# sigma_w_min = 80.
# sigma_w_max = 100.
# elif 255.*sigma_w < 150.:
# sigma_w_min = 100.
# sigma_w_max = 150.
# elif 255.*sigma_w < 300.:
# sigma_w_min = 150.
# sigma_w_max = 300.
# else:
# sigma_w_min = 300.
# sigma_w_max = 500.
sigma_w_min = sigma_w * 255.
sigma_w_max = sigma_w * 255.
save_name = LDAMP.GenDnCNNFilename(sigma_w_min / 255.,
sigma_w_max / 255.,
useSURE=useSURE)
save_name_chckpt = save_name + ".ckpt"
saver.restore(sess, save_name_chckpt)
print("Reconstructing Signal")
start_time = time.time()
[reconstructed_test_images] = sess.run([x_hat],
feed_dict={y_measured: y_test})
time_taken = time.time() - start_time
fig1 = plt.figure()
plt.imshow(np.transpose(np.reshape(x_test[:, 0], (height_img, width_img))),
interpolation='nearest',
cmap='gray')
plt.show()
fig2 = plt.figure()