## Initialize the variable theta which stores the weights and biases
theta_dncnn = LDAMP.init_vars_DnCNN(init_mu, init_sigma)
## Construct the reconstruction model
#x_hat = LDAMP.DnCNN(y_measured,None,theta_dncnn,training=training_tf)
[x_hat, div_overN] = LDAMP.DnCNN_wrapper(y_measured,
None,
theta_dncnn,
training=training_tf)
## Define loss and optimizer
nfp = np.float32(height_img * width_img)
if useSURE:
cost = LDAMP.MCSURE_loss(x_hat, div_overN, y_measured, sigma_w_tf)
else:
cost = tf.nn.l2_loss(x_true - x_hat) * 1. / nfp
LDAMP.CountParameters()
## Load and Preprocess Training Data
#Training data was generated by GeneratingTrainingImages.m and ConvertImagestoNpyArrays.py
train_images = np.load('./TrainingData/TrainingData_patch' + str(height_img) +
'.npy')
train_images = train_images[range(n_Train_Images), 0, :, :]
assert (len(train_images) >=
n_Train_Images), "Requested too much training data"
val_images = np.load('./TrainingData/ValidationData_patch' + str(height_img) +
'.npy')
[A_handle, At_handle, A_val, A_val_tf] = LDAMP.GenerateMeasurementOperators(measurement_mode)
y_measured = LDAMP.GenerateNoisyCSData_handles(x_true, A_handle, sigma_w, A_val_tf)
## Construct the reconstruction model
if alg=='DAMP':
(x_hat, MSE_history, NMSE_history, PSNR_history, r_final, rvar_final, div_overN) = LDAMP.LDAMP(y_measured,A_handle,At_handle,A_val_tf,theta,x_true,tie=tie_weights,training=training_tf,LayerbyLayer=LayerbyLayer)
elif alg=='DIT':
(x_hat, MSE_history, NMSE_history, PSNR_history) = LDAMP.LDIT(y_measured,A_handle,At_handle,A_val_tf,theta,x_true,tie=tie_weights,training=training_tf,LayerbyLayer=LayerbyLayer)
else:
raise ValueError('alg was not a supported option')
## Define loss and determine which variables to train
nfp = np.float32(height_img * width_img)
if loss_func=='SURE':
assert alg=='DAMP', "Only LDAMP supports training with SURE"
cost = LDAMP.MCSURE_loss(x_hat, div_overN, r_final, tf.sqrt(rvar_final))
elif loss_func=='GSURE':
assert alg == 'DAMP', "Only LDAMP currently supports training with GSURE"
temp0=tf.matmul(A_val_tf,A_val_tf,transpose_b=True)
temp1=tf.matrix_inverse(temp0)
pinv_A=tf.matmul(A_val_tf,temp1,transpose_a=True)
P=tf.matmul(pinv_A,A_val_tf)
#Treat LDAMP/LDIT as a function of A^ty to calculate the divergence
Aty_tf=At_handle(A_val_tf,y_measured)
#Overwrite existing x_hat def
(x_hat, _, _, _, _, _, _) = LDAMP.LDAMP_Aty(Aty_tf, A_handle,At_handle,A_val_tf, theta, x_true,tie=tie_weights,training=training_tf,LayerbyLayer=LayerbyLayer)
if sigma_w==0.:#Not sure if TF is smart enough to avoid computing MCdiv when it doesn't have to
MCdiv=0.
else:
#Calculate MC divergence of P*LDAMP(Aty)
epsilon = tf.maximum(.001 * tf.reduce_max(Aty_tf, axis=0), .00001)