def testAll():
## test all the testing samples
iniA = 0
iniUp3x3 = 0
if FLAGS.Ynormalize:
data = sio.loadmat("CAVEdata/Ynum")
Ynum = data['Ynum']
else:
Ynum = 1
Y = tf.placeholder(tf.float32, shape=(1, 512, 512, 3)) # supervised data (None,64,64,3)
Z = tf.placeholder(tf.float32, shape=(1, 512/32, 512/32, FLAGS.outDim))
outX, X1, YA, _, HY = MHFnet.HSInet(Y, Z, iniUp3x3,iniA,FLAGS.upRank,FLAGS.outDim,FLAGS.HSInetL,FLAGS.subnetL)
config = tf.ConfigProto(allow_soft_placement=True,log_device_placement=True)
config.gpu_options.allow_growth = True
saver = tf.train.Saver(max_to_keep = 5)
save_path = FLAGS.train_dir
ML.mkdir(FLAGS.test_dir)
with tf.Session(config=config) as sess:
ckpt = tf.train.latest_checkpoint(save_path)
saver.restore(sess, ckpt)
for root, dirs, files in os.walk('CAVEdata/X/'):
for i in range(32):
data = sio.loadmat("CAVEdata/Y/"+files[i])
inY = data['RGB']
inY = np.expand_dims(inY, axis = 0)*Ynum
data = sio.loadmat("CAVEdata/Z/"+files[i])
inZ = data['Zmsi']
inZ = np.expand_dims(inZ, axis = 0)
pred_X,ListX,pred_HY,pred_YA = sess.run([outX, X1, HY, YA],feed_dict={Y:inY,Z:inZ})
pred_Lr = ListX[FLAGS.HSInetL-2]
sio.savemat(FLAGS.test_dir+files[i], {'outX': pred_X,'outLR': pred_Lr,'outHY': pred_HY, 'outYA':pred_YA})
print(files[i] + ' done!')
def test():
data = sio.loadmat(FLAGS.test_data_name)
Y = data['RGB']
Z = data['Zmsi']
X = data['msi']
if FLAGS.Ynormalize:
Ynum = 1
# Depending on the situation, this line may need to be changed as following:
# data = sio.loadmat("CAVEdata/Ynum")
# Ynum = data['Ynum']
else:
Ynum = 1
## banchsize H W C
inY = np.expand_dims(Y, axis = 0)
inY = tf.to_float(inY)*Ynum;
inZ = np.expand_dims(Z, axis = 0)
inZ = tf.to_float(inZ);
inX = np.expand_dims(X, axis = 0)
inX = tf.to_float(inX);
iniA = 0
iniUp3x3 = 0
outX, X1, YA, _, HY = MHFnet.HSInet(inY,inZ, iniUp3x3, iniA,FLAGS.upRank,FLAGS.outDim,FLAGS.HSInetL,FLAGS.subnetL)
config = tf.ConfigProto(allow_soft_placement=True,log_device_placement=True)
config.gpu_options.allow_growth = True
saver = tf.train.Saver(max_to_keep = 5)
save_path = FLAGS.train_dir
with tf.Session(config=config) as sess:
ckpt = tf.train.latest_checkpoint(save_path)
saver.restore(sess, ckpt)
pred_X,pred_YA,pred_HY,inX = sess.run([outX, YA, HY, inX])
toshow = np.hstack((ML.normalized(ML.get3band_of_tensor(pred_HY)),ML.get3band_of_tensor(pred_YA)))
toshow2 = np.hstack((ML.get3band_of_tensor(pred_X),ML.get3band_of_tensor(inX)))
toshow = np.vstack((toshow,toshow2))
print('The vasaul result of Y_hat (left upper), Y*A (right upper), fusion result (left lower) and ground truth (right lower)')
ML.imshow(toshow)
ML.imwrite(toshow)
def test():
data = sio.loadmat(FLAGS.test_data_name)
Y = data['RGB']
Z = data['Zmsi']
X = data['msi']
## banchsize H W C
inY = np.expand_dims(Y, axis=0)
inY = tf.to_float(inY)
inZ = np.expand_dims(Z, axis=0)
inZ = tf.to_float(inZ)
inX = np.expand_dims(X, axis=0)
inX = tf.to_float(inX)
iniA = 0
iniUp3x3 = 0
outX, X1, YA, _, HY = MHFnet.HSInet(inY, inZ, iniUp3x3, iniA, FLAGS.upRank,
FLAGS.outDim, FLAGS.HSInetL,
FLAGS.subnetL)
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
config.gpu_options.allow_growth = True
saver = tf.train.Saver(max_to_keep=5)
save_path = FLAGS.train_dir
with tf.Session(config=config) as sess:
ckpt = tf.train.latest_checkpoint(save_path)
saver.restore(sess, ckpt)
pred_X, pred_YA, pred_HY, inX = sess.run([outX, YA, HY, inX])
toshow = np.hstack((ML.normalized(ML.get3band_of_tensor(pred_HY)),
ML.get3band_of_tensor(pred_YA)))
toshow2 = np.hstack(
(ML.get3band_of_tensor(pred_X), ML.get3band_of_tensor(inX)))
toshow = np.vstack((toshow, toshow2))
ML.imshow(toshow)
ML.imwrite(toshow)
def train():
Crd.PrepareDataAndiniValue()
random.seed(1)
## 变为4D张量 banchsize H W C
iniData1 = sio.loadmat("CAVEdata/iniA")
iniA = iniData1['iniA']
iniData2 = sio.loadmat("CAVEdata/iniUp")
iniUp3x3 = iniData2['iniUp1']
X = tf.placeholder(tf.float32,
shape=(None, FLAGS.image_size, FLAGS.image_size,
FLAGS.outDim)) # HrHS (None,96,96,31)
Y = tf.placeholder(tf.float32,
shape=(None, FLAGS.image_size, FLAGS.image_size,
3)) # HrMS (None,96,96,3)
Z = tf.placeholder(tf.float32,
shape=(None, FLAGS.image_size / 32,
FLAGS.image_size / 32,
FLAGS.outDim)) # LrHS (None,3,3,31)
outX, ListX, YA, E, HY = MHFnet.HSInet(Y, Z, iniUp3x3, iniA, FLAGS.upRank,
FLAGS.outDim, FLAGS.HSInetL,
FLAGS.subnetL)
# loss function
loss = tf.reduce_mean(tf.square(X - outX)) + FLAGS.alpha * tf.reduce_mean(
tf.square(X - YA)) + FLAGS.beta * tf.reduce_mean(
tf.square(E)) # supervised MSE loss
for i in range(FLAGS.HSInetL - 1):
loss = loss + FLAGS.alpha * tf.reduce_mean(tf.square(X - ListX[i]))
lr_ = FLAGS.learning_rate
lr = tf.placeholder(tf.float32, shape=[])
g_optim = tf.train.AdamOptimizer(lr).minimize(
loss) # Optimization method: Adam
# saver setting
saver = tf.train.Saver(max_to_keep=5)
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
save_path = FLAGS.train_dir
ML.mkdir(save_path)
epoch = int(FLAGS.epoch)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
if tf.train.get_checkpoint_state(
'temp/TrainedNet/'): # load previous trained model
ckpt = tf.train.latest_checkpoint('temp/TrainedNet/')
saver.restore(sess, ckpt)
ckpt_num = re.findall(r"\d", ckpt)
if len(ckpt_num) == 3:
start_point = 100 * int(ckpt_num[0]) + 10 * int(
ckpt_num[1]) + int(ckpt_num[2])
elif len(ckpt_num) == 2:
start_point = 10 * int(ckpt_num[0]) + int(ckpt_num[1])
else:
start_point = int(ckpt_num[0])
print("Load success")
else:
print("re-training")
start_point = 0
allX, allY = Crd.all_train_data_in()
val_h5_X, val_h5_Y, val_h5_Z = Crd.eval_data_in(20)
for j in range(start_point, epoch):
if j + 1 > (4 * epoch / 5):
lr_ = FLAGS.learning_rate * 0.1
Training_Loss = 0
for num in range(FLAGS.BatchIter):
batch_X, batch_Y, batch_Z = Crd.train_data_in(
allX, allY, FLAGS.image_size, FLAGS.batch_size)
_, lossvalue = sess.run([g_optim, loss],
feed_dict={
X: batch_X,
Y: batch_Y,
Z: batch_Z,
lr: lr_
})
Training_Loss += lossvalue # training loss
# visual output
_, ifshow = divmod(num + 1, 200)
if ifshow == 1:
pred_X, pred_ListX, pred_HY, Pred_YA = sess.run(
[outX, ListX, HY, YA],
feed_dict={
Y: batch_Y,
Z: batch_Z
})
psnr = skimage.measure.compare_psnr(batch_X, pred_X)
ssim = skimage.measure.compare_ssim(batch_X,
pred_X,
multichannel=True)
CurLoss = Training_Loss / (num + 1)
model_name = 'model-epoch' # save model
print('...Training with the %d-th banch ....' % (num + 1))
print(
'.. %d epoch training, learning rate = %.8f, Training_Loss = %.4f, PSNR = %.4f, SSIM = %.4f..'
% (j + 1, lr_, CurLoss, psnr, ssim))
showX = ML.get3band_of_tensor(batch_X,
nbanch=0,
nframe=[0, 15, 30])
maxS = np.max(showX)
minS = np.min(showX)
toshow = np.hstack(
(ML.setRange(
ML.get3band_of_tensor(Pred_YA,
nbanch=0,
nframe=[0, 15, 30]), maxS,
minS),
ML.setRange(
ML.get3band_of_tensor(
pred_ListX[FLAGS.HSInetL - 2],
nbanch=0,
nframe=[0, 15, 30]), maxS, minS),
ML.setRange(
ML.get3band_of_tensor(pred_X,
nbanch=0,
nframe=[0, 15, 30]), maxS,
minS)))
toshow2 = np.hstack(
(ML.setRange(
ML.normalized(
ML.get3band_of_tensor(batch_Y,
nbanch=0,
nframe=[2, 1, 0]))),
ML.setRange(
ML.normalized(ML.get3band_of_tensor(pred_HY))),
ML.setRange(showX, maxS, minS)))
toshow = np.vstack((toshow, toshow2))
ML.imshow(toshow)
# ML.imwrite(toshow,('tempIm_train/epoch%d_num%d.png'%(j+1,num+1)))
CurLoss = Training_Loss / (num + 1)
model_name = 'model-epoch' # save model
save_path_full = save_path + model_name
saver.save(sess, save_path_full, global_step=j + 1)
ckpt = tf.train.latest_checkpoint(save_path)
saver.restore(sess, ckpt)
Validation_Loss, pred_val = sess.run([loss, outX],
feed_dict={
X: val_h5_X,
Y: val_h5_Y,
Z: val_h5_Z,
lr: lr_
})
psnr_val = skimage.measure.compare_psnr(val_h5_X, pred_val)
ssim_val = skimage.measure.compare_ssim(val_h5_X,
pred_val,
multichannel=True)
toshow = np.hstack(
(ML.normalized(
ML.get3band_of_tensor(pred_val,
nbanch=18,
nframe=[0, 15, 30])),
ML.normalized(
ML.get3band_of_tensor(val_h5_X,
nbanch=18,
nframe=[0, 15, 30]))))
print(
'The %d epoch is finished, learning rate = %.8f, Training_Loss = %.4f, Validation_Loss = %.4f, PSNR = %.4f, SSIM = %.4f, PSNR_Valid = %.4f,SSIM_Valid = %.4f.'
% (j + 1, lr_, CurLoss, Validation_Loss, psnr, ssim, psnr_val,
ssim_val))
ML.imshow(toshow)
print('=========================================')
print('*****************************************')