def testAll():
if not os.path.exists(param['test_dir']):
os.makedirs(param['test_dir'])
I = tf.placeholder(tf.float32, shape=(None, 512, 512, 34))
I_out, _ = PNN(I, param)
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)
with tf.Session(config=config) as sess:
ckpt = tf.train.latest_checkpoint(param['train_dir'])
saver.restore(sess, ckpt)
files = os.listdir('CAVEdata/X/')
files.sort()
for i in range(32):
I_HS, I_MS = Crd.generate_test_data2(param['ratio'], files[i])
I_in = input_prep(I_HS, I_MS, ratio=param['ratio'])
I_in = np.transpose(I_in, [0, 2, 3, 1])
I_pred = sess.run([I_out], {I: I_in})
if param['residual']:
I_res = I_in[:, :, :, :param['channel2']]
I_pred = np.squeeze(I_pred) + np.squeeze(I_res)
else:
I_pred = np.squeeze(I_pred)
sio.savemat(param['test_dir'] + files[i], {'outX': I_pred})
print(files[i] + ' done!')
def testAll():
if not os.path.exists(param['test_dir']):
os.makedirs(param['test_dir'])
I = tf.placeholder(tf.float32, shape=(None, 1676, 2632, 11))
I_out, _ = PNN(I, param)
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)
with tf.Session(config=config) as sess:
ckpt = tf.train.latest_checkpoint(param['train_dir'])
saver.restore(sess, ckpt)
data = sio.loadmat("WV2Data/testYZ.mat")
I_MS = data['Y']
I_HS = data['Z']
I_MS = np.expand_dims(I_MS, axis=0)
I_MS = np.transpose(I_MS, (0, 3, 1, 2))
I_HS = np.expand_dims(I_HS, axis=0)
I_HS = np.transpose(I_HS, (0, 3, 1, 2))
I_in = input_prep(I_HS, I_MS, ratio=param['ratio'])
I_in = np.transpose(I_in, [0, 2, 3, 1])
I_pred = sess.run([I_out], {I: I_in})
if param['residual']:
I_res = I_in[:, :, :, :param['channel2']]
I_pred = np.squeeze(I_pred) + np.squeeze(I_res)
else:
I_pred = np.squeeze(I_pred)
sio.savemat(param['test_dir'] + 'WV2_result.mat', {'outX': I_pred})
print("Done.")
def testAll():
if not os.path.exists(param['test_dir']):
os.makedirs(param['test_dir'])
I = tf.placeholder(tf.float32,
shape=(None, 448, 544,
param['channel1'] + param['channel2']))
I_out, _ = PNN(I, param)
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)
data = sio.loadmat('ChikuseiData/V')
V = np.float32(data['V'])
with tf.Session(config=config) as sess:
ckpt = tf.train.latest_checkpoint(param['train_dir'])
saver.restore(sess, ckpt)
files = os.listdir('ChikuseiData/test/')
files.sort()
for file in [i for i in files if i[0] != '.']:
I_HS, I_MS = Crd.generate_test_data(file)
I_in = input_prep(I_HS, I_MS, ratio=param['ratio'])
I_in = np.transpose(I_in, [0, 2, 3, 1])
I_pred = sess.run([I_out], {I: I_in})
if param['residual']:
I_res = I_in[:, :, :, :param['channel2']]
I_pred = np.squeeze(I_pred) + np.squeeze(I_res)
else:
I_pred = np.squeeze(I_pred)
I_pred = np.tensordot(I_pred, V.transpose(), ([2], [0]))
sio.savemat(param['test_dir'] + file, {'outX': I_pred})
print(file + ' done!')
def train():
if not os.path.exists(param['train_dir']):
os.makedirs(param['train_dir'])
random.seed(1)
img_size = param['img_size']
ch1 = param['channel1']
ch2 = param['channel2']
I = tf.placeholder(tf.float32, shape=(None, img_size, img_size, ch1 + ch2))
I_g = tf.placeholder(tf.float32, shape=(None, img_size, img_size, ch2))
I_out, reg = PNN(I, param)
lr_ = param['lr']
lr = tf.placeholder(tf.float32, shape=[])
loss1 = tf.reduce_mean(tf.abs(I_out - I_g))
if param['regol']:
loss2 = param['reg_weight'] * reg
else:
loss2 = 0
loss = loss1 + loss2
g_optim = tf.train.AdamOptimizer(lr).minimize(loss)
if param['tensorboard']:
if param['regol']:
tf.summary.scalar('loss1', loss1)
tf.summary.scalar('loss2', loss2)
tf.summary.scalar('loss', loss)
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter('logs', graph=tf.get_default_graph())
saver = tf.train.Saver(max_to_keep=5)
save_path = param['train_dir'] + param['save_model_name']
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
if tf.train.get_checkpoint_state(save_path):
ckpt = tf.train.latest_checkpoint(save_path)
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()
for j in range(start_point, param['epoch']): #start point
if j + 1 > (param['epoch'] / 3):
lr_ = param['lr'] * 0.1
if j + 1 > (2 * param['epoch'] / 3):
lr_ = param['lr'] * 0.01
for num in range(param['batch_iter']):
print(
"...Training with the %d-th batch of the %d-th epoch... " %
(num + 1, j + 1))
batch_X, batch_Y, batch_Z = Crd.train_data_in(
allX, allY, param['img_size'], param['batch_size'])
I_HS_HR = np.transpose(batch_X, (0, 3, 1, 2))
I_HS = np.transpose(batch_Z, (0, 3, 1, 2))
I_MS = np.transpose(batch_Y, (0, 3, 1, 2))
I_input = input_prep(I_HS, I_MS, ratio=param['ratio'])
I_in = np.transpose(I_input, [0, 2, 3, 1])
I_ref = np.transpose(I_HS_HR, [0, 2, 3, 1])
if param['residual']:
I_ref = I_ref - I_in[:, :, :, :param['channel2']]
if param['regol']:
_, lossvalue, lossvalue1, lossvalue2 = sess.run(
[g_optim, loss, loss1, loss2], {
I: I_in,
I_g: I_ref,
lr: lr_
})
print("loss: {0}, loss1: {1}, loss2: {2}".format(
lossvalue, lossvalue1, lossvalue2))
else:
_, lossvalue = sess.run([g_optim, loss], {
I: I_in,
I_g: I_ref,
lr: lr_
})
print("loss: {0}".format(lossvalue))
if param['tensorboard'] and num % 100 == 99:
result = sess.run(merged,
feed_dict={
I: I_in,
I_g: I_ref,
lr: lr_
})
writer.add_summary(
result, num // 100 +
param['batch_iter'] * j // 100) #将日志数据写入文件
saver.save(sess, save_path, global_step=j + 1)
ckpt = tf.train.latest_checkpoint(param['train_dir'])
saver.restore(sess, ckpt)