def get_result_from_log():
# ## data 2------------------------
dataname = Config.paviaU
class_num = Config.paviaU_class_num
# ## data 2------------------------
# ## data 2------------------------
# dataname = Config.ksc
# class_num =Config.ksc_class_num
# ## data 2------------------------
# log
log = Config.spectral_log + dataname + '/'
with tf.Graph().as_default():
pd = Data(data_name=dataname)
data_sets = pd.get_train_valid_test_of_spectral_feature()
# init class
train_data = Data_Set([data_sets[0], data_sets[1]])
valid_data = Data_Set([data_sets[2], data_sets[3]])
test_data = Data_Set([data_sets[4], data_sets[5]])
images_placeholder, labels_placeholder = \
placeholder_inputs(input_dim=data_sets[0].shape[1])
logits = sAE.inference(images_placeholder, class_num,
Config.encoder_layers)
loss = sAE.loss(logits, labels_placeholder)
train_op = sAE.training(loss, Config.learn_rate)
eval_correct = sAE.evaluation(logits, labels_placeholder)
model_file = tf.train.latest_checkpoint('./spectral_log/')
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_file)
for step in range(Config.max_steps):
if (step + 1) % 100 == 0 or (step + 1) == Config.max_steps:
# checkpoint_file = os.path.join(log, 'valid_model.ckpt')
# saver.save(sess, checkpoint_file, global_step=step)
print(' Valid data evaluation')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, valid_data)
if (step + 1) % 100 == 0 or (step + 1) == Config.max_steps:
# checkpoint_file = os.path.join(log, 'test_model.ckpt')
# tf.saver.save(sess, checkpoint_file, global_step=step)
print(' Test data evaluation')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, test_data)
def run_trainning():
# # ## data 1------------------------
#dataname = Config.paviaU
# class_num =Config.paviaU_class_num
# # ## data 1------------------------
## data 2------------------------
dataname = Config.ksc
#class_num =Config.ksc_class_num
## data 2------------------------
## data 2------------------------
dataname = Config.Salinas
# class_num =Config.ksc_class_num
## data 2------------------------
# log
spatial_log='/'+dataname+'/'+'spatial'
final_ckpt_dir=Config.final_ckpt_dir+spatial_log
if not os.path.exists(final_ckpt_dir):
os.makedirs(final_ckpt_dir)
pre_train_ckpt_dir = Config.pretrain_ckpt+spatial_log
if not os.path.exists(pre_train_ckpt_dir):
os.makedirs(pre_train_ckpt_dir)
pre_need_train = True
final_need_train = True
with tf.Graph().as_default() as gad:
pd = Data(data_name = dataname)
data_sets = pd.get_train_valid_test_of_spatial_feature()
# init class
train_data = Data_Set([data_sets[0],data_sets[1]])
valid_data = Data_Set([data_sets[2],data_sets[3]])
test_data = Data_Set([data_sets[4],data_sets[5]])
with tf.Session(graph = gad) as sess:
input_dim = [train_data.feature_dim,Config.batch_size]
epoch_size = data_sets[0].shape[0]
# int
sae = new_SAE(input_dim, encoder_shape=Config.encoder_layers)
pre_trian_saver = tf.train.Saver() # after graph, sometime with variable name ,so so
fianl_saver = tf.train.Saver() # after graph, sometime with variable name ,so so
writer = tf.summary.FileWriter(final_ckpt_dir,gad)
init = tf.global_variables_initializer() # order is important,after graph
sess.run(init) #after graph
# if pretrain ready
if pre_need_train == True:
path = os.path.join(pre_train_ckpt_dir,'SAE_graph')
sae.pre_train(train_data,sess,path)
pre_trian_saver.save(sess, os.path.join(pre_train_ckpt_dir, 'pre_train_model.ckpt'))
else:
ckpt = tf.train.get_checkpoint_state(pre_train_ckpt_dir)
pre_trian_saver.restore(sess, ckpt.model_checkpoint_path)
# train real model
if final_need_train == True:
for step in range(int(Config.epoch_final_train_times * epoch_size / Config.batch_size)):
start_time = time.time()
feed_dict = fill_feed_dict(train_data, sae.x, sae.y)
cost,_= sae.train_final_model(feed_dict,sess)
duration = time.time() - start_time
if step % 500 == 0:
print('final model train : Step %d: loss = %.2f(%.3f sec)' % (step, cost, duration))
summary_ = sess.run(sae.final_merged, feed_dict=feed_dict)
writer.add_summary(summary_,step)
fianl_saver.save(sess, os.path.join(final_ckpt_dir, 'final_model.ckpt'))
print(' Train data evaluation')
do_eval(sess,
sae.correct,
sae.x,
sae.y,
train_data)
if (step + 1) % 500 == 0 or (step + 1) == Config.max_steps:
print(' Valid data evaluation')
do_eval(sess,
sae.correct,
sae.x,
sae.y,
valid_data)
if (step + 1) % 500 == 0 or (step + 1) == Config.max_steps:\
print(' Test data evaluation')
do_eval(sess,
sae.correct,
sae.x,
sae.y,
test_data)
else:
ckpt = tf.train.get_checkpoint_state(final_ckpt_dir)
fianl_saver.restore(sess, ckpt.model_checkpoint_path)
# v = tf.get_collection(tf.GraphKeys.VARIABLES)
print(' Train data evaluation')
do_eval(sess,
sae.correct,
sae.x,
sae.y,
train_data)
print(' Valid data evaluation')
do_eval(sess,
sae.correct,
sae.x,
sae.y,
valid_data)
print(' Test data evaluation')
do_eval(sess,
sae.correct,
sae.x,
sae.y,
test_data)
# ## data 1------------------------
dataname = Config.paviaU
class_num = Config.paviaU_class_num
# ## data 1------------------------
# set log dir
log = './test/'
with tf.Graph().as_default() as gr:
pd = Data(dataname)
data_sets = pd.get_train_valid_test_of_spectral_feature()
# init data
train_data = Data_Set([data_sets[0], data_sets[1]])
valid_data = Data_Set([data_sets[2], data_sets[3]])
test_data = Data_Set([data_sets[4], data_sets[5]])
with tf.Session(graph=gr) as sess:
img_placeholder = tf.placeholder(
tf.float32,
[None, data_sets[0].shape[1]]) ## important to save real data
sae = new_SAE(img_placeholder,
encoder_shape=[60, 60, 60, 60],
sess=sess)
saver = tf.train.Saver(
) # after graph, sometime with variable name ,so so
init = tf.global_variables_initializer(
) # order is important,after graph
def run_trainning():
# ## data 1------------------------
# dataname = Config.ksc
# class_num =Config.ksc_class_num
# ## data 1------------------------
## data 2------------------------
dataname = Config.ksc
class_num = Config.ksc_class_num
## data 2------------------------
#set log dir
log = Config.mix_feature_log+dataname+'/'
with tf.Graph().as_default():
pd = Data(dataname)
data_sets = pd.get_train_valid_test_of_mix_feature()
# init class
train_data = Data_Set([data_sets[0],data_sets[1]])
valid_data = Data_Set([data_sets[2],data_sets[3]])
test_data = Data_Set([data_sets[4],data_sets[5]])
images_placeholder, labels_placeholder = \
placeholder_inputs(input_dim=data_sets[0].shape[1])
logits = sAE.inference(images_placeholder,class_num,Config.encoder_layers)
loss = sAE.loss(logits,labels_placeholder)
train_op = sAE.training(loss,Config.learn_rate)
eval_correct = sAE.evaluation(logits,labels_placeholder)
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
summary_writer = tf.summary.FileWriter(log,sess.graph)
sess.run(init)
for step in range(Config.max_steps):
start_time = time.time()
feed_dict = fill_feed_dict(train_data,images_placeholder,labels_placeholder)
# only return loss inside run([])
_,loss_value = sess.run([train_op,loss],feed_dict=feed_dict)
duration = time.time()-start_time
if step%100 == 0:
print('Step %d: loss = %.2f(%.3f sec)'%(step,loss_value,duration))
#print(sess.run(sAE.evaluation(logits, train_data[1])))
summary_str = sess.run(summary,feed_dict=feed_dict)
summary_writer.add_summary(summary_str,step)
summary_writer.flush()
if(step+1)%100 == 0 or (step+1)==Config.max_steps:
checkpoint_file =os.path.join(log,'valid_model.ckpt')
saver.save(sess,checkpoint_file,global_step = step)
print(' Valid data evaluation')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
valid_data)
if (step + 1) % 100 == 0 or (step + 1) == Config.max_steps:
checkpoint_file = os.path.join(log, 'test_model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
print(' Test data evaluation')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
test_data)
def get_results_from_exist_model():
# ## data 2------------------------
dataname = Config.paviaU
class_num = Config.paviaU_class_num
# ## data 2------------------------
# ## data 2------------------------
# dataname = Config.ksc
# class_num =Config.ksc_class_num
# ## data 2------------------------
#set log dir
log = './'+dataname+'/'+Config.spatial_log
with tf.Graph().as_default():
pd = Data(data_name=dataname)
data_sets = pd.get_train_valid_test_of_spatial_feature()
# init class
train_data = Data_Set([data_sets[0], data_sets[1]])
valid_data = Data_Set([data_sets[2], data_sets[3]])
test_data = Data_Set([data_sets[4], data_sets[5]])
images_placeholder, labels_placeholder = \
placeholder_inputs(input_dim=data_sets[0].shape[1])
logits = sAE.inference(images_placeholder, class_num, Config.encoder_layers)
loss = sAE.loss(logits, labels_placeholder)
train_op = sAE.training(loss, Config.learn_rate)
eval_correct = sAE.evaluation(logits, labels_placeholder)
saver = tf.train.Saver()
# valid_saver = tf.train.Saver()
# test_saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
path = log+'/test/test_model.ckpt-99'
model_file = tf.train.latest_checkpoint(log+'/test')
saver.restore(sess, model_file)
print('success')
for step in range(Config.max_steps):
feed_dict = fill_feed_dict(train_data,images_placeholder,labels_placeholder)
# only return loss inside run([])
_,loss_value = sess.run([train_op,loss],feed_dict=feed_dict)
if (step + 1) % 100 == 0 or (step + 1) == Config.max_steps:
# checkpoint_file = os.path.join(log, 'valid_model.ckpt')
# valid_saver.restore(sess, checkpoint_file)
# #saver.save(sess, checkpoint_file, global_step=step)
print(' Valid data evaluation')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
valid_data)
if (step + 1) % 100 == 0 or (step + 1) == Config.max_steps:
# checkpoint_file1 = os.path.join(log, 'test_model.ckpt')
# test_saver.restore(sess, checkpoint_file1)
# # tf.saver.save(sess, checkpoint_file, global_step=step)
print(' Test data evaluation')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
test_data)