#
model = ModelRecog()
#
# data
print('loading data ...')
data_train = model_data.load_data(meta.dir_data_train)
data_valid = model_data.load_data(meta.dir_data_valid)
print('load finished.')
#
# train
model.train_and_valid(data_train, data_valid)
#
#
# predict
model.load_pb_for_prediction()
sess = model.create_session_for_prediction()
#
list_images_valid = model_data.getFilesInDirect(meta.dir_images_valid,
meta.str_dot_img_ext)
for img_file in list_images_valid:
#
# img_file = './data_test/images/bkgd_1_0_generated_0.png'
#
print(img_file)
#
model.predict(sess, img_file, out_dir='./results_prediction')
#
# input-output, graph
x = tf.placeholder(tf.float32, (None, None, None, 3), name='x-input')
yT = tf.sparse_placeholder(tf.int32, shape=(None, None), name='y-input')
w = tf.placeholder(tf.int32, (None, ), name='w-input')
#
features, sequence_length = model_recog.conv_feat_layers(
x, w, learn.ModeKeys.TRAIN) #INFER
result_logits = model_recog.rnn_recog_layers(features, sequence_length,
num_classes)
#
loss = model_recog.ctc_loss_layer(yT, result_logits, sequence_length)
#
print('graph loaded')
#
# get test images
list_images = model_recog_data.getFilesInDirect(dir_images, str_dot_img_ext)
#
# test_result save-path
if os.path.exists(dir_results): shutil.rmtree(dir_results)
time.sleep(0.1)
os.mkdir(dir_results)
#
# to process
saver = tf.train.Saver()
with tf.Session() as sess:
#
tf.global_variables_initializer().run()
#
# restore with saved data
ckpt = tf.train.get_checkpoint_state(model_dir)
#
def train_and_valid(self,
data_train,
data_valid,
load_from_pretrained=True):
#
# model save-path
if not os.path.exists(meta.model_recog_dir):
os.mkdir(meta.model_recog_dir)
#
# training graph
self.z_graph = tf.Graph()
self.z_define_graph_all(self.z_graph, True)
#
# load from pretained
list_ckpt = model_recog_data.getFilesInDirect(meta.model_recog_dir,
'.meta')
#
print(' ')
#
if len(list_ckpt) > 0:
print(
'model_recog ckpt already exists, no need to load common tensors.'
)
elif load_from_pretrained == False:
print('not to load common tensors, by manual setting.')
else:
print('load common tensors from pretrained detection model.')
self.z_load_from_pretrained_detection_model()
print(' ')
#
# restore and train
with self.z_graph.as_default():
#
saver = tf.train.Saver()
with tf.Session(config=self.z_sess_config) as sess:
#
tf.global_variables_initializer().run()
#
# restore with saved data
ckpt = tf.train.get_checkpoint_state(meta.model_recog_dir)
#
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
#
# variables
#
x = self.z_graph.get_tensor_by_name('x-input:0')
w = self.z_graph.get_tensor_by_name('w-input:0')
#
y_s = self.z_graph.get_tensor_by_name('y-input/shape:0')
y_i = self.z_graph.get_tensor_by_name('y-input/indices:0')
y_v = self.z_graph.get_tensor_by_name('y-input/values:0')
#
# <tf.Operation 'y-input/shape' type=Placeholder>,
# <tf.Operation 'y-input/values' type=Placeholder>,
# <tf.Operation 'y-input/indices' type=Placeholder>]
#
#conv_feat = self.z_graph.get_tensor_by_name('conv_comm/conv_feat:0')
#
loss = self.z_graph.get_tensor_by_name('loss:0')
#
global_step = self.z_graph.get_tensor_by_name('global_step:0')
learning_rate = self.z_graph.get_tensor_by_name(
'learning_rate:0')
train_op = self.z_graph.get_tensor_by_name(
'train_op/control_dependency:0')
#
print('begin to train ...')
#
num_samples = len(data_train['x'])
#
# start training
start_time = time.time()
begin_time = start_time
step = 0
#
for curr_iter in range(TRAINING_STEPS):
#
# save and validate
if step % self.z_valid_freq == 0:
#
# ckpt
print('save model to ckpt ...')
saver.save(sess, os.path.join(meta.model_recog_dir, meta.model_recog_name), \
global_step = step)
#
# validate
print('validating ...')
self.validate(data_valid, step)
#
#
# train
index_batch = random.sample(range(num_samples),
self.z_batch_size)
#
images = [data_train['x'][i] for i in index_batch]
targets = [data_train['y'][i] for i in index_batch]
w_arr = np.ones((self.z_batch_size, ),
dtype=np.float32) * meta.width_norm
#
# targets_sparse_value
tsv = model_def.convert2SparseTensorValue(targets)
#
#
feed_dict = {
x: images,
w: w_arr,
y_s: tsv.dense_shape,
y_i: tsv.indices,
y_v: tsv.values
}
#
#print(width)
#conv_v = sess.run(conv_feat, feed_dict)
#print(len(conv_v))
# sess.run
_, loss_value, step, lr = sess.run([train_op, loss, global_step, learning_rate], \
feed_dict)
#
if step % 1 == 0:
#
curr_time = time.time()
#
print(
'step: %d, loss: %g, lr: %g, sect_time: %.1f, total_time: %.1f'
% (step, loss_value, lr, curr_time - begin_time,
curr_time - start_time))
#
begin_time = curr_time