def build_model_m1(self):
self.xu = tf.placeholder(tf.float32, shape=[None,] + self.input_shape, name='xu_input')
###########################################################################
# network define
#
# x_encoder : x -> hx
self.config['x encoder params']['name'] = 'EncoderHX_X'
self.config['x encoder params']["output dims"] = self.hx_dim
self.x_encoder = get_encoder(self.config['x encoder'],
self.config['x encoder params'], self.is_training)
# decoder : hx -> x
self.config['hx decoder params']['name'] = 'DecoderX_HX'
# if self.config
# self.config['hx decoder params']["output dims"] = int(np.product(self.input_shape))
self.hx_decoder = get_decoder(self.config['hx decoder'], self.config['hx decoder params'], self.is_training)
###########################################################################
# for unsupervised training:
#
# xu --> mean_hxu, log_var_hxu ==> kl loss
# |
# sample_hxu --> xu_decode ==> reconstruction loss
mean_hxu, log_var_hxu = self.x_encoder(self.xu)
sample_hxu = self.draw_sample(mean_hxu, log_var_hxu)
xu_decode = self.hx_decoder(sample_hxu)
self.m1_loss_kl_z = (get_loss('kl', 'gaussian', {'mean' : mean_hxu, 'log_var' : log_var_hxu})
* self.m1_loss_weights.get('kl z loss weight', 1.0))
self.m1_loss_recon = (get_loss('reconstruction', 'mse', {'x' : self.xu, 'y' : xu_decode})
* self.m1_loss_weights.get('reconstruction loss weight', 1.0))
self.m1_loss = self.m1_loss_kl_z + self.m1_loss_recon
###########################################################################
# optimizer configure
self.m1_global_step, m1_global_step_update = get_global_step('m1_step')
(self.m1_train_op,
self.m1_learning_rate,
_) = get_optimizer_by_config(self.config['m1 optimizer'], self.config['m1 optimizer params'],
self.m1_loss, self.m1_vars, self.m1_global_step, m1_global_step_update)
def train_single(logger, tb_log_dir):
"""
:param logger:
:param tb_log_dir:
:return:
"""
# prepare dataset
if config.USE_STN:
tfrecords_dir = config.DATA_DIR + '_stn'
else:
tfrecords_dir = config.DATA_DIR
# Set up summary writer
global_summaries = set([])
summary_writer = tf.summary.FileWriter(tb_log_dir)
# Set up data provider
char_dict_path = config.CHAR_DICT
ord_map_dict_path = config.ORD_MAP_DICT
train_images, train_labels, train_images_paths, \
val_images, val_labels, val_images_paths = data.get_data(tfrecords_dir, char_dict_path, ord_map_dict_path)
# Set up convert
convert = tf_io_pipline_fast_tools.FeatureReader(
char_dict_path=char_dict_path, ord_map_dict_path=ord_map_dict_path)
# Set up network graph
train_encoder_net = get_encoder('train')
train_decoder_net = get_decoder('train')
val_encoder_net = get_encoder('val')
val_decoder_net = get_decoder('val')
gpu_list = config.GPUS
gpu_list = gpu_list.split(',')
device_name = '/gpu:{}'.format(gpu_list[0])
with tf.device(device_name):
with tf.name_scope('Train') as train_scope:
with tf.variable_scope(tf.get_variable_scope(), reuse=False):
train_encoder_out = train_encoder_net.forward(
train_images, name=config.ENCODER.NETWORKTYPE + 'Encoder')
train_decoder_net.set_label(train_labels)
train_decoder_out = train_decoder_net.predict(
train_encoder_out)
train_loss = train_decoder_net.loss(train_decoder_out)
train_sequence_dist = train_decoder_net.sequence_dist(
train_decoder_out)
with tf.name_scope('Val') as _:
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
val_encoder_out = val_encoder_net.forward(
val_images, name=config.ENCODER.NETWORKTYPE + 'Encoder')
val_decoder_net.set_label(val_labels)
val_decoder_out = val_decoder_net.predict(val_encoder_out)
val_loss = val_decoder_net.loss(val_decoder_out)
val_sequence_dist = val_decoder_net.sequence_dist(
val_decoder_out)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# set optimizer
optimizer = optimizer_builder.build(global_summaries)
optimizer = optimizer.minimize(
loss=train_loss,
global_step=tf.train.get_or_create_global_step())
# Gather initial summaries.
global_summaries.add(
tf.summary.scalar(name='train_loss', tensor=train_loss))
global_summaries.add(tf.summary.scalar(name='val_loss', tensor=val_loss))
global_summaries.add(
tf.summary.scalar(name='train_seq_distance',
tensor=train_sequence_dist))
global_summaries.add(
tf.summary.scalar(name='val_seq_distance', tensor=val_sequence_dist))
# Set saver configuration
saver = tf.train.Saver()
train_start_time = time.strftime('%Y-%m-%d-%H-%M-%S',
time.localtime(time.time()))
model_name = '{:s}_{:s}.ckpt'.format(config.MODEL.NAME,
str(train_start_time))
model_save_path = os.path.join(config.MODEL_SAVE_DIR, model_name)
# Set sess configuration
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.per_process_gpu_memory_fraction = config.TRAIN.GPU_MEMORY_FRACTION
sess_config.gpu_options.allow_growth = config.TRAIN.TF_ALLOW_GROWTH
#
sess = tf.Session(config=sess_config)
# Merge all summaries together.
summaries = set(tf.get_collection(tf.GraphKeys.SUMMARIES, train_scope))
summaries |= global_summaries
merge_summary_op = tf.summary.merge(list(summaries), name='summary_op')
summary_writer.add_graph(sess.graph)
# Set the training parameters
train_epochs = config.TRAIN.EPOCHS
with sess.as_default():
epoch = 0
tf.train.write_graph(graph_or_graph_def=sess.graph,
logdir='',
name='{:s}/single_model.pb'.format(
config.MODEL_SAVE_DIR))
if not config.RESUME_PATH:
logger.info('Training from scratch')
init = tf.global_variables_initializer()
sess.run(init)
else:
logger.info('Restore model from {:s}'.format(config.RESUME_PATH))
saver.restore(sess=sess, save_path=config.TRAIN.RESUME)
epoch = sess.run(tf.train.get_global_step())
patience_counter = 1
cost_history = [np.inf]
while epoch < train_epochs:
epoch += 1
# setup early stopping
if epoch > 1 and config.TRAIN.EARLY_STOPPING:
# We always compare to the first point where cost didn't improve
if cost_history[-1 - patience_counter] - cost_history[
-1] > config.TRAIN.PATIENCE_DELTA:
patience_counter = 1
else:
patience_counter += 1
if patience_counter > config.TRAIN.PATIENCE_EPOCHS:
logger.info(
"Cost didn't improve beyond {:f} for {:d} epochs, stopping early."
.format(config.TRAIN.PATIENCE_DELTA, patience_counter))
break
if config.TRAIN.DECODE and epoch % 500 == 0:
# train part
_, train_loss_value, train_decoder_out_dict, train_seq_dist_value, train_labels_sparse, \
merge_summary_value = sess.run(
[optimizer, train_loss, train_decoder_out, train_sequence_dist, train_labels, merge_summary_op])
avg_train_accuracy = compute_avg_accuracy(
convert, train_labels_sparse, train_decoder_out_dict)
if epoch % config.TRAIN.DISPLAY_STEP == 0:
logger.info(
'Epoch_Train: {:d} cost= {:9f} seq distance= {:9f} train accuracy= {:9f}'
.format(epoch + 1, train_loss_value,
train_seq_dist_value, avg_train_accuracy))
# validation part
val_loss_value, val_decoder_out_dict, val_seq_dist_value, val_labels_sparse = sess.run(
[val_loss, val_decoder_out, val_sequence_dist, val_labels])
avg_val_accuracy = compute_avg_accuracy(
convert, val_labels_sparse, val_decoder_out_dict)
if epoch % config.TRAIN.VAL_DISPLAY_STEP == 0:
logger.info(
'Epoch_Val: {:d} cost= {:9f} seq distance= {:9f} val accuracy= {:9f}'
.format(epoch + 1, val_loss_value, val_seq_dist_value,
avg_val_accuracy))
summary_fly = tf.Summary(value=[
tf.Summary.Value(tag='acc_train',
simple_value=avg_train_accuracy),
tf.Summary.Value(tag='acc_val',
simple_value=avg_val_accuracy),
])
summary_writer.add_summary(summary=summary_fly,
global_step=epoch)
else:
_, train_loss_value, merge_summary_value = sess.run(
[optimizer, train_loss, merge_summary_op])
if epoch % config.TRAIN.DISPLAY_STEP == 0:
logger.info('Epoch_Train: {:d} cost= {:9f}'.format(
epoch + 1, train_loss_value))
# record history train ctc loss
cost_history.append(train_loss_value)
# add training sumary
summary_writer.add_summary(summary=merge_summary_value,
global_step=epoch)
if epoch % 2000 == 0:
saver.save(sess=sess,
save_path=model_save_path,
global_step=epoch)
return np.array(cost_history[1:]) # Don't return the first np.inf
def train_multi(logger, tb_log_dir):
"""
:param logger:
:param tb_log_dir:
:return:
"""
with tf.Graph().as_default(), tf.device('/cpu:0'):
# prepare dataset
if config.USE_STN:
tfrecords_dir = config.DATA_DIR + '_stn'
else:
tfrecords_dir = config.DATA_DIR
# Set up summary writer
global_summaries = set([])
summary_writer = tf.summary.FileWriter(tb_log_dir)
# Set up data provider
char_dict_path = config.CHAR_DICT
ord_map_dict_path = config.ORD_MAP_DICT
train_images, train_labels, train_images_paths, \
val_images, val_labels, val_images_paths = data.get_data(tfrecords_dir, char_dict_path, ord_map_dict_path)
# Set up network graph
train_encoder_net = get_encoder('train')
train_decoder_net = get_decoder('train')
val_encoder_net = get_encoder('val')
val_decoder_net = get_decoder('val')
# set average container
tower_grads = []
train_scopes = []
train_tower_loss = []
val_tower_loss = []
batchnorm_updates = None
# Set up optimizer
optimizer = optimizer_builder.build(global_summaries)
gpu_list = config.GPUS
gpu_list = gpu_list.split(',')
# set distributed train op
is_network_initialized = False
for i, gpu_id in enumerate(gpu_list):
with tf.device('/gpu:{}'.format(gpu_id)):
with tf.name_scope('Train_{:d}'.format(i)) as train_scope:
with tf.variable_scope(tf.get_variable_scope(),
reuse=is_network_initialized):
train_loss, grads = compute_net_gradients(
train_images, train_labels, train_encoder_net,
train_decoder_net, optimizer)
is_network_initialized = True
train_scopes.append(train_scope)
# Only use the mean and var in the first gpu tower to update the parameter
if i == 0:
batchnorm_updates = tf.get_collection(
tf.GraphKeys.UPDATE_OPS)
tower_grads.append(grads)
train_tower_loss.append(train_loss)
with tf.name_scope('Val_{:d}'.format(i)) as _:
with tf.variable_scope(tf.get_variable_scope(),
reuse=is_network_initialized):
val_loss, _ = compute_net_gradients(
val_images, val_labels, val_encoder_net,
val_decoder_net)
val_tower_loss.append(val_loss)
with tf.name_scope('Average_Grad'):
grads = average_gradients(tower_grads)
with tf.name_scope('Average_Loss'):
avg_train_loss = tf.reduce_mean(train_tower_loss)
avg_val_loss = tf.reduce_mean(val_tower_loss)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
global_summaries.add(tf.summary.histogram(var.op.name, var))
# Track the moving averages of all trainable variables
variable_averages = tf.train.ExponentialMovingAverage(
config.TRAIN.MOVING_AVERAGE_DECAY,
num_updates=tf.train.get_or_create_global_step())
variables_to_average = tf.trainable_variables(
) + tf.moving_average_variables()
variables_averages_op = variable_averages.apply(variables_to_average)
# Group all the op needed for training
batchnorm_updates_op = tf.group(*batchnorm_updates)
apply_gradient_op = optimizer.apply_gradients(
grads, global_step=tf.train.get_or_create_global_step())
train_op = tf.group(apply_gradient_op, variables_averages_op,
batchnorm_updates_op)
global_summaries.add(
tf.summary.scalar(name='average_train_loss',
tensor=avg_train_loss))
# Merge all summaries together.
summaries = set(
tf.get_collection(tf.GraphKeys.SUMMARIES, train_scopes[0]))
summaries |= global_summaries
train_merge_summary_op = tf.summary.merge(list(summaries),
name='train_summary_op')
val_merge_summary_op = tf.summary.merge(
[tf.summary.scalar(name='average_val_loss', tensor=avg_val_loss)],
name='val_summary_op')
# Set saver configuration
saver = tf.train.Saver()
train_start_time = time.strftime('%Y-%m-%d-%H-%M-%S',
time.localtime(time.time()))
model_name = '{:s}_{:s}.ckpt'.format(config.MODEL.NAME,
str(train_start_time))
model_save_path = os.path.join(config.MODEL_SAVE_DIR, model_name)
# set sess config
sess_config = tf.ConfigProto(device_count={'GPU': len(gpu_list)},
allow_soft_placement=True)
sess_config.gpu_options.per_process_gpu_memory_fraction = config.TRAIN.GPU_MEMORY_FRACTION
sess_config.gpu_options.allow_growth = config.TRAIN.TF_ALLOW_GROWTH
sess_config.gpu_options.allocator_type = 'BFC'
# Set the training parameters
train_epochs = config.TRAIN.EPOCHS
logger.info('Global configuration is as follows:')
logger.info(config)
sess = tf.Session(config=sess_config)
summary_writer.add_graph(sess.graph)
epoch = 0
tf.train.write_graph(graph_or_graph_def=sess.graph,
logdir='',
name='{:s}/multi_model.pb'.format(
config.MODEL_SAVE_DIR))
if not config.RESUME_PATH:
logger.info('Training from scratch')
init = tf.global_variables_initializer()
sess.run(init)
else:
logger.info('Restore model from {:s}'.format(config.RESUME_PATH))
saver.restore(sess=sess, save_path=config.TRAIN.RESUME)
epoch = sess.run(tf.train.get_global_step())
train_cost_time_mean = []
val_cost_time_mean = []
while epoch < train_epochs:
epoch += 1
# training part
t_start = time.time()
_, train_loss_value, train_summary = sess.run(
fetches=[train_op, avg_train_loss, train_merge_summary_op])
if math.isnan(train_loss_value):
raise ValueError('Train loss is nan')
cost_time = time.time() - t_start
train_cost_time_mean.append(cost_time)
summary_writer.add_summary(summary=train_summary,
global_step=epoch)
# validation part
t_start_val = time.time()
val_loss_value, val_summary = sess.run(
fetches=[avg_val_loss, val_merge_summary_op])
cost_time_val = time.time() - t_start_val
val_cost_time_mean.append(cost_time_val)
summary_writer.add_summary(val_summary, global_step=epoch)
if epoch % config.TRAIN.DISPLAY_STEP == 0:
logger.info(
'Epoch_Train: {:d} total_loss= {:6f} mean_cost_time= {:5f}s '
.format(epoch + 1, train_loss_value,
np.mean(train_cost_time_mean)))
train_cost_time_mean.clear()
if epoch % config.TRAIN.VAL_DISPLAY_STEP == 0:
logger.info(
'Epoch_Val: {:d} total_loss= {:6f} mean_cost_time= {:5f}s '
.format(epoch + 1, val_loss_value,
np.mean(val_cost_time_mean)))
val_cost_time_mean.clear()
if epoch % 5000 == 0:
saver.save(sess=sess,
save_path=model_save_path,
global_step=epoch)
sess.close()
return
def build_model(self):
self.xu = tf.placeholder(tf.float32,
shape=[
None,
] + self.input_shape,
name='xu_input')
self.xl = tf.placeholder(tf.float32,
shape=[
None,
] + self.input_shape,
name='xl_input')
self.yl = tf.placeholder(tf.float32,
shape=[None, self.nb_classes],
name='yl_input')
###########################################################################
# network define
#
# x_encoder : x -> hx
self.config['x encoder params']['name'] = 'EncoderHX_X'
self.config['x encoder params']["output dims"] = self.hx_dim
self.x_encoder = get_encoder(self.config['x encoder'],
self.config['x encoder params'],
self.is_training)
# hx_y_encoder : [hx, y] -> hz
self.config['hx y encoder params']['name'] = 'EncoderHZ_HXY'
self.config['hx y encoder params']["output dims"] = self.hz_dim
self.hx_y_encoder = get_encoder(self.config['hx y encoder'],
self.config['hx y encoder params'],
self.is_training)
# hz_y_decoder : [hz, y] -> x_decode
self.config['hz y decoder params']['name'] = 'DecoderX_HZY'
self.config['hz y decoder params']["output dims"] = int(
np.product(self.input_shape))
self.hz_y_decoder = get_decoder(self.config['hz y decoder'],
self.config['hz y decoder params'],
self.is_training)
# x_classifier : hx -> ylogits
self.config['x classifier params']['name'] = 'ClassifierX'
self.config['x classifier params']["output dims"] = self.nb_classes
self.x_classifier = get_classifier(self.config['x classifier'],
self.config['x classifier params'],
self.is_training)
###########################################################################
# for supervised training:
#
# xl --> mean_hxl, log_var_hxl
# |
# sample_hxl --> yllogits ==> classification loss
# |
# [sample_hxl, yl] --> mean_hzl, log_var_hzl ==> kl loss
# | |
# [yl, sample_hzl] --> xl_decode ==> reconstruction loss
#
hxl = self.x_encoder(self.xl)
mean_hzl, log_var_hzl = self.hx_y_encoder(
tf.concat([hxl, self.yl], axis=1))
sample_hzl = self.draw_sample(mean_hzl, log_var_hzl)
decode_xl = self.hz_y_decoder(tf.concat([sample_hzl, self.yl], axis=1))
# decode_xl = self.hx_decoder(decode_hxl)
yllogits = self.x_classifier(self.xl)
self.su_loss_kl_z = (get_loss('kl', 'gaussian', {
'mean': mean_hzl,
'log_var': log_var_hzl,
}) * self.loss_weights.get('kl z loss weight', 1.0))
self.su_loss_recon = (get_loss('reconstruction', 'mse', {
'x': self.xl,
'y': decode_xl
}) * self.loss_weights.get('reconstruction loss weight', 1.0))
self.su_loss_cls = (get_loss('classification', 'cross entropy', {
'logits': yllogits,
'labels': self.yl
}) * self.loss_weights.get('classiciation loss weight', 1.0))
self.su_loss_reg = (
get_loss('regularization', 'l2',
{'var_list': self.x_classifier.vars}) *
self.loss_weights.get('regularization loss weight', 0.0001))
self.su_loss = ((self.su_loss_kl_z + self.su_loss_recon +
self.su_loss_cls + self.su_loss_reg) *
self.loss_weights.get('supervised loss weight', 1.0))
###########################################################################
# for unsupervised training:
#
# xu --> mean_hxu, log_var_hxu
# |
# sample_hxu --> yulogits --> yuprobs
# |
# [sample_hxu, y0] --> mean_hzu0, log_var_hzu0 ==> kl_loss * yuprobs[0]
# | | |
# | [y0, sample_hzu0] --> decode_hxu0 ==> reconstruction loss * yuprobs[0]
# |
# [sample_hxu, y1] --> mean_hzu1, log_var_hzu1 ==> kl_loss * yuprobs[1]
# | | |
# | [y1, sample_hzu1] --> decode_hxu1 ==> reconstruction loss * yuprobs[1]
# .......
#
hxu = self.x_encoder(self.xu)
yulogits = self.x_classifier(self.xu)
yuprobs = tf.nn.softmax(yulogits)
unsu_loss_kl_z_list = []
unsu_loss_recon_list = []
for i in range(self.nb_classes):
yu_fake = tf.ones([
tf.shape(self.xu)[0],
], dtype=tf.int32) * i
yu_fake = tf.one_hot(yu_fake, depth=self.nb_classes)
mean_hzu, log_var_hzu = self.hx_y_encoder(
tf.concat([hxu, yu_fake], axis=1))
sample_hzu = self.draw_sample(mean_hzu, log_var_hzu)
decode_xu = self.hz_y_decoder(
tf.concat([sample_hzu, yu_fake], axis=1))
# decode_xu = self.hx_decoder(decode_hxu)
unsu_loss_kl_z_list.append(
get_loss(
'kl', 'gaussian', {
'mean': mean_hzu,
'log_var': log_var_hzu,
'instance_weight': yuprobs[:, i]
}))
unsu_loss_recon_list.append(
get_loss('reconstruction', 'mse', {
'x': self.xu,
'y': decode_xu,
'instance_weight': yuprobs[:, i]
}))
self.unsu_loss_kl_y = (
get_loss('kl', 'bernoulli', {'probs': yuprobs}) *
self.loss_weights.get('kl y loss weight', 1.0))
self.unsu_loss_kl_z = (tf.reduce_sum(unsu_loss_kl_z_list) *
self.loss_weights.get('kl z loss weight', 1.0))
self.unsu_loss_recon = (
tf.reduce_sum(unsu_loss_recon_list) *
self.loss_weights.get('reconstruction loss weight', 1.0))
self.unsu_loss_reg = (
get_loss('regularization', 'l2',
{'var_list': self.x_classifier.vars}) *
self.loss_weights.get('regularization loss weight', 0.0001))
self.unsu_loss = (
(self.unsu_loss_kl_z + self.unsu_loss_recon + self.unsu_loss_kl_y +
self.unsu_loss_reg) *
self.loss_weights.get('unsupervised loss weight', 1.0))
self.xt = tf.placeholder(tf.float32,
shape=[
None,
] + self.input_shape,
name='xt_input')
###########################################################################
# for test models
#
# xt --> mean_hxt, log_var_hxt
# |
# sample_hxt --> ytlogits --> ytprobs
# | |
# [sample_hxt, ytprobs] --> mean_hzt, log_var_hzt
#
hxt = self.x_encoder(self.xt)
ytlogits = self.x_classifier(self.xt)
self.ytprobs = tf.nn.softmax(ytlogits)
self.mean_hzt, self.log_var_hzt = self.hx_y_encoder(
tf.concat([hxt, self.ytprobs], axis=1))
###########################################################################
# optimizer configure
global_step, global_step_update = get_global_step()
(self.supervised_train_op, self.supervised_learning_rate,
_) = get_optimizer_by_config(self.config['optimizer'],
self.config['optimizer params'],
self.su_loss, self.vars, global_step,
global_step_update)
(self.unsupervised_train_op, self.unsupervised_learning_rate,
_) = get_optimizer_by_config(self.config['optimizer'],
self.config['optimizer parmas'],
self.unsu_loss, self.vars, global_step,
global_step_update)
###########################################################################
# model saver
self.saver = tf.train.Saver(self.vars + [
self.global_step,
])
def build_model(self):
self.x_real = tf.placeholder(
tf.float32,
shape=[None, np.product(self.input_shape)],
name='x_input')
self.y_real = tf.placeholder(tf.float32,
shape=[None, self.nb_classes],
name='y_input')
# self.encoder_input_shape = int(np.product(self.input_shape))
self.config['encoder parmas']['name'] = 'EncoderX'
self.config['encoder params']["output dims"] = self.z_dim
self.encoder = get_encoder(self.config['x encoder'],
self.config['encoder params'],
self.is_training)
self.config['decoder params']['name'] = 'Decoder'
self.config['decoder params']["output dims"] = self.encoder_input_shape
# self.y_encoder = get_encoder(self.config['y encoder'], self.config['y encoder params'], self.is_training)
self.decoder = get_decoder(self.config['decoder'],
self.config['decoder params'],
self.is_training)
# build encoder
self.z_mean, self.z_log_var = self.x_encoder(
tf.concatenate([self.x_real, self.y_real]))
self.z_mean_y = self.y_encoder(self.y_real)
# sample z from z_mean and z_log_var
self.z_sample = self.draw_sample(self.z_mean, self.z_log_var)
# build decoder
self.x_decode = self.decoder(self.z_sample)
# build test decoder
self.z_test = tf.placeholder(tf.float32,
shape=[None, self.z_dim],
name='z_test')
self.x_test = self.decoder(self.z_test, reuse=True)
# loss function
self.kl_loss = (get_loss(
'kl', self.config['kl loss'], {
'z_mean': (self.z_mean - self.z_mean_y),
'z_log_var': self.z_log_var
}) * self.config.get('kl loss prod', 1.0))
self.xent_loss = (
get_loss('reconstruction', self.config['reconstruction loss'], {
'x': self.x_real,
'y': self.x_decode
}) * self.config.get('reconstruction loss prod', 1.0))
self.loss = self.kl_loss + self.xent_loss
# optimizer configure
self.global_step, self.global_step_update = get_global_step()
if 'lr' in self.config:
self.learning_rate = get_learning_rate(self.config['lr_scheme'],
float(self.config['lr']),
self.global_step,
self.config['lr_params'])
self.optimizer = get_optimizer(
self.config['optimizer'],
{'learning_rate': self.learning_rate}, self.loss,
self.decoder.vars + self.x_encoder.vars + self.y_encoder.vars)
else:
self.optimizer = get_optimizer(
self.config['optimizer'], {}, self.loss,
self.decoder.vars + self.x_encoder.vars + self.y_encoder.vars)
self.train_update = tf.group([self.optimizer, self.global_step_update])
# model saver
self.saver = tf.train.Saver(self.x_encoder.vars + self.y_encoder.vars,
self.decoder.vars + [
self.global_step,
])
def _build_encoder(self, name, params=None):
net_config = self.config[name + ' params'].copy()
if params is not None:
net_config.update(params)
return get_encoder(self.config[name], net_config, self.is_training)
def recognize(image_path, weights_path, is_vis):
"""
:param image_path:
:param weights_path:
:param is_vis:
:return:
"""
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if config.USE_STN:
new_heigth = config.STN.IH
new_width = config.STN.IW
image = cv2.resize(image,
dsize=(config.STN.IW, config.STN.IH),
interpolation=cv2.INTER_LINEAR)
else:
new_heigth = config.ENCODER.IH
new_width = config.ENCODER.IW
image = cv2.resize(image,
dsize=(config.ENCODER.IW, config.ENCODER.IH),
interpolation=cv2.INTER_LINEAR)
image_vis = image
# Set up data placeholder
char_dict_path = config.CHAR_DICT
ord_map_dict_path = config.ORD_MAP_DICT
inputdata = tf.placeholder(
dtype=tf.float32,
shape=[1, new_heigth, new_width, config.ENCODER.INPUT_CHANNELS],
name='input')
# Set up convert
convert = tf_io_pipline_fast_tools.FeatureReader(
char_dict_path=char_dict_path, ord_map_dict_path=ord_map_dict_path)
# Set up network graph
encoder_net = get_encoder('test')
decoder_net = get_decoder('test')
gpu_list = config.GPUS
gpu_list = gpu_list.split(',')
device_name = '/gpu:{}'.format(gpu_list[0])
with tf.device(device_name):
with tf.name_scope('Test'):
encoder_out = encoder_net.forward(inputdata,
name=config.ENCODER.NETWORKTYPE +
'Encoder')
decoder_out = decoder_net.predict(encoder_out)
# Config tf saver
saver = tf.train.Saver()
# config tf session
sess_config = tf.ConfigProto(
allow_soft_placement=True) # 允许tf自动选择一个存在并且可用的设备来运行操作
sess_config.gpu_options.per_process_gpu_memory_fraction = config.TEST.GPU_MEMORY_FRACTION
sess_config.gpu_options.allow_growth = config.TEST.TF_ALLOW_GROWTH
sess = tf.Session(config=sess_config)
with sess.as_default():
saver.restore(sess=sess, save_path=weights_path)
decoder_out_value = sess.run(decoder_out,
feed_dict={inputdata: [image]})
if config.DECODER_MODEL in config.VALID_DECODER_MODEL:
if config.DECODER_MODEL == 'normal_ctc' or config.DECODER_MODEL == 'normal_sts':
preds = convert.sparse_tensor_to_str(
decoder_out_value['labels'])
elif config.DECODER_MODEL == 'reverse_sts':
preds = convert.sparse_tensor_to_str(
decoder_out_value['labels'], reverse=True)
elif config.DECODER_MODEL == 'bidirection_sts':
normal_preds = convert.sparse_tensor_to_str(
decoder_out_value['normal']['labels'])
reverse_preds = convert.sparse_tensor_to_str(
decoder_out_value['reverse']['labels'], reverse=True)
if decoder_out_value['normal']['scores'] > decoder_out_value[
'reverse']['scores']:
preds = normal_preds
else:
preds = reverse_preds
else:
raise ValueError('Unknown decoder model: {}'.format(
config.DECODER_MODEL))
else:
raise ValueError('Unknown decoder model: {}'.format(
config.DECODER_MODEL))
logger.info('Predict image {:s} result {:s}'.format(
ops.split(image_path)[1], preds[0]))
if is_vis:
plt.figure('CRNN Model Demo')
plt.imshow(image_vis)
plt.show()
sess.close()
return