def tower_loss(scope, keep_prob):
images, labels = captcha.inputs(train=True, batch_size=FLAGS.batch_size)
logits = captcha.inference(images, keep_prob)
_ = captcha.loss(logits, labels)
losses = tf.get_collection('losses', scope)
total_loss = tf.add_n(losses, name='total_loss')
return total_loss
def run_eval():
with tf.Graph().as_default(), tf.device('/cpu:0'):
images, labels = captcha.inputs(train=False,
batch_size=FLAGS.batch_size)
logits = captcha.inference(images, keep_prob=1)
eval_correct = captcha.evaluation(logits, labels)
sess = tf.Session()
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(FLAGS.checkpoint_dir))
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
num_iter = int(math.ceil(FLAGS.num_examples / FLAGS.batch_size))
true_count = 0
total_true_count = 0
total_sample_count = num_iter * FLAGS.batch_size
step = 0
log('loop: %d, total_sample_count: %d' %
(num_iter, total_sample_count))
while step < num_iter and not coord.should_stop():
true_count = sess.run(eval_correct)
total_true_count += true_count
precision = true_count / FLAGS.batch_size
log('Step %d: true/total: %d/%d precision @ 1 = %.3f' %
(step, true_count, FLAGS.batch_size, precision))
step += 1
precision = total_true_count / total_sample_count
log('true/total: %d/%d precision @ 1 = %.3f' %
(total_true_count, total_sample_count, precision))
except Exception as e:
coord.request_stop(e)
finally:
coord.request_stop()
coord.join(threads)
sess.close()
def predict(self, base64_image_str):
with tf.Graph().as_default(), tf.device('/cpu:0'):
input_images = self.input_data(base64_image_str)
images = tf.constant(input_images)
logits = captcha.inference(images, keep_prob=1)
result = captcha.output(logits)
saver = tf.train.Saver()
sess = tf.Session()
# log(checkpoint_dir)
saver.restore(sess, tf.train.latest_checkpoint(checkpoint_dir))
# log(tf.train.latest_checkpoint(checkpoint_dir))
recog_result = sess.run(result)
sess.close()
text = self.one_hot_to_texts(recog_result)
return text[0]
def run_train():
"""Train CAPTCHA for a number of steps."""
with tf.Graph().as_default():
images, labels = captcha.inputs(train=True,
batch_size=FLAGS.batch_size)
logits = captcha.inference(images, keep_prob=0.5)
loss = captcha.loss(logits, labels)
train_op = captcha.training(loss)
saver = tf.train.Saver(tf.global_variables())
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess = tf.Session()
sess.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
step = 0
while not coord.should_stop():
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
if step % 10 == 0:
logger.info('Step %d run_train: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
if step % 100 == 0:
logger.info('%s Saving in %s' %
(datetime.now(), FLAGS.checkpoint))
saver.save(sess, FLAGS.checkpoint, global_step=step)
step += 1
except Exception as e:
logger.info('%s Saving in %s' % (datetime.now(), FLAGS.checkpoint))
saver.save(sess, FLAGS.checkpoint, global_step=step)
coord.request_stop(e)
finally:
coord.request_stop()
coord.join(threads)
sess.close()
def run_predict():
with tf.Graph().as_default(), tf.device('/cpu:0'):
input_images, input_filenames = input_data(FLAGS.captcha_dir)
logger.info(input_filenames)
images = tf.constant(input_images)
logits = captcha.inference(images, keep_prob=1)
result = captcha.output(logits)
saver = tf.train.Saver()
sess = tf.Session()
saver.restore(sess, tf.train.latest_checkpoint(FLAGS.checkpoint_dir))
logger.info(tf.train.latest_checkpoint(FLAGS.checkpoint_dir))
recog_result = sess.run(result)
sess.close()
text = one_hot_to_texts(recog_result)
total_count = len(input_filenames)
true_count = 0.
for i in range(total_count):
logger.info('image ' + input_filenames[i] + " recognize ----> '" +
text[i] + "'")
if text[i] in input_filenames[i]:
true_count += 1
precision = true_count / total_count
logger.info('true/total: %d/%d recognize @ 1 = %.3f' %
(true_count, total_count, precision))