def gen_valid_input(self, inputs, decode):
#---------------------- valid
validset = list_files(FLAGS.valid_input)
logging.info('validset:{} {}'.format(len(validset), validset[:2]))
eval_image_name, eval_text, eval_text_str, eval_input_text, eval_input_text_str = inputs(
validset,
decode=decode,
batch_size=FLAGS.eval_batch_size,
num_threads=FLAGS.num_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.eval_shuffle,
seed=FLAGS.eval_seed,
fix_random=FLAGS.eval_fix_random,
num_prefetch_batches=FLAGS.num_prefetch_batches,
min_after_dequeue=FLAGS.min_after_dequeue,
fix_sequence=FLAGS.fix_sequence,
name=self.input_valid_name)
eval_batch_size = FLAGS.eval_batch_size
eval_result = eval_image_name, eval_text, eval_text_str, eval_input_text, eval_input_text_str
eval_show_result = None
if FLAGS.show_eval:
eval_fixed = bool(FLAGS.fixed_valid_input)
self.eval_fixed = eval_fixed
if eval_fixed:
assert FLAGS.fixed_eval_batch_size >= FLAGS.num_fixed_evaluate_examples, '%d %d'%(FLAGS.fixed_eval_batch_size, FLAGS.num_fixed_evaluate_examples)
logging.info('fixed_eval_batch_size:{}'.format(FLAGS.fixed_eval_batch_size))
logging.info('num_fixed_evaluate_examples:{}'.format(FLAGS.num_fixed_evaluate_examples))
logging.info('num_evaluate_examples:{}'.format(FLAGS.num_evaluate_examples))
#------------------- fixed valid
fixed_validset = list_files(FLAGS.fixed_valid_input)
logging.info('fixed_validset:{} {}'.format(len(fixed_validset), fixed_validset[:2]))
fixed_image_name, fixed_text, fixed_text_str, fixed_input_text, fixed_input_text_str = inputs(
fixed_validset,
decode=decode,
batch_size=FLAGS.fixed_eval_batch_size,
fix_sequence=True,
num_prefetch_batches=FLAGS.num_prefetch_batches,
min_after_dequeue=FLAGS.min_after_dequeue,
name=self.fixed_input_valid_name)
#-------------shrink fixed image input as input batch might large then what we want to show, only choose top num_fixed_evaluate_examples
fixed_image_name = melt.first_nrows(fixed_image_name, FLAGS.num_fixed_evaluate_examples)
fixed_text = melt.first_nrows(fixed_text, FLAGS.num_fixed_evaluate_examples)
fixed_text_str = melt.first_nrows(fixed_text_str, FLAGS.num_fixed_evaluate_examples)
fixed_input_text = melt.first_nrows(fixed_input_text, FLAGS.num_fixed_evaluate_examples)
fixed_input_text_str = melt.first_nrows(fixed_input_text_str, FLAGS.num_fixed_evaluate_examples)
#notice read data always be FLAGS.fixed_eval_batch_size, if only 5 tests then will wrapp the data
eval_image_name = tf.concat([fixed_image_name, eval_image_name], axis=0)
#melt.align only need if you use dynamic batch/padding
if FLAGS.dynamic_batch_length:
fixed_text, eval_text = melt.align_col_padding2d(fixed_text, eval_text)
fixed_input_text, eval_input_text = melt.align_col_padding2d(fixed_input_text, eval_input_text)
eval_text = tf.concat([fixed_text, eval_text], axis=0)
eval_text_str = tf.concat([fixed_text_str, eval_text_str], axis=0)
eval_input_text = tf.concat([fixed_input_text, eval_input_text], axis=0)
eval_input_text_str = tf.concat([fixed_input_text_str, eval_input_text_str], axis=0)
eval_batch_size = FLAGS.num_fixed_evaluate_examples + FLAGS.eval_batch_size
#should aways be FLAGS.num_fixed_evaluate_examples + FLAGS.num_evaluate_examples
num_evaluate_examples = min(eval_batch_size, FLAGS.num_fixed_evaluate_examples + FLAGS.num_evaluate_examples)
print('----num_evaluate_examples', num_evaluate_examples)
print('----eval_batch_size', eval_batch_size)
#------------combine valid and fixed valid
evaluate_image_name = melt.first_nrows(eval_image_name, num_evaluate_examples)
evaluate_text_str = melt.first_nrows(eval_text_str, num_evaluate_examples)
evaluate_text = melt.first_nrows(eval_text, num_evaluate_examples)
evaluate_input_text_str = melt.first_nrows(eval_input_text_str, num_evaluate_examples)
evaluate_input_text = melt.first_nrows(eval_input_text, num_evaluate_examples)
self.num_evaluate_examples = num_evaluate_examples
eval_result = eval_image_name, eval_text, eval_text_str, eval_input_text, eval_input_text_str
eval_show_result = evaluate_image_name, evaluate_text, evaluate_text_str, evaluate_input_text, evaluate_input_text_str
return eval_result, eval_show_result, eval_batch_size
def gen_valid_input(self, inputs, decode_fn):
#---------------------- valid
validset = list_files(FLAGS.valid_input)
logging.info('validset:{} {}'.format(len(validset), validset[:2]))
eval_image_name, eval_image_feature, eval_text, eval_text_str, eval_input_text, eval_input_text_str = inputs(
validset,
decode_fn=decode_fn,
batch_size=FLAGS.eval_batch_size,
num_threads=FLAGS.num_threads,
batch_join=FLAGS.batch_join,
shuffle_files=FLAGS.eval_shuffle_files,
seed=FLAGS.eval_seed,
fix_random=FLAGS.eval_fix_random,
num_prefetch_batches=FLAGS.num_prefetch_batches,
min_after_dequeue=FLAGS.min_after_dequeue,
fix_sequence=FLAGS.fix_sequence,
name=self.input_valid_name)
eval_batch_size = FLAGS.eval_batch_size
eval_result = eval_image_name, eval_text, eval_text_str, eval_input_text, eval_input_text_str
eval_show_result = None
if FLAGS.show_eval:
eval_fixed = bool(FLAGS.fixed_valid_input)
self.eval_fixed = eval_fixed
if eval_fixed:
assert FLAGS.fixed_eval_batch_size >= FLAGS.num_fixed_evaluate_examples, '%d %d' % (
FLAGS.fixed_eval_batch_size,
FLAGS.num_fixed_evaluate_examples)
logging.info('fixed_eval_batch_size:{}'.format(
FLAGS.fixed_eval_batch_size))
logging.info('num_fixed_evaluate_examples:{}'.format(
FLAGS.num_fixed_evaluate_examples))
logging.info('num_evaluate_examples:{}'.format(
FLAGS.num_evaluate_examples))
#------------------- fixed valid
fixed_validset = list_files(FLAGS.fixed_valid_input)
logging.info('fixed_validset:{} {}'.format(
len(fixed_validset), fixed_validset[:2]))
fixed_image_name, fixed_image_feature, fixed_text, fixed_text_str, fixed_input_text, fixed_input_text_str = inputs(
fixed_validset,
decode_fn=decode_fn,
batch_size=FLAGS.fixed_eval_batch_size,
fix_sequence=True,
num_prefetch_batches=FLAGS.num_prefetch_batches,
min_after_dequeue=FLAGS.min_after_dequeue,
name=self.fixed_input_valid_name)
#-------------shrink fixed image input as input batch might large then what we want to show, only choose top num_fixed_evaluate_examples
fixed_image_name = melt.first_nrows(
fixed_image_name, FLAGS.num_fixed_evaluate_examples)
fixed_image_feature = melt.first_nrows(
fixed_image_feature, FLAGS.num_fixed_evaluate_examples)
fixed_text = melt.first_nrows(
fixed_text, FLAGS.num_fixed_evaluate_examples)
fixed_text = melt.make_batch_compat(fixed_text)
fixed_text_str = melt.first_nrows(
fixed_text_str, FLAGS.num_fixed_evaluate_examples)
fixed_input_text = melt.first_nrows(
fixed_input_text, FLAGS.num_fixed_evaluate_examples)
fixed_input_text = melt.make_batch_compat(fixed_input_text)
fixed_input_text_str = melt.first_nrows(
fixed_input_text_str, FLAGS.num_fixed_evaluate_examples)
#notice read data always be FLAGS.fixed_eval_batch_size, if only 5 tests then will wrapp the data
eval_image_name = tf.concat(
[fixed_image_name, eval_image_name], axis=0)
eval_image_feature = tf.concat(
[fixed_image_feature, eval_image_feature], axis=0)
#melt.align only need if you use dynamic batch/padding
#work well before but for imtxt2txt seems has bug here, strange TODO FIXME in imtxt2txt sh ./train/fixme.sh will reproduce bug
#Maybe write sequence exmaple will be fine ?
if FLAGS.dynamic_batch_length:
fixed_text, eval_text = melt.align_col_padding2d(
fixed_text, eval_text)
fixed_input_text, eval_input_text = melt.align_col_padding2d(
fixed_input_text, eval_input_text)
eval_text = tf.concat([fixed_text, eval_text], axis=0)
eval_text_str = tf.concat([fixed_text_str, eval_text_str],
axis=0)
eval_input_text = tf.concat(
[fixed_input_text, eval_input_text], axis=0)
eval_input_text_str = tf.concat(
[fixed_input_text_str, eval_input_text_str], axis=0)
eval_batch_size = FLAGS.num_fixed_evaluate_examples + FLAGS.eval_batch_size
tf.add_to_collection('fixed_input_text', fixed_input_text)
tf.add_to_collection('fixed_text', fixed_text)
tf.add_to_collection('eval_text', eval_text)
#should aways be FLAGS.num_fixed_evaluate_examples + FLAGS.num_evaluate_examples
num_evaluate_examples = min(
eval_batch_size, FLAGS.num_fixed_evaluate_examples +
FLAGS.num_evaluate_examples)
print('----num_evaluate_examples', num_evaluate_examples)
print('----eval_batch_size', eval_batch_size)
#------------combine valid and fixed valid
evaluate_image_name = melt.first_nrows(eval_image_name,
num_evaluate_examples)
evaluate_image_feature = melt.first_nrows(eval_image_feature,
num_evaluate_examples)
evaluate_text_str = melt.first_nrows(eval_text_str,
num_evaluate_examples)
evaluate_text = melt.first_nrows(eval_text, num_evaluate_examples)
evaluate_input_text_str = melt.first_nrows(eval_input_text_str,
num_evaluate_examples)
evaluate_input_text = melt.first_nrows(eval_input_text,
num_evaluate_examples)
self.num_evaluate_examples = num_evaluate_examples
eval_result = eval_image_name, eval_image_feature, eval_text, eval_text_str, eval_input_text, eval_input_text_str
eval_show_result = evaluate_image_name, evaluate_image_feature, evaluate_text, evaluate_text_str, evaluate_input_text, evaluate_input_text_str
return eval_result, eval_show_result, eval_batch_size