def parse_text_file(text_file):
if not os.path.exists(text_file):
print('No text file', text_file, file=sys.stderr)
return
num_lines = gezi.get_num_lines(text_file)
pb = ProgressBar(num_lines, 'parse text file %s' % text_file)
for line in open(text_file):
pb.progress()
l = line.split('\t')
image = l[0]
image = image[:image.index('#')]
text = l[-1].strip()
#why text and ori_text ? because fo cn corpus text will be \x01 seperated(segged text)
#for en corpus text and ori_text is the same
ori_text = l[FLAGS.ori_text_index].strip()
if text == '':
continue
if image not in text_map:
text_map_[image] = set([text])
text_map[image] = [(text, ori_text)]
else:
if text not in text_map_:
text_map_[image].add(text)
text_map[image].append((text, ori_text))
for image in text_map:
text_map[image] = list(text_map[image])
def convert_to(feat_file, name):
num_shards = FLAGS.shards
num_threads = FLAGS.threads
if FLAGS.threads > 1:
assert(num_threads == num_shards)
f = open(feat_file).readlines()
num_lines = len(f)
if FLAGS.debug:
num_lines = NUM_DEBUG_LINES
shard_ranges = np.linspace(0,
num_lines,
num_shards + 1).astype(int)
record = []
for i in xrange(num_threads):
args = (f, name, i, shard_ranges[i], shard_ranges[i + 1])
process = multiprocessing.Process(target=_convert_to,args=args)
process.start()
record.append(process)
for process in record:
process.join()
return
#--------------single thread
num_lines = gezi.get_num_lines(feat_file)
if FLAGS.debug:
num_lines = NUM_DEBUG_LINES
shard_ranges = np.linspace(0,
num_lines,
num_shards + 1).astype(int)
pb = ProgressBar(num_lines, "convert")
shard = 0
count = 0;
output_filename = '%s-%.5d-of-%.5d' % (name, shard, num_shards)
output_file = os.path.join(FLAGS.output_directory, output_filename)
print('Writing', output_file, count)
writer = tf.python_io.TFRecordWriter(output_file)
for line in open(feat_file):
pb.progress()
if count >= shard_ranges[shard + 1]:
shard += 1
writer.close()
output_filename = '%s-%.5d-of-%.5d' % (name, shard, num_shards)
output_file = os.path.join(FLAGS.output_directory, output_filename)
print('Writing', output_file, count)
writer = tf.python_io.TFRecordWriter(output_file)
_parse_line(line, writer)
count += 1
if FLAGS.debug and count >= NUM_DEBUG_LINES:
break
writer.close()
def _convert_to(f, name, thread_index, start, end):
num_shards = FLAGS.shards
shard = thread_index
output_filename = '%s-%.5d-of-%.5d' % (name, shard, num_shards)
output_file = os.path.join(FLAGS.output_directory, output_filename)
print('Writing', output_file, start)
writer = tf.python_io.TFRecordWriter(output_file)
num_lines = end - start
pb = ProgressBar(num_lines)
for i in xrange(start, end):
pb.progress()
line = f[i]
_parse_line(line, writer, thread_index)
texts_dict[thread_index] = gtexts[thread_index]
text_strs_dict[thread_index] = gtext_strs[thread_index]
import conf
from conf import IMAGE_FEATURE_LEN
import json
import nltk.tokenize
from libprogress_bar import ProgressBar
START_WORD = '<S>'
END_WORD = '</S>'
with tf.gfile.FastGFile(FLAGS.captions_file, "r") as f:
caption_data = json.load(f)
pb = ProgressBar(len(caption_data["annotations"]))
id_to_filename = [(x["id"], x["file_name"]) for x in caption_data["images"]]
print(len(id_to_filename))
ids = set()
for x in caption_data["images"]:
ids.add(x["id"])
print(len(ids))
print(len(caption_data["annotations"]))
caption_data = None
for annotation in caption_data["annotations"]: