tensorflow - How to create dataset in the same format as the FSNS dataset?

Question

I'm working on this project based on TensorFlow.

I just want to train an OCR model by attention_ocr based on my own datasets, but I don't know how to store my images and ground truth in the same format as FSNS datasets.

Is there anybody also work on this project or know how to solve this problem?

Answer 1

The data format for storing training/test is defined in the FSNS paper https://arxiv.org/pdf/1702.03970.pdf (Table 4).

To store tfrecord files with tf.Example protos you can use tf.python_io.TFRecordWriter. There is a nice tutorial, an existing answer on the stackoverflow and a short gist.

Assume you have an numpy ndarray img which has num_of_views images stored side-by-side (see Fig. 3 in the paper): and a corresponding text in a variable text. You will need to define some function to convert a unicode string into a list of character ids padded to a fixed length and unpadded as well. For example:

char_ids_padded, char_ids_unpadded = encode_utf8_string(
   text='abc', 
   charset={'a':0, 'b':1, 'c':2},
   length=5,
   null_char_id=3)

the result should be:

char_ids_padded = [0,1,2,3,3]
char_ids_unpadded = [0,1,2]

If you use functions _int64_feature and _bytes_feature defined in the gist you can create a FSNS compatible tf.Example proto using a following snippet:

char_ids_padded, char_ids_unpadded = encode_utf8_string(
   text, charset, length, null_char_id)
example = tf.train.Example(features=tf.train.Features(
  feature={
    'image/format': _bytes_feature("PNG"),
    'image/encoded': _bytes_feature(img.tostring()),
    'image/class': _int64_feature(char_ids_padded),
    'image/unpadded_class': _int64_feature(char_ids_unpadded),
    'height': _int64_feature(img.shape[0]),
    'width': _int64_feature(img.shape[1]),
    'orig_width': _int64_feature(img.shape[1]/num_of_views),
    'image/text': _bytes_feature(text)
  }
))