Changes needed to Converting *.xml to *.csv

[jatansharma]

Following changes I had to make in "generate_tfrecord.py" file to make it work with the given cmd command:

1. Define flag for image path which is passed through command line
   flags.DEFINE_string('img_path', '', 'Path to the images')

2. Inside create_tf_example definition change path to FLAGS.img_path
   with tf.gfile.GFile(os.path.join(FLAGS.img_path, '{}'.format(group.filename)), 'rb') as fid:

Also, I got Nonetype error: None has type NoneType, but expected one of: int, long. To fix this change the else return to 0 instead of null
`
def class_text_to_int(row_label):

if row_label == FLAGS.label:  # 'ship':

    return 1

else:

    return 0

`

PS: let me know if there is a better solution
PPS: Still waiting on label map implementation

[sglvladi]

@jatansharma I am not quite sure why you had to implement the mentioned changes. Could you be a bit more specific? I can admit I haven't used this piece of code in a while (2-3 months), but generally I never had any issues.

Now, when it comes to the class_text_to_int(row_label), as I have mentioned in #1, the generate_tfrecord.py code has not been written by me, but I would disagree with changing the function to return 0 instead of None. I will explain why:

• In general, if your code reaches the point of returning None, it means that the given row_label does not correspond to any previously defined labels.
• Now assume that in your label_map.pbtxt file you have a label which is given an id: 0, eg:

item {
    id: 0
    name: 'cat'
}

• If you apply your change to the code (i.e. to return 0 instead of None), every time you call the class_text_to_int(row_label) function with an unknown row_label, the function will return the value of 0, thus falsely pointing to the cat label.

So, instead of changing the function, I would recommend that you add some code which detects that a None is returned from the function. For example, you could simply add to lines to the following code in generate_tfrecord.py, which will ignore any unknown labels:

for index, row in group.object.iterrows():
    # Add this:
    if class_text_to_int(row['class']) is None:
        continue
    # ===========>
    xmins.append(row['xmin'] / width)
    xmaxs.append(row['xmax'] / width)
    ymins.append(row['ymin'] / height)
    ymaxs.append(row['ymax'] / height)
    classes_text.append(row['class'].encode('utf8'))
    classes.append(class_text_to_int(row['class']))

In terms of a label map implementation, I have not actually done it, but in #1 I briefly explain how this could be achieved. Unfortunately, I don't currently have time to implement/test the change.

Hope the above proves helpful!

[sglvladi]

Closing as there has been no correspondence in 3 months. Feel free to re-open if necessary.