The generate_tfrecord.py should read the image width and height from the csv, not from the image itself.