Image preprocessing is INCONSISTENT at train and evaluate stage causes InvalidArgumentError

[ytliang97]

Hi, first thanks for getting this done.

Recently I encountered following problem when I try to run evaluate.py:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Number of ways to split should evenly divide the split dimension, but got split_dim 2 (size = 1) and num_split 3

I've noticed that the preprocessing codes in train.py use tf.reshape()before decoding images :
use tf.reshape() can convert image from grayscale to rgb.

tensorflow-deeplab-v3/train.py

Lines 140 to 141 in a5d7ff2

	image = tf.image.decode_image(
	tf.reshape(parsed['image/encoded'], shape=[]), _DEPTH)

tensorflow-deeplab-v3/train.py

Lines 145 to 146 in a5d7ff2

	label = tf.image.decode_image(
	tf.reshape(parsed['label/encoded'], shape=[]), 1)

but in evaluate.py, the preprocessing codes didn't use tf.reshape() :

tensorflow-deeplab-v3/utils/preprocessing.py

Line 233 in a5d7ff2

image = tf.image.decode_image(image_string)

tensorflow-deeplab-v3/utils/preprocessing.py

Line 243 in a5d7ff2

label = tf.image.decode_image(label_string)

This will raise error if there is a grayscale image in dataset. So I edit code like this:

image = tf.image.decode_image( tf.reshape(image_string, shape=[]), 3 )

label = tf.image.decode_image( tf.reshape(label_string, shape=[]), 1 )

It works!