CODEWITHSUNDEEP
tensorflow
Clash
Clash

Reputation: 5025

Why doesn't `tf.image.resize_images` set the image shape?

My objective is to introduce random scaling and translation for data augmentation purposes.

distorted_image = tf.image.resize_images(distorted_image, random_scale, random_scale)
distorted_image = tf.image.crop_to_bounding_box(distorted_image, random_y, random_x, 299, 299)

This fails with 'image' must be fully defined. Swapping the lines works, but doesn't do what I really need.

distorted_image = tf.image.crop_to_bounding_box(distorted_image, random_y, random_x, 299, 299)  
distorted_image = tf.image.resize_images(distorted_image, random_scale, random_scale)

So it seems like resize_images loses the shape of the image tensor and then crop_to_bouding_box fails. Is this on purpose, am I missing something? How come random_crop works after a resize but crop_to_bounding_box doesn't?

Upvotes: 1

Views: 1942

Answers (1)

mrry
mrry

Reputation: 126194

The tf.image.resize_images() op does set the image shape, on this line of the implementation. (This was added in TensorFlow 0.7.)

However, if either of the new_height or new_width arguments is a dynamic value, then TensorFlow cannot infer a single shape for that dimension, and so uses None for that dimension. I notice in your code that the new height and width values are called random_scale: if a new random value is drawn on each step, then the shape will have None for the height and width dimensions.

Note that in this case, the tf.image.crop_to_bounding_box() op will not work because—as the error message indicates—the current implementation requires that the shape of the input be fully defined. As I noted in a recent answer, the best workaround might be to use the lower level ops from which tf.image.crop_to_bounding_box() is implemented (in particular tf.slice() with computed indices).

Upvotes: 3

Related Questions