CODEWITHSUNDEEP
tensorflow
Łukasz Sromek
Łukasz Sromek

Reputation: 3677

Mapping tensors to operations in tensorflow

I'm trying to combine a number of tf.write_file ops into single op to save a batch of images inside single Session.run() call, with no luck.

This code:

r = tf.range(tf.shape(images)[0])
crops = tf.image.crop_and_resize(images, boxes, r, [160, 160])
crops = tf.image.convert_image_dtype(crops, dtype=tf.uint8)

def write_ith_image(i):
    nonlocal out_file_tensor, crops
    encoded_image = tf.image.encode_jpeg(crops[i])
    return tf.write_file(out_file_tensor[i], encoded_image)

encode_all = tf.map_fn(write_ith_image, r)

results in:

TypeError: Can't convert Operation 'map/while/WriteFile' to Tensor (target dtype=None, name='value', as_ref=False)

tf.while_loop gives similar results. Is there a way to map a tensor to a group of operations that can be executed in single tf.Session.run() call? Why this works for tensors, but not for operations with side effects?

Upvotes: 1

Views: 226

Answers (1)

P-Gn
P-Gn

Reputation: 24581

You could return a dummy value.

def write_ith_image(i):
    nonlocal out_file_tensor, crops
    encoded_image = tf.image.encode_jpeg(crops[i])
    with tf.control_dependencies([tf.write_file(out_file_tensor[i], encoded_image)]):
      dummy = tf.constant([0])
    return dummy

encode_all = tf.map_fn(write_ith_image, r)

Not pretty, but gets the job done.

You could get something slightly more satisfying with tf.while_loop (no dummy values), but it would be more verbose and probably not any more efficient.

Upvotes: 2

Related Questions