decode_image return tensor without shape in python

[ZhangBowen]

System Information

• OS Platform and Distribution (i.e. Linux Ubuntu 16.0): CentOS Linux release 7.0.1406 (Core)
• TensorFlow installed from (source or binary)?: python3 pip
• TensorFlow version : v1.0.0-65-g4763edf-dirty 1.0.1
• Python version : Python 3.5.2 :: Anaconda custom (64-bit)

Describe the problem clearly

When I decode a jpeg file to a tensor with decode_jpeg, the code ran normally.
but if I use decode_image instead of decode_jpeg for more compatibilities, It raised a ValueError as follow.
Is It a bug?
Thanks

Source Code / Logs

def preprocess_image(image_path):
    file_content = tf.read_file(image_path)
    
    #image = tf.image.decode_jpeg(file_content)
    image = tf.image.decode_image(file_content)
    image = tf.image.per_image_standardization(image)
    image = tf.image.resize_images(image, [480, 640])

output :

Traceback (most recent call last):
  File "/data/home/zhangbowen/is_vehicle/classify.py", line 29, in <module>
    main()
  File "/data/home/zhangbowen/is_vehicle/classify.py", line 26, in main
    image = preprocess_image(image_file)
  File "/data/home/zhangbowen/is_vehicle/classify.py", line 15, in preprocess_image
    image = tf.image.resize_images(image, [480, 640])
  File "/data/home/zhangbowen/anaconda3/lib/python3.5/site-packages/tensorflow/python/ops/image_ops_impl.py", line 643, in resize_images
    raise ValueError('\'images\' contains no shape.')
ValueError: 'images' contains no shape.

[drpngx]

The decode_image is a cascade of tf.cond.
resize_images needs to pull the shape object. However, we don't know which one it's going to be upfront.

@cwhipkey any comment here?

[cwhipkey]

I think the issue is that in the decode_image cond, one branch leads to decode_gif. decode_gif returns a 4-d tensor: [num_frames, height, width, channels], which is not the same as the 3-d tensor returned by decode_png and decode_jpeg.

One possible fix would be to make decode_gif have an attribute controlling whether to handle animated gifs ,and use that in decode_image? (possibly also with an option? It's not clear to me someone would want to use decode_image with a mix of animated and non-animated images - so maybe just changing it would be fine?).

+raingo and wicke

[cwhipkey]

@raingo @martinwicke

[martinwicke]

We could definitely add "frame=None" which is without effect for png and jpg, but for gif will return all frames if None, of the given frame number if not None and the gif is animated. This would work for other animated formats if those become available.

If frame != None we can set the rank to [None, None, None], maybe even [None, None, 3] -- not sure that'll help.

[raingo]

@martinwicke I can send PR to add the frame=None to https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/image_ops_impl.py#L1338.

      with ops.control_dependencies([assert_decode, assert_channels]):
        res = gen_image_ops.decode_gif(contents)
        if frame is not None:
           res = res[frame]
        return res

It makes the least changes to the code, and add some flexibility.

Not sure if that is the best way to do it.

[girving]

I think the right way to fix this is to make the image resizing machinery not need static shape. I'll take a look at how difficult that is.

[cwhipkey]

Having decode_image sometimes return 4-d and sometimes return 3-d does not seem useful though - should it always just return frame 0 of a gif? (this was a change in behavior of decode_image when animated gif support was added)

[girving]

@cwhipkey It was a backwards incompatible change?

[cwhipkey]

No, I don't think so. DecodeGif always returned 4D tensor. decode_image in python was added after that, and always returned a 4D for GIF and 3D for png or jpg. So nm on what I wrote earlier about having it always return frame 0.

[drpngx]

CC: @girving

[girving]

I'm going to restrict my current change to unify decode_jpeg and decode_png: both ops will handle both formats.

[girving]

decode_png and decode_jpeg now decode all formats.

[ZhangBowen]

thanks

[lizhengwei1992]

I'm going to restrict my current change to unify decode_jpeg and decode_png: both ops will handle both formats.

how to judge the format of the image ?

[azizullah2017]

I found the solution

tf.enable_eager_execution()

add this to the top of your code.

or

img_tensor = tf.image.decode_jpeg(img_raw,channels=3)
img_final = tf.image.resize_images(img_tensor, [192, 192])
img_final = img_final/255.0
print(img_final.shape)

[reubencmoyo]

I found the solution

tf.enable_eager_execution()

add this to the top of your code.

or

img_tensor = tf.image.decode_jpeg(img_raw,channels=3)
img_final = tf.image.resize_images(img_tensor, [192, 192])
img_final = img_final/255.0
print(img_final.shape)

This worked for me. Thanks

tf.enable_eager_execution()