How to resize (downsample) 5D samples in keras?

[innat]

(Reposting from here)

---

Currently, for 5D data (batch_size, h, w, depth, channel), the tf.keras.backend.resize_volumes or UpSampling3D can be used to upsampling purpose. For example, I can do

a  = tf.ones(shape=(1, 100, 100, 64, 1))

tf.keras.backend.resize_volumes(
       a, depth_factor=2, 
       height_factor=2, 
       width_factor=2, 
       data_format="channels_last").shape
TensorShape([1, 200, 200, 128, 1])

These factor values (above) should be an integer. https://github.com/keras-team/keras/blob/master/keras/backend.py#L3441-L3444 - in that case, how can I downsample the input sample, ie.

a  = tf.ones(shape=(1, 100, 100, 64, 1))

tf.keras.backend.resize_volumes(
       a, depth_factor=0.5, 
       height_factor=0.5, 
       width_factor=0.5 
       data_format="channels_last").shape

TypeError: 'float' object cannot be interpreted as an integer

# EXPECTED
TensorShape([1, 50, 50, 32, 1])

And HERE https://stackoverflow.com/q/57341504/9215780, another scenario where the factor needed to be fractional.

---

( PS. downsample or upsample can be done properly in the same manner with scipy.ndimage.zoom )

[bhack]

The big question here is about Keras vs Tensorflow ownership.

Could this be achieved composing TF ops calls available in TF? If not it is a high level API request for Keras but it requires another ticket/PR in TF.

Also more in general I don't know what is the most update plan/scope for the tf.keras.backed namespace from an user perspective (probably @qlzh727 @fchollet could help on this point).

[innat]

It's might be tricky or specialized cases. Normally, in volumetric CT samples, the spatial information mostly appears in the middle of the depth range.

In most cases, the beginning and ending slices are completely black - that makes them useless to use. In that case, in order to downsample the volume depth size, it's needed to resample from the middle range of slices to get the most informative slices. But I'm not sure if this is a general scenario for all volumetric data.

Now, in tf, if we want to downsample volumetric data and set a specific middle-range for the depth axis, we can do as follow - though not sure if it's optimal.

# (input data: 1, 50, 50, 20, 4)
# (desired output: 1, 25, 25, 10, 4)

a = tf.ones(shape=(1, 50, 50, 20, 4))
a.shape # TensorShape([1, 50, 50, 20, 4])

a2 = tf.reshape(a, [-1, 50, 50, 20*4])
a2.shape # TensorShape([1, 50, 50, 80])

a3 = tf.image.resize(a2, [25, 25])
a3.shape # TensorShape([1, 25, 25, 80])

a4 = tf.reshape(a3, [-1, 25, 25, 20, 4])
a4.shape # TensorShape([1, 25, 25, 20, 4])

# HERE, Picking some middle slices -
# - Assuming that by this we may get relevant slices. 
# How convenient is this? # May not general! 
a5 = a4[...,  5:15, :]
a5.shape
TensorShape([1, 25, 25, 10, 4])

[bhack]

a5 = a4[..., 5:15, :]

Is this pseudocode? As I think it will not work with Tensor.

[qlzh727]

Agreed with @bhack. I think this kind of implementation should be pushed to ops level (rather than implement on keras side with py) to achieve the best performance. We aren't going to add more API to keras.backend since it was historically a layering between keras and different backends, and TF is the only backend now.

[innat]

@bhack It worked on eager mode. Maybe will fail on graph mode, not sure I didn't try in actual training.

tf.config.run_functions_eagerly(False)
tf.compat.v1.disable_eager_execution()

a5 = a4[...,  5:15, :] # OK

[innat]

@qlzh727 With the current tf ops, is it possible effectively and safely to do this operation? Or, should I open a ticket in tf.

[bhack]

@bhack It worked on eager mode. Maybe will fail on graph mode, not sure I didn't try in actual training.

tf.config.run_functions_eagerly(False)
tf.compat.v1.disable_eager_execution()

a5 = a4[...,  5:15, :] # OK

Oh sorry I supposed that you need to do a slice assigment later.

[bhack]

It really depend on what you want to do. If the interpolation included in tf.image.resize needs to just work with the current dimensions that It already support It could be ok.
Do we have cases where we need to interpolate across multiple dimensions?

[innat]

Do we have cases where we need to interpolate across multiple dimensions?

Not sure if there are many causes for this. For medical cases, I've seen that people often adopt a resampling approach over all relevant dimensions (h, w, depth).

[bhack]

Probably an interface like:

https://www.mathworks.com/help/images/ref/imresize3.html