Implement advanced indexing (and mixed basic/advanced)

[aselle]

NumPy style advanced indexing is documented here http://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#advanced-indexing

We currently support basic indexing http://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#basic-slicing-and-indexing
using StridedSlice.

e.g.

foo = Some tensor
idx1=[1,2,3]
idx2=[3,4,5]
foo[idx1, idx2, 3]

[aselle]

Broken off of #206

[shoyer]

I mentioned this in #206, but I wanted to remind again that mixed indexing with slices/arrays has some really strange/unpredictable behavior for the order of result axes in NumPy. So it might be better to hold off on implementing that in TensorFlow until we're sure we're doing it right.

[aselle]

Yes, I am aware of how the mixed advanced/basic indexing is super nonintuitive. I could break the mixed indexing into its own issue.

[yaroslavvb]

idx1/idx2 in your example could be Tensor objects, right?

[Rob-Haslinger-Bose]

does .10 currently support negative indexing? as in X[:,-1] for example?

[aselle]

@yaroslavvb: yes, idx1 andidx2 can be tensor objects. even in what is already implemented for baisc indexing, you can do

a=tf.constant(3); 
b=tf.constant(6);
foo[a:b]

@robsync: negative indices have a bug in 0.10 and work in 0.11rc0

[Fenugreek]

@aselle: I saw in #206 according to your 7/12 comment that lvalue basic indexing has been implemented. But what does that mean? Because whenever I do any kind of lvalue indexing

e.g.

foo = Some tensor
foo[:3] = tf.zeros(3)

I get TypeError: 'Tensor' object does not support item assignment.
Thanks.

[shoyer]

@Fenugreek Tensors are immutable. But you can do this sort of thing if foo is a tf.Variable by writing foo[:3].assign(tf.zeros(3))

[danijar]

...and making sure that it gets executed:

with tf.control_dependencies([foo[:3].assign(tf.zeros(3))]):
    foo = tf.identity(foo)

[Fenugreek]

Great, thanks @shoyer and @danijar . I am past that problem now, but have another error with the basic indexing:

        print pool.get_shape(), args.get_shape(), values.get_shape()
        tf.assign(pool[args], values)

gives

(1024000,) (512000,) (512000,)
Traceback (most recent call last):
...
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/array_ops.py", line 1644, in _DelegateStridedSliceShape
    return common_shapes.call_cpp_shape_fn(op, input_tensors_needed=[1, 2, 3])
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/common_shapes.py", line 596, in call_cpp_shape_fn
    raise ValueError(err.message)
ValueError: Shape must be rank 1 but is rank 2

There are no rank 2 shapes anywhere, so I am confused. The same error results even if I just do a return pool[args], so has nothing to do with the assignment but with the basic indexing itself. Thanks again.

[charlienash]

I'm having a similar error to @Fenugreek:

A = tf.constant([[1,2],[3,4],[5,6]])
id_rows = tf.constant([0,2])
A[id_rows, :]

gives an error:

ValueError: Shape must be rank 1 but is rank 2 for 'strided_slice' 
(op: 'StridedSlice') with input shapes: [3,2], [1,2], [1,2], [1].

This is using version 1.0.0. Thanks.

[aselle]

You need to manually broadcast it i.e. values.get_shape() must match pool[args].get_shape. This is a limitation in the current implementation.

[Kublai-Jing]

@aselle Can you elaborate more on how to manually broadcast to make it work ? Thanks.

[shoyer]

StackOverflow would be a great place to ask for tips on how to use broadcasting to make gather_nd work

…

On Mon, Apr 3, 2017 at 6:56 PM, Kublai-Jing ***@***.***> wrote: @aselle <https://github.com/aselle> Can you elaborate more on how to manually broadcast to make it work ? Thanks. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4638 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABKS1tCXk0PpQctDC7VHNcWE18soT37sks5rsaNkgaJpZM4KJUyn> .

[cancan101]

I posted the issue to SO: https://stackoverflow.com/questions/44793286/slicing-tensorflow-tensor-with-tensor.

[cancan101]

@itsmeolivia why was the issue closed? It doesn't seem like the example in the OP works.

[itsmeolivia]

This question is better asked on StackOverflow since it is not a bug or feature request. There is also a larger community that reads questions there. Thanks!

[cancan101]

I'm asking why the ticket was closed given the op was in fact a feature. Was it implemented? Or decided to not implement. There is no info in the ticket as to why it was closed.

[brianwa84]

@aselle probably want to keep this open right?

[SpinachR]

I have realize the advanced indexing like the same way in numpy. Please refer to:
https://github.com/SpinachR/ubuntuTest/blob/master/beautifulCodes/tensorflow_advanced_index_slicing.ipynb

[tensorflowbutler]

Nagging Assigneee: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[tensorflowbutler]

Nagging Assignee: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[yifeif]

@aselle do you know if we have plan to implement this? If not, should we mark this as contribution welcome?

[asimshankar]

CC @saxenasaurabh

[traveller59]

use gather to do advanced index need much more code than numpy/pytorch, and make code hard to read. please implement this if possible.
I have implemented a simple function to do some numpy-style advanced indexing based on tf.gather_nd and tf.transpose.

[tensorflowbutler]

Nagging Assignee @aselle: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[tensorflowbutler]

Nagging Assignee @aselle: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[tensorflowbutler]

Nagging Assignee @aselle: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[tensorflowbutler]

Nagging Assignee @aselle: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[aselle]

@saxenasaurabh, could you take this issue over?

[aselle]

@saxenasaurabh, could you please take this issue over?

[tensorflowbutler]

Please remove the assignee, as this issue is inviting external contributions. Otherwise, remove the contributions welcome label. Thank you.

[tensorflowbutler]

Please remove the assignee, as this issue is inviting external contributions. Otherwise, remove the contributions welcome label. Thank you.

[tensorflowbutler]

Please remove the assignee, as this issue is inviting external contributions. Otherwise, remove the contributions welcome label. Thank you.

[tensorflowbutler]

Please remove the assignee, as this issue is inviting external contributions. Otherwise, remove the contributions welcome label. Thank you.

[tensorflowbutler]

Please remove the assignee, as this issue is inviting external contributions. Otherwise, remove the contributions welcome label. Thank you.

[gokul-uf]

@aselle Is there active development on this front or is using tf Eager the less painful way to go?

Thanks!

[asimshankar]

@gokul-uf - Nope, this is not being actively developed at this time. Not quite sure if using eager helps you either. You could easily convert a Tensor to numpy and use advanced indexing, but if you want to compute gradients through that indexing, that won't work.

[shoyer]

On a related note, it might be worth reviewing our proposal for how to change indexing behavior in NumPy moving forward: http://www.numpy.org/neps/nep-0021-advanced-indexing.html

…

On Thu, Jul 26, 2018 at 10:44 AM Asim Shankar ***@***.***> wrote: @gokul-uf <https://github.com/gokul-uf> - Nope, this is not being actively developed at this time. Not quite sure if using eager helps you either. You could easily convert a Tensor to numpy and use advanced indexing, but if you want to compute gradients through that indexing, that won't work. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4638 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABKS1lwYXKQeWKJ36lMhrDG_rxT9eTAtks5uKf_jgaJpZM4KJUyn> .

[sathyarr]

@Fenugreek Tensors are immutable. But you can do this sort of thing if foo is a tf.Variable by writing foo[:3].assign(tf.zeros(3))

I have been trying to replace one of the tf.py_func. Please refer this for clear description of the problem.

I'm stuck on replacing tensor's values which is created as an Object and not a Variable.
Any workarounds possible for this case?

Similar Stack Overflow Question

[un-lock-me]

Is it doable something like this in tensorflow:

out[a2[y], x[:, None]] = alp[a2[y], x[:, None]]

please refer to this post for reproducble example.

Thanks

[aselle]

This issue is way to broad and stale. However, anyone needing advanced indexing functionality should consider the excellent tf.experimental.numpy` module by @wangpengmit

[wangpengmit]

(CCing @cantonios)