fix issue #27985

[pminervini]

The dresult variable may be an IndexedSlices object, which does not have a _copy method.

Link to the issue: #27985

[googlebot]

Thanks for your pull request. It looks like this may be your first contribution to a Google open source project (if not, look below for help). Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

📝 Please visit https://cla.developers.google.com/ to sign.

Once you've signed (or fixed any issues), please reply here (e.g. I signed it!) and we'll verify it.

---

What to do if you already signed the CLA

Individual signers

• It's possible we don't have your GitHub username or you're using a different email address on your commit. Check your existing CLA data and verify that your email is set on your git commits.

Corporate signers

• Your company has a Point of Contact who decides which employees are authorized to participate. Ask your POC to be added to the group of authorized contributors. If you don't know who your Point of Contact is, direct the Google project maintainer to go/cla#troubleshoot (Public version).
• The email used to register you as an authorized contributor must be the email used for the Git commit. Check your existing CLA data and verify that your email is set on your git commits.
• The email used to register you as an authorized contributor must also be attached to your GitHub account.

ℹ️ Googlers: Go here for more info.

[pminervini]

I signed it!

[googlebot]

CLAs look good, thanks!

ℹ️ Googlers: Go here for more info.

[alextp]

Thanks for the contribution!

I'd like a test which fails before this change and passes after, so we don't accidentally break this again.

[pminervini]

@alextp yeah let me briefly look into that

[pminervini]

I managed to reproduce the problem with a small self-contained script:

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import tensorflow as tf
import tensorflow.contrib.eager as tfe

tf.enable_eager_execution()

shape = (5, 5)
initializer = tf.random_uniform_initializer(-1.0, 1.0)
var = tfe.Variable(initializer(shape), name='tensor')

with tf.GradientTape() as tape:
        a = tf.gather(tf.gather(var, [0, 1]), [0, 1])
        b = tf.gather(tf.gather(var, [2, 3]), [0, 1])
        r = tf.einsum('ij,ij->i', a, b)
g = tape.gradient(r, [var])

Here's the output - I'll commit a unit test shortly.

$ ./fun.py 
2019-04-19 20:30:43.159014: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-04-19 20:30:43.240709: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-04-19 20:30:43.241170: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 0 with properties: 
name: TITAN X (Pascal) major: 6 minor: 1 memoryClockRate(GHz): 1.531
pciBusID: 0000:02:00.0
totalMemory: 11.91GiB freeMemory: 5.75GiB
2019-04-19 20:30:43.241190: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1511] Adding visible gpu devices: 0
2019-04-19 20:30:43.517288: I tensorflow/core/common_runtime/gpu/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-04-19 20:30:43.517315: I tensorflow/core/common_runtime/gpu/gpu_device.cc:988]      0 
2019-04-19 20:30:43.517324: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1001] 0:   N 
2019-04-19 20:30:43.517479: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 5526 MB memory) -> physical GPU (device: 0, name: TITAN X (Pascal), pci bus id: 0000:02:00.0, compute capability: 6.1)
Traceback (most recent call last):
  File "./fun.py", line 17, in <module>
    g = tape.gradient(r, [var])
  File "/home/anaconda/anaconda3/lib/python3.6/site-packages/tensorflow/python/eager/backprop.py", line 901, in gradient
    output_gradients=output_gradients)
  File "/home/anaconda/anaconda3/lib/python3.6/site-packages/tensorflow/python/eager/imperative_grad.py", line 64, in imperative_grad
    output_gradients)
  File "/home/anaconda/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 858, in grad_fun
    return [dresult._copy(device_name=self_device)]
AttributeError: 'IndexedSlices' object has no attribute '_copy'

[pminervini]

@alextp this should do the trick - what do you think?

[pminervini]

@alextp link to the test: https://github.com/pminervini/tensorflow/blob/master/tensorflow/python/framework/ops_test.py#L161

Is there a way to enforce this to run on GPU?

[alextp]

Use the @test_util.run_gpu_only decorator

[alextp]

The linter will fail on this line of code.

Also, you should probably put the whole thing inside the [], like return [dresult._copy() if hasattr(dresult, '_copy') else dresult] which will make it easier to split into lines to fit in 80 columns.

[pminervini]

@alextp done!

[pminervini]

@alextp there was one line exceeding 80 chars in ops_test.py, it should be fine now

[pminervini]

@alextp mmm where is the problem?

[alextp]

Nothing wrong, the build is just currently broken.