Fix wrong accuracy fetches when single_session is used.

[gyeongchan-yun]

Hello,

I'm wondering whether accuracy op is correct or not in 'distributed_all_reduce'.
To support my doubt, the log message is here. Top-5 accuracy is over 1.0!

I run it on 4 servers with 4 GPUs (total GPUs: 16) with the following configuration.
Model: resenet56
Dataset: cifar10 (synthetic)
Mode: BenchmarkMode.TRAIN
SingleSess: True
Variables: distributed_all_reduce
AllReduce: xring

Step Img/sec total_loss top_1_accuracy top_5_accuracy
1 images/sec: 5.6 +/- 0.0 (jitter = 0.0) 3.286 0.445 1.906
10 images/sec: 52.6 +/- 81.8 (jitter = 29.5) 2.580 0.484 1.945
20 images/sec: 99.3 +/- 42.3 (jitter = 29.1) 2.522 0.500 2.164
30 images/sec: 141.0 +/- 28.7 (jitter = 19.9) 2.519 0.398 1.852
40 images/sec: 178.4 +/- 21.7 (jitter = 17.6) 2.511 0.391 1.992
50 images/sec: 212.3 +/- 17.7 (jitter = 24.9) 2.507 0.375 2.047
60 images/sec: 242.8 +/- 14.8 (jitter = 25.6) 2.509 0.312 1.891
70 images/sec: 270.8 +/- 12.7 (jitter = 27.8) 2.503 0.391 2.094
80 images/sec: 297.0 +/- 11.4 (jitter = 29.8) 2.502 0.414 2.008
90 images/sec: 321.3 +/- 10.4 (jitter = 34.8) 2.503 0.391 1.914
100 images/sec: 343.0 +/- 9.5 (jitter = 38.9) 2.505 0.391 1.883

In benchmark_cnn.py,

  def _build_fetches(self, global_step, all_logits, losses, device_grads,
                     enqueue_ops, update_ops, all_accuracy_ops, phase_train):
    """Complete construction of model graph, populating the fetches map."""
    fetches = {}
    if enqueue_ops:
      fetches['enqueue_ops'] = enqueue_ops
    for name, ops in all_accuracy_ops.items():
      if name.startswith(constants.UNREDUCED_ACCURACY_OP_PREFIX):
        key = name[len(constants.UNREDUCED_ACCURACY_OP_PREFIX):]
        fetches[key] = tf.concat(ops, 0)
      else:
        # The problem if single_session is used.
        fetches[name] = tf.reduce_sum(ops) / self.batch_size  
        if self.task_index == 0 and self.params.summary_verbosity >= 1:
          tf.summary.scalar(name, fetches[name])

I think this is correct accuracy_ops if single_session is used.

fetches[name] = (tf.reduce_sum(ops) / self.batch_size * 
                 (self.num_workers  if self.single_session else 1))

Best regards,
Gyeongchan Yun

[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 with @googlebot 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.

[gyeongchan-yun]

@googlebot I signed it!

[googlebot]

CLAs look good, thanks!

ℹ️ Googlers: Go here for more info.

[reedwm]

I think you want to divide by num_workers here, not multiply. Or alternatively, put parenthesis around self.batch_size * (self.num_workers if self.single_session else 1)

[gyeongchan-yun]

You're right. I want to divide num_workers if single_session is used. Your alternative suggestion is right. 👍

[gyeongchan-yun]

I've updated it.

Best regards,
Gyeongchan Yun

[reedwm]

You are failing some tests, like testReplicated. To reproduce, run:

python -m unittest -v benchmark_cnn_test.TfCnnBenchmarksTest.testReplicated

The issue is self.single_session is not set during evaluation. Probably the best way to fix this would be to set self.single_session in the constructor:

self.single_session = params.variable_update == 'distributed_all_reduce'

Then, change the if-statement here to check if self.single_session is True instead, and no longer set self.single_session in the _build_graph function.

[gyeongchan-yun]

You are failing some tests, like testReplicated. To reproduce, run:

python -m unittest -v benchmark_cnn_test.TfCnnBenchmarksTest.testReplicated

The issue is self.single_session is not set during evaluation. Probably the best way to fix this would be to set self.single_session in the constructor:

self.single_session = params.variable_update == 'distributed_all_reduce'

Then, change the if-statement here to check if self.single_session is True instead, and no longer set self.single_session in the _build_graph function.

Thanks for your kind suggestion!
I've updated it and confirmed passing the test.

Best Regards,
Gyeongchan Yun