Models: [object detection] Why is the num_examples 8000 in mask_rcnn_resnet101_atrous_coco.config?

[haichaoyu]

As far as I know, the number of eval images in mscoco is 5000. Where does 8000 come? Thanks.

[tensorflowbutler]

Thank you for your post. We noticed you have not filled out the following field in the issue template. Could you update them if they are relevant in your case, or leave them as N/A? Thanks.
What is the top-level directory of the model you are using
Have I written custom code
OS Platform and Distribution
TensorFlow installed from
TensorFlow version
Bazel version
CUDA/cuDNN version
GPU model and memory
Exact command to reproduce

[pkulzc]

8000 happens to be the number we use in internal split. This fields will be deprecated after this PR . You can remove it and by default the the eval job will eval all images.

[eewindfly]

@pkulzc thanks your explanation.
However, after I try to remove "num_examples" argument, it is still default by 5000 examples instead of eval all images. Is this an intentional behavior? I found the assumption of all images evaluated is ok only when total number of eval images is below 5000, that repeated images are ignored because of the same image ID detected.

[pkulzc]

If you sync to latest, num_examples is already deprecated so all images will be eval'ed. There is another way to control the rough number of eval images though. Try using sample_1_of_n_examples in input_reader in the config file. See detailed explanation for it here.

[eewindfly]

@pkulzc , thanks so much for your tips.
But some error occurred when I try to set run_once flag for evaluating.
Therefore, I send out a PR for fixing this problem.
Could you kindly review https://github.com/tensorflow/models/pull/6010? Thanks.

[tispratik]

@pkulzc can you please advise on how to check version of Tensorflow this was fixed?