Add image-pipe ops for zero-copy image generation

[ghostplant]

An ops in place of tf.keras.preprocessing.ImageDataGenerator with very high-performance data generation performance loading from on-disk original image directories to GPU directly with ZeroCopy, which could achieve ~96% performance of synthetic dataset training for modern models like Resnet50/Inception3/..

1. Multi-worker deterministic image input by configuration of seed, which is not supported by tf.keras.preprocessing.ImageDataGenerator;
2. Support direct image generation with either NCHW or NHWC format;
3. Support target image resize in place and interleaving generation;
4. Reference of internal image directory format -

/train/
    /class-monkey/
        aug_1.jpg
        aug_2.jpg
    /class-bird/
        aug_1.jpg
        aug_2.jpg

[mrry]

Thanks for the contribution! Since we are in the process of sunsetting tf.contrib, we are not currently accepting new submodules.

However, this contribution might make sense as part of the Special Interest Group on IO, which is planned to host a repository of community-maintained I/O-related code (primarily Dataset and FileSystem implementations).

I'm going to close this PR for now, but please consider joining the SIG IO mailing list, and contributing to that repository when it is set up.

/cc @ewilderj

[ghostplant]

OK, Thanks.

[fangjyshanghai]

Thanks

[yongtang]

@ghostplant The sig-io and its repo is in place now: https://github.com/tensorflow/io

You can consider open a PR in the repo and join the discussion in sig-io groups: https://groups.google.com/a/tensorflow.org/forum/#!forum/io

[ghostplant]

@yongtang Yeah, that's great!

[ghostplant]

@yongtang Do you know there is a C-based header library included in Tensorflow that could iterate filesystem's directories and file elements on any generic OS platform (e.g. Linux, Windows, MacOS) like what os.walk in python does?

[ghostplant]

OK, I found a good reference here: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/platform/posix/posix_file_system.cc

[feihugis]

@ghostplant You can also check MatchingFilesDatasetOp, which can iterate filesystem's directories and files on Linux, Windows, and Mac as well. Its Python API is here.