WebP support

[sjincho]

Describe the problem

It would be nice if tensorflow supports webp image files.
https://developers.google.com/speed/webp/
Our image dataset gets 2-3 times smaller when using webp compared to png.

[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.
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

[sjincho]

It's about a feature request, so I think they are irrelevant for my case.
Have I written custom code: No
OS Platform and Distribution: macOS, Ubuntu 16.04
TensorFlow installed from: pip
TensorFlow version: 1.4, 1.7
Bazel version: N/A
CUDA/cuDNN version: 8/6
GPU model and memory: Nvidia 1080ti, 11GB
Exact command to reproduce: N/A

[yongtang]

@w4-sjcho A PR has been opened in tensorflow/io (tensorflow/io#43) to add the WebP support through tf.data interface.

The above mentioned PR does not match the decode_png/jpeg/bmp API in the tensorflow repo.

If api review team could confirm the decode_webp API, I could also create a PR in tensorflow repo to implement the decode_webp (similar to decode_png/jpeg/bmp) to bring WebP support to tf.

[yongtang]

@w4-sjcho WebP support has been added to tensorflow/io (tensorflow/io#43). It allows you to feed WebP image format into tf.data and used in tensorflow. You could give it a try if interested.

For decode_webp type of API, will look into api review team first.

[alextp]

decode_webp should also go first in the SIG IO

[yongtang]

@w4-sjcho the decode_webp op was implemented in SIG IO: tensorflow/io#86

You can follow the progress there.

[wingman-jr-addon]

@yongtang I see that this was indeed implemented in tensorflow/io, but it's not clear to me how we track the progress from the SIG IO back to tensorflow here. Presumably after a year, webp support would be considered stable and ready for tensorflow? It seems like it might be cumbersome to use decode_webp without fully plumbing into decode_image in the situation where there are mixed image types in the input set so I'm hopeful that this will land in Tensorflow once it is considered ready. Do you know what the status is? Thanks!

[yongtang]

@wingman-jr-addon It is likely decode_webp will remain in SIG IO in the foreseeable future, given TF is trying to modularize the components to different repos and tensorflow/io is part of the modularization process.

With respect to decode_image, a likely path would be to have a decode_image in tensorflow/io instead, that covers a large superset of the supported image format.

I have created a tensorflow/io#930 to track the issue there.

[wingman-jr-addon]

Thank you for taking the time to respond and explain the current situation! I will watch the newly created issue.