TensorFlowJS Unsupported Ops needed for Yahoo's Open NSFW Model

[tomslick]

To get help from the community, check out our Google group.

TensorFlow.js version

tensorflowjs 0.4.1
(installed via pip install tensorflowjs on June 13, 2018)

Browser version

N/A

Describe the problem or feature request

I attempted to convert the TensorFlow model of Yahoo's Open NSFW model found here with tensorflowjs-convert. (First, I exported the TensorFlow .npy file into a Frozen Model and also a Saved Model.)

For the Frozen Model, I ran:

tensorflowjs_converter --input_format=tf_frozen_model --output_node_names='fc_nsfw/BiasAdd' --saved_model_tags=serve frozen_open_nsfw.pb web_model

and received the following message:

Unsupported Ops in the model
TensorArrayWriteV3, TensorArrayV3, DecodeBase64, TensorArrayScatterV3, TensorArrayReadV3, EncodeJpeg, Assert, All, TensorArraySizeV3, DecodeJpeg, Unpack, TensorArrayGatherV3

Running on the Saved Model also results in the same Unsupported Ops list.

Request: Please add my vote to support these Ops

• TensorArrayWriteV3
• TensorArrayV3
• DecodeBase64
• TensorArrayScatterV3
• TensorArrayReadV3
• EncodeJpeg
• Assert
• All
• TensorArraySizeV3
• DecodeJpeg
• Unpack (I noticed this was added earlier today but not yet in the latest pip installed version)
• TensorArrayGatherV3

Thanks!

[pyu10055]

The TensorArray related ops are almost ready, see tensorflow/tfjs-converter#170
All op has been added. The ops left are:
DecodedBase64, EncodeJpeg, Assert
We will add those to our priority list.

[tomslick]

Thanks! That's amazing progress!

Also, don't forget DecodeJpeg please.

[rstml]

Same issue here. The only outstanding ops are: DecodeBase64, DecodeJpeg, EncodeJpeg

[xue160709]

Same issue here. The only outstanding ops are: DecodeBase64, DecodeJpeg, EncodeJpeg

[SSaishruthi]

Working on it.

[rstml]

@SSaishruthi Frankly, if OpenNSFW is the only reason for supporting those operations, you may reconsider it.

For everyone else requesting this:

OpenNSFW has not been updated since its announcement in 2016. There are now alternative models which you can use.

We trained our alternative, MobileNetV2 based model which you can find here: https://github.com/purify-ai/geacc-models
TFJS example using this model is available here: https://github.com/purify-ai/geacc-tfjs-demo
Chrome extension which uses TFJS and our model: https://github.com/purify-ai/geacc-crx
We plan to release updated dataset and model soon (potentially 100K images).

There's also NsfwJS which is based on InceptionV3: https://github.com/infinitered/nsfwjs

[pyu10055]

DecodeJpeg op has been added in tfjs-node.