Will Detectron2go be released / implemented?

[ArtanisCV]

❓ Will Detectron2go be released / implemented

In the official blog, a library named Detectron2go is introduced. It seems that this library is very helpful for model deployment.

By now, I cannot find related codes in the Detectron2 repo. Will Detectron2 provide similar features in the near future?

[wat3rBro]

@ArtanisCV Detectron2Go will also be open-sourced soon, stay tuned!

[choasup]

When will Detectron2go release in plan?

[wat3rBro]

@choasup we're targeting in November.

[cpoptic]

Hi Yuxin, any update on the release date for Detectron2Go? Very excited for this

[wat3rBro]

@cpoptic we'll release the model exporting very soon and then efficient models for mobile after that, hopefully within next 1-2 weeks.

[lucasjinreal]

@wat3rBro Exporting model to which format? Is it onnx?

[no-1ne]

@jinfagang i beleive caffe2( via onnx ) and mobile models would be with shufflenet or fbnet backbone.

@wat3rBro So many possibilities, looking forward

[MentalGear]

@cpoptic we'll release the model exporting very soon and then efficient models for mobile after that, hopefully within next 1-2 weeks.

That be awesome. Currently deciding between Tensorflow Object Detection API and Detectron2. If mobile inference was possible on Dectectron, my decision would be made.

[wat3rBro]

@jinfagang, @startupgurukul is correct, with FBNet backbone.

[no-1ne]

Can you pls elaborate if export will be targeting pytorch mobile and caffe2 or only caffe2.

Thanks in advance.

On Wed, Dec 18, 2019, 6:46 AM wat3rBro notifications@github.com wrote:

@jinfagang https://github.com/jinfagang, @startupgurukul https://github.com/startupgurukul is correct, with FBNet backbone.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/detectron2/issues/12?email_source=notifications&email_token=ACUEFOKMFP6TKIYKBZXUXEDQZF2X7A5CNFSM4I7YR7J2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEHEQODQ#issuecomment-566822670, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACUEFOKS6EVTKYFL7NU6RRLQZF2X7ANCNFSM4I7YR7JQ .

[wat3rBro]

@startupgurukul only caffe2 and "caffe2-style" torch script for now.

[no-1ne]

Thanks for the response. Waiting to check if it's convertible to tensorflow pb model via this, https://github.com/onnx/onnx-tensorflow

If that can happen, detectron2 can work on browsers along with iOS and Android via tensorflow.js. (onnx.js seems unmaintained)

Considering the way Detectron2 is making image models accessible to everyone, JS compatibility would be super awesome 🤞 and in line with the vision of detectron2.

On Fri, Dec 20, 2019, 8:13 AM wat3rBro notifications@github.com wrote:

@startupgurukul https://github.com/startupgurukul only caffe2 and "caffe2-style" torch script for now.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/detectron2/issues/12?email_source=notifications&email_token=ACUEFOJPQK4MV66OC2R7BQTQZQWN5A5CNFSM4I7YR7J2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEHLWFJQ#issuecomment-567763622, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACUEFOICILFB4RXGLL3AG3DQZQWN5ANCNFSM4I7YR7JQ .

[ppwwyyxx]

@startupgurukul

There is already a (very limited) tool that converts some standard detectron2 models to tensorflow, as mentioned in https://detectron2.readthedocs.io/notes/compatibility.html#compatibility-with-tensorflow

It does not use onnx. I believe most detectron2 models cannot successfully convert to tensorflow through onnx.

[no-1ne]

It is just awesome, going to the extent of supporting other frameworks, just increases the respect towards Detectron Team and can only imagine this being to Portal to visual apps of the future.

Coming to getting to run in a browser, while converting to tf.js format got a complain that ValueError: Unsupported Ops in the model before optimization, MirrorPad, InvertPermutation

By setting --skip_op_check flag, it was bypassed and conversion succeded, yet to try the converted model, will keep you posted with the updates

Looking forward to come back with the news that detectron2 can also infer in the browser as well.

[mathpopo]

@wat3rBro

@startupgurukul only caffe2 and "caffe2-style" torch script for now.

hello,i have tried to use maskrcnn2go (caffe2 style),very great job,when pytorch version come ?

[no-1ne]

My guess is pytorch mobile is still in active development and is yet to get operator compatibility with pytorch, but that shouldn't delay detectron2 team from moving to production so for now they must have picked caffe2 export,

As of now it's now very kind of them to open source it, the and continue adding more like export and help anyone working on visual models quickly research and move to production super quickly also while providing mobile friendly privacy preserving models(current ETA is Jan 2020 as per @wat3rBro )

On Wed, Dec 25, 2019, 12:07 PM mathpopo notifications@github.com wrote:

@startupgurukul https://github.com/startupgurukul only caffe2 and "caffe2-style" torch script for now.

hello,i try to use maskrcnn2go (caffe2 style),when pyotrch version come ?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/detectron2/issues/12?email_source=notifications&email_token=ACUEFOJOXHDIWKLXMVBXEYDQ2L5TVA5CNFSM4I7YR7J2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEHT7LJA#issuecomment-568849828, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACUEFOKI747Z47AXSWVLSBLQ2L5TVANCNFSM4I7YR7JQ .

[shuxp]

any news about detectron2go...

[zhiqwang]

It seems that the newest PR is about detectron2go?

[lucasjinreal]

So, one year after does there any updates on such detectron2go?

[dragen1860]

Detectron2go still has a long way to go!

[kewin1807]

when will Detectron2go be released??? I am very expected it.

[ppwwyyxx]

Exporting and running a model in caffe2 is already open source for a while. See https://detectron2.readthedocs.io/tutorials/deployment.html

The only other feature we currently plan to open source from detectron2go is efficient FBNet-based models, tracked in #696.

[shuxp]

Exporting and running a model in caffe2 is already open source for a while. See https://detectron2.readthedocs.io/tutorials/deployment.html

The only other feature we currently plan to open source from detectron2go is efficient FBNet-based models, tracked in #696.

Is there a demo of how to use model.pb and model_init.pb? After pytorch merged caffe2, it is not wasy to find a caffe2 C++ example...

[lucasjinreal]

Does this pb model is essentially onnx model? Can it port to other framwork through this model such as TensorRT?