Mobilenet /Inception v3

[idata]

Hi there

I used to attend the challenge but somehow miss the submit date as in travel quite often. Here is my repo can support to train image classify based on tensorflow r1.4.0 from scratch, support mobilenet v1, inception v3 so far.

After training finished , run the freeze script and the pb file can be used for compile graph need for Movidus usb stick, the average inference for mobilenet about 30-40 ms , for the inception v3 , takes about 300-400 ms.

https://github.com/chesterkuo/imageclassify-movidius

[idata]

@chester Sorry that you missed the submission date for the competition. Thanks for dropping by and sharing your work with the community.

[idata]

Feel free to use, also pull request.

I'd like to show people how they can benifet tensorflow framework + Movidius who plan to use TF here , as this is easy and act like one click solution without any big changes.

[idata]

Thanks for this, had been trying to get a model made with 1.3.1 as the compatible RPI version for Stretch only goes that far as far as I know, 1.3.1 does not support Slim (as far as I know). Ditching the RPI and NUC for now and just going to use the Movidius with 1.4.0 on a desktop. Hopefully my previous code will work now :)

[idata]

Hi @AdamMiltonBarker

The TF Slim had been supported for long time , as checking from github results, slim be included into r1.3 also.

https://github.com/tensorflow/tensorflow/tree/r1.3/tensorflow/contrib/slim

You may try this if needs, but my case is based on r1.4 by using pip with GPU enabled.

[idata]

Hi @chester thanks I did try for a couple of days and had all sorts of issues, possibly could of been due to the NUC. I had found a discussion on Github that said slim was not compatible, I may have mixed up convos though. Your code helped me piece some pieces together so thanks for that. Training on a GPU (Only 750 ti but going to try on my 1050 ti on Windows) with TF 1.4.0 and was able to get it working. I forked your project and fixed it for Python 3 and made it all run off one shell script including compiling with NCS and reloading the model, it also has IoT connectivity, I can upload it to my fork and submit a pull request if you like? I so far have made a facial recognition project and also a breast cancer detector out of it, breast cancer detector is doing really well so far :)

[idata]

I also found a new github repo with all the versions of TF for RPI, not tested them yet but will share the link when I find it in my bookmarks.

[idata]

@AdamMiltonBarker , feel free to create pull request for python3 or others, i will merge .

As for TF RPi, here is link i used to download, search for "rpi"

http://ci.tensorflow.org/view/All/

[idata]

Also , i had inception-resnet-v2 network with Tensorflow , can be applied also.

However this network need lots of time to train but it have very good accuracy on top-5 image classification.

[idata]

@chester Thanks, these are the ones I found:

https://github.com/lhelontra/tensorflow-on-arm/releases

And here is the first tutorial, will add it and the facial recognition one as soon as possible:

https://github.com/AdamMiltonBarker/BoobAIs/tree/master/V1/Python/Tensorflow/Inception-V3

[idata]

@chester Thanks, these are the ones I found:

https://github.com/lhelontra/tensorflow-on-arm/releases

And here is the first tutorial, will add it and the facial recognition one as soon as possible:

https://github.com/AdamMiltonBarker/BoobAIs/tree/master/V1/Python/Tensorflow/Inception-V3

[idata]

@AdamMiltonBarker

That's cool.

I'm working on case to see if i can add spot word wake engine into Movidius , as all of those are implemented with CNN.

This had been verfied on RP3 with good performance.

I do a network change to apply inception block in order to increase accuracy but still keep model as small as possible (5MB total) and inference time about 75ms on RP3 platform

https://github.com/chesterkuo/kaggle-speech-challenge-1