Has Mobilenet-SSD been supported on iOS yet ? about forge HOT 9 CLOSED

Just so people know, there is a partial port made here:
https://github.com/vonholst/SSDMobileNet_CoreML

The end of the network output has been re-implemented in swift code.

from forge.

There is an example project using TensorFlow 1.4.0 that runs the following models:
ssd_mobilenet_v1_coco, ssd_inception_v2_coco, faster_rcnn_resnet101_coco

https://github.com/JieHe96/iOS_Tensorflow_ObjectDetection_Example

There are some issues currently, I needed to recompile the protbuff headers with protoc version 3.4.0.

Get the correct 3.4.0 binary for your system here:
https://github.com/google/protobuf/releases

Get the original .proto definition file here:
https://github.com/tensorflow/models/blob/master/research/object_detection/protos/string_int_label_map.proto

Then run something like:
./protoc --cpp_out=./ ./string_int_label_map.proto

Make sure to use the protoc binary for 3.4.0 that you downloaded above. Then replace the files in the ex_SSD_Mobilenet_TF folder:
string_int_label_map.pb.cc
string_int_label_map.pb.h

I haven't yet hooked this up to video input yet but I will be trying shortly and I might commit and push some changes to the repo.

from forge.

@madhavajay Recommend trying https://github.com/tf-coreml/tf-coreml instead of. It's support using GPU on ios device so the performance would be better than.

from forge.

@madhavajay cool I have not had oppotinity to check it out yet. If you did tried it please highlight something
Thanks 👍

from forge.

I don't know if CoreML supports it. It looks your model was made using a custom version of Caffe, while coremltools only supports the original version of Caffe.

I've implemented (the TensorFlow version of) MobileNet+SSD in Metal for a client and that works fine, but it did require a few custom Metal kernels.

from forge.

So is it possible to convert my model to metal ?

from forge.

Yes, that should be possible but it might involve writing custom compute kernels.

from forge.

Yes, I just saw that and will give it a try but wanted to provide a working solution (even if its slow) that I found in case someone else finds it useful. Seems there are few and far between solutions for a lot of this stuff at the moment.

Should we all be looking at NASnet now though?
https://research.googleblog.com/2017/11/automl-for-large-scale-image.html
https://github.com/tensorflow/models/tree/master/research/slim

Anyone got a working implementation for mobile?

from forge.

Yes, it works straight from the master branch on load in xcode, just change your Developer team. 👍

from forge.