It looks like you're new here. If you want to get involved, click one of these buttons!
Please feel free to add any comments by replying to this thread.
Api corresponding to distributed processing to multiple sticks is necessary, not cyclic processing.
Support for Tensorflow mobilenet SSD. The blog posts at movidius.github.io are quite good - additional topics by the same authors would be helpful.
Why does nobody follow the issues and pull requests on the Github code repo? You have zero resources dedicated to supporting the project that you're actively selling new hardware upgrades for? Terrible strategy.
I hope to see support for Mask R-CNN very soon.
@GoldenWings Thanks for the input. Any other suggestions or features you guys would want to see?
Would like to see support for squeezenet and Tensorflow mobilenet SSD object detection.
I would in general like to see more support for different tensorflow ops. This would enable greater freedom in design of networks which would be really cool. Also more tutorials for maybe custom Networks.
Tutorials on getting new networks not already in the ncappzoo working on the sticks. Like DSOD - Densely Supervised Object Detectors. MobileNets-SSD is slow to train on Caffe. (and will remain until they start supporting an efficient Depthwise Separable Layer, in Caffe-SSD)
@trevor you should be able to use this link for training SSD with efficient DWS convolution with CUDNN9: https://github.com/listenlink/caffe/tree/ssd
Please support Keras. I have everything setup with Keras in Ubuntu. I do not see particular reason to add Caffe in my system other than Movidius stick. Untill then example graph files that can be deployed in RPi would be appreciated.
How to use retrain.py and your own classifications, mnvcompile and use it on the stick.
Release date for Movidius X ??? ( this year !? )
It would be great if the stick will support Pytorch in the future
1) NCS kit in general is messy and loaded with Python dependencies. Straight C/C++ solutions would be best and a little python maybe for conversion / support.
2) install.sh just fails in so many ways on a vanilla clean install Ubuntu 16.04 - like it was never really tested.
3) find a way to enable a broader support community by open sourcing more tools (e.g. myriad compilers) and selecting a few outside partners to support smaller hobbyist / maker types.
4) get a partner program going - important to all of the above items
It would also be nice for better Raspberry Pi / TinkerBoard support during installation and compilation.
Internal processing FP32 compatible
@Tome_at_Intel please merge outstanding pull requests and support the Github code! I just made a PR to fix a simple syntax bug, and I now see someone made the exact same pull request on the 27th Feb, 8 MONTHS ago, and it hasn't been merged!
I'm working on face recognition using MTCNN and Sphereface, and found that
1. both v1 and v2 cannot change the input shape/ cannot get multi-output *(and when I try to concatenate the outputs.. bugs come, even worse is that the outputs can't be concatenate..)
2. v2 doesn't support MTCNN, there's always info saying that no weights provided, list index out of range when I try to compile the graph. (btw, somehow many warning messages aren't that specific, they are even misleading)
(2.1 No demo for loading multi-graph in one ncs which should be a huge achievement in v2(sorry that I don't know the fifo is bound to the graph or the device? Could you give more details about the fifo?))
(2.2 plz try to support more and more layers that is normal in tf or caffe afap..thx.. orz.. )
Multioutput Support many awesome tensorflow models has many outputs, and the only model that the documentation teach you to compile to graph is incepction v3
why not just limit your guys support to only one framework (TensorFlow)? That way you guys can consolidate your efforts and get us great support!