- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I finished the work to translante Mobilenetv2 to NCS stick,
links here:
https://github.com/xufeifeiWHU/Mobilenet-v2-on-Movidius-stick
~~~~
Hope it can help you something.
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@zufeifei
Good Job!!!
I will try it immediately!!!
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@PINTO
OH !
you will find it run no faster than v1 ,
and the accurancy may improve a little,
So, the better choice may be using mobilenet v1 instead,
mobilenet v2 only run faster than v1 on mobile phone,
and movidius stick is more likely similiar to a GPU,
I would try to translante tiny-yolo-v3 next,
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@zufeifei
you will find it run no faster than v1 ,
I know that.
I'd like to increase the variety of options I have on hand.
I would try to translante tiny-yolo-v3 next,
Great!
I am looking forward.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
oh I do not know that,
I had thought it could run faster than moibilenet v1 before,
so sad.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@zufeifei
I was looking at the comparative article below.
Performance seems to be reversed by width multiplier.
https://qiita.com/HiromuMasuda0228/items/d1de49b0620b1a3f0317#object-detection
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@PINTO
OH!
Thanks a lot for your helpful link,
I would translate and read it conscientiously。
~~~~~~~~
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Did you train the weights using Caffe or import them across from TF? If you imported them from the TF model can you show how you did that?
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@madhavajay
I used caffe ,
and I get the .caffemodel and the .prototxt files here,
https://github.com/shicai/MobileNet-Caffe
I thought tensorflow work the same as caffe,
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@zufeifei The reason why I ask is the same github user who created the Caffe SSD MobileNetv1 used in the NCSDK App Zoo has this:
https://github.com/chuanqi305/MobileNetv2-SSDLite
It contains code to take the weights directly from the TensorFlow model and write them into a caffe model.
I will do some experimentation but people have mentioned anecdotally that the same model in Caffe under performs.
I have done training in Caffe with this model on COCO for the suggested 120000 epochs with fine-tuning on the existing mobilenet_iter_73000.caffemodel file.
https://github.com/chuanqi305/MobileNet-SSD
The results were far worse than the Stock TF MobileNet SSD v1 when run through TensorFlow.
And got substantially worse results than the default TF MobileNet SSD
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@madhavajay
I also try to train SSD-Mobilenet on mu own dataset with the scripts from https://github.com/chuanqi305/MobileNet-SSD
but it seems performs well,
And you mean, mobilenet v2 may perform better in tensorflow than caffe?
if so, I should have a try in tensorflow to test it
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@zufeifei No im not saying its about V1 vs V2 models, im saying that the same architecture when implemented in Caffe (which doesnt support all the identical ops) and trained in Caffe (again which uses different optimisers and default settings), seems to under perform. I don't know if its a result of the Training process or some other issue, I need to investigate closer but because I know I can train good results in TF it would be nicer to know I can simply convert to NCS from a good source rather than messing around in Caffe hoping to replicate those results.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page