Hi, I am planning to develop object detection (preferably SSD/mobilenet) on snap dragon 820 or 835 using neural processing engine sdk. I would like to know the following:
1. which 820 or 835 kit should i buy? what is the difference?
2. can neural processing engine sdk support mobilenet? I am planning to use caffe2 framework. I am concerned with seprable/depthwise convolution like 1x1, 3x1, 1x3 and 3x3 with grouping = input channels. (I know some chip vendors does not support such convolution)
3. is there support for ssd? i only read that faster-rcnn is supported in NPE. What are the consideration if i want to implement SSD myself? one concern is NMS (non max suppresion), besides that is there any other things that i should take care of?
Thank you very much!
We had mixed results. CPU-only version of our model running on 835 was noticeably slower with snpe-1.2.2 compared to 820. GPU-only version ran slightly faster on 835.
I am greatly disappointed that Qualcomm hasn't released the 64-bit version of SNPE SDK for Android.
Hi KingCobra,
May I ask that, do you run SSD/mobilenet with SNPE SDK? If does, then you should convert model file to a dlc file, have you met errors when converting model file to a dlc file, and what about the solutions?
I met the error list below:
.......
File "/home/damon/work/snpe-sdk/snpe-1.6.0/lib/python/converters/tensorflow/converter.py", line 500, in _create_sorted_node_input_graph
if input_node_name not in nodes_map or input_node_name in ordered_nodes_map:
RuntimeError: maximum recursion depth exceeded in cmp
Do you meet the same issue before? Any comments appreciated!
Thanks.
Hi, KingCobra
Currently snpe1.6 support MobileNet-SSD priorbox? Have you already put the MobileNet-SSD model into a .dlc file? Can tell me how to convert it, or send me a successful conversion. Dlc format file.
Look at the hyperlink below:
https://developer.qualcomm.com/forum/qdn-forums/software/snapdragon-neur...
Hi, DamonZhou
Currently snpe1.6 support MobileNet-SSD priorbox? Have you already put the MobileNet-SSD model into a .dlc file?
PriorBox layers are not supported. https://developer.qualcomm.com/forum/qdn-forums/software/snapdragon-neur...
Hi,
Did you manage to run mobilenet-ssd? If so, could you let me know how? If not, is there any other objection detection network that works and is supported? I am after object detection / localisation.
Thanks,
Martin Peniak
Hi qinglong,
Unfortunitely I failed to convert mobilenet-ssd, now I am working on convert tiny-yolo, but the convert tool always reports add/mul/max are not consumed, still have to try it.
Hi,dear
Now, does snpe support ssd mobilenet? And can you offer me the paper about the ssd mobilenet testing on hexagon682/685? For example, the user guide about snapdragon820/835(with snpe doc).
Could you be kind enough to send me the two papers to my email [email protected]?
Tks!