tomg008 / skynet Goto Github PK

Dear author, I follow the train code in the ismart2 repo and I can generate the weight bin file. Can you tell me how do you generate the two weight files weights_fixed.bin and weights_floating.bin here in the HLS dir?

profiled skynet further, found that each skynet call to the pl costs 0.165ms， as such 24fps；
and the first image_stitch opeartion adds a cost, and the following ones cost are very low.
Nice design!
There still exists one timing problem:
Under 214MHz, there exists -3.014ns WNS unmet timing constrains in vivado 2019.2, in 3×3 conv modules. Is it normal or the Vivado version differences?

If design focus speed performance only , what other optimization could take? Does AXI bus clock alone to speed up the dram access? Appreciate if you could give some advice, thanks.

Hello SkyNet,

I want to deploy the code on Ultra96-V2.
I started over and built the bitstream and weights from scratch

After finishing building the bitstream and weights, I changed the file name to "SkyNet.bin", "SkyNet.hwh", and "SkyNet.bit", and then ran the script "SkyNet.py" provided by the team.

But I always got the message KeyError: 'power1'

I tried to print out pynq.get_rails() and I only have these followings

{'in1': Rail {name=in1, voltage=Sensor {name=in1_voltage, value=1.193V}}, 
'in2': Rail {name=in2, voltage=Sensor {name=in2_voltage, value=1.797V}}, 
'in3': Rail {name=in3, voltage=Sensor {name=in3_voltage, value=0.844V}}, 
'in4': Rail {name=in4, voltage=Sensor {name=in4_voltage, value=0.852V}}, '
in5': Rail {name=in5, voltage=Sensor {name=in5_voltage, value=1.798V}}, 
'in6': Rail {name=in6, voltage=Sensor {name=in6_voltage, value=1.794V}}, 
'in7': Rail {name=in7, voltage=Sensor {name=in7_voltage, value=0.842V}}, 
'in8': Rail {name=in8, voltage=Sensor {name=in8_voltage, value=0.841V}}, 
'in9': Rail {name=in9, voltage=Sensor {name=in9_voltage, value=0.844V}}, 
'in10': Rail {name=in10, voltage=Sensor {name=in10_voltage, value=1.796V}}, 
'in11': Rail {name=in11, voltage=Sensor {name=in11_voltage, value=1.092V}}}

No "power1" is in pynq.get_rails().
Could you kindly suggest which part I might go wrong so that I could fix it?

Thank you.

Hi,

I tried to run the demo, but got the following error when I tried to load the Overlay.
Please help me resolve it.

Thanks,
Akshay

hi, will you release the training code ?

@TomG008 @DNN-Accelerators
Release the training code

Hello!I am a novice, I would like to ask if I can use my own dataset to do target detection training with skynet?I can't seem to find the code.
Thanks!@TomG008

Hi, I don't know how to generate weights_float.bin when I wanna to trian skynet on new dataset. I guess the skynet is trained by pytorch, whose product should be *.pth file. I appreciate that if you can share the method about how to generate weights_float.bin. Thanks a lot.

i download all the files, change the target platform from 3eg to 9eg in script file.
in 2019.2 vivado hls command line,use
vivado_hls -f script.tcl
and get the syntheis report saying that the lut costs 188% of what's available in xczu9eg.

Skynet should be self aware, and maybe Blood thirsty

I use the dac.weights and only obtained 0.46 IOU. So the dac.weights is the final weight file?
I use the train.py file of iSmart2 and modify some parameters to train SkyNet, obtain 0.63 IOU. I can not get the 0.731 IOU. Are there some tricks in the training?

I find that only 10000 images of training dataset are randomly chosen in a training epoch. Why don't use all training datasets in a training epoch?

i run like readme.md

there is an error:

Hey there, thanks for the great work.

Is it possible to detect multiple objects using your design?

Successfully uninstalled pynq-2.5.1
Successfully installed pynq-2.5.3
(base) cao@cao-OMEN-by-HP-Desktop-PC-880-p0xx:/media/cao/SkyNet-master/FPGA/Deploy$ sudo python3 SkyNet.py

**** Running SkyNet
Traceback (most recent call last):
File "SkyNet.py", line 20, in
xlnk = Xlnk()
File "/usr/local/lib/python3.8/dist-packages/pynq/xlnk.py", line 136, in init
Xlnk._open_library(Xlnk.libxlnk_path)
File "/usr/local/lib/python3.8/dist-packages/pynq/xlnk.py", line 122, in _open_library
cls.libxlnk = cls.ffi.dlopen(path)
File "/usr/local/lib/python3.8/dist-packages/cffi/api.py", line 150, in dlopen
lib, function_cache = _make_ffi_library(self, name, flags)
File "/usr/local/lib/python3.8/dist-packages/cffi/api.py", line 832, in _make_ffi_library
backendlib = _load_backend_lib(backend, libname, flags)
File "/usr/local/lib/python3.8/dist-packages/cffi/api.py", line 827, in _load_backend_lib
raise OSError(msg)
OSError: cannot load library '/usr/lib/libcma.so': /usr/lib/libcma.so: cannot open shared object file: No such file or directory. Additionally, ctypes.util.find_library() did not manage to locate a library called '/usr/lib/libcma.so'
Exception ignored in: <function Xlnk.del at 0x7f7df96fc700>
Traceback (most recent call last):
File "/usr/local/lib/python3.8/dist-packages/pynq/xlnk.py", line 150, in del
AttributeError: 'Xlnk' object has no attribute 'bufmap'

I tried to run the demo, but got the following error.Can you help me solve this problem?

Thanks!

Greetings,

I actually wanted to integrate with mmdetection so that I could train qdtrack model. Would be really great if you could provide me assistance of adding SkyNet to mmdetection repo.

Thanks

I'm running SkyNet on a Pynq-Z2, so I realize that is where this issue might come from.

I have changed the things that need to be changed in the HLS/RTL portion and got it to synthesize, implement, and generate the bitstream.

The issue is when I am deploying it in the notebook, I get to the part where we start to write.

SkyNet = overlay.SkyNet_0
SkyNet.write(0x10, img.physical_address)
SkyNet.write(0x18, conv_weight_1x1_all.physical_address)
SkyNet.write(0x20, conv_weight_3x3_all.physical_address)
SkyNet.write(0x28, bias_all.physical_address)
SkyNet.write(0x30, DDR_pool_3_out.physical_address)
SkyNet.write(0x38, DDR_pool_6_out.physical_address)
SkyNet.write(0x40, DDR_buf.physical_address)
SkyNet.write(0x50, predict_boxes.physical_address)
SkyNet.write(0x58, constant.physical_address)

The error I receive is:

AttributeError: Could not find IP or hierarchy write in overlay

Any idea?