hirokinakahara/guinness

GUINNESS: A GUI-based binarized deep Neural NEtwork SyntheSizer toward an FPGA

guinness's Issues

RuntimeError: `nvcc` command returns non-zero exit status.

konと申します。
お世話になります。
手順通り環境構築し、学習開始した矢先に下記のエラーで詰まっています。

File "cupy/core/elementwise.pxi", line 552, in cupy.core.core.ElementwiseKernel.call (cupy/core/core.cpp:43113)
File "cupy/util.pyx", line 39, in cupy.util.memoize.decorator.ret (cupy/util.cpp:1480)
File "cupy/core/elementwise.pxi", line 409, in cupy.core.core._get_elementwise_kernel (cupy/core/core.cpp:41459)
File "cupy/core/elementwise.pxi", line 12, in cupy.core.core._get_simple_elementwise_kernel (cupy/core/core.cpp:34090)
File "cupy/core/elementwise.pxi", line 32, in cupy.core.core._get_simple_elementwise_kernel (cupy/core/core.cpp:33912)
File "cupy/core/carray.pxi", line 86, in cupy.core.core.compile_with_cache (cupy/core/core.cpp:33567)
File "/home/kono/.pyenv/versions/anaconda3-4.1.0/lib/python3.5/site-packages/cupy/cuda/compiler.py", line 156, in compile_with_cache
cubin = nvcc(source, options, arch)
File "/home/kono/.pyenv/versions/anaconda3-4.1.0/lib/python3.5/site-packages/cupy/cuda/compiler.py", line 78, in nvcc
_run_nvcc(cmd, root_dir)
File "/home/kono/.pyenv/versions/anaconda3-4.1.0/lib/python3.5/site-packages/cupy/cuda/compiler.py", line 56, in _run_nvcc
raise RuntimeError(msg)
RuntimeError: nvcc command returns non-zero exit status.
command: ['nvcc', '--cubin', '-arch', 'sm_61', '/tmp/tmpuo9118zz/kern.cu']
return-code: 1
stdout/stderr:
b'/tmp/tmpuo9118zz/kern.cu(14): error: identifier "__float2half_rn" is undefined\n\n/tmp/tmpuo9118zz/kern.cu(17): error: identifier "__float2half_rn" is undefined\n\n/tmp/tmpuo9118zz/kern.cu(20): error: identifier "__float2half_rn" is undefined\n\n/tmp/tmpuo9118zz/kern.cu(23): error: identifier "__float2half_rn" is undefined\n\n/tmp/tmpuo9118zz/kern.cu(26): error: identifier "__float2half_rn" is undefined\n\n/tmp/tmpuo9118zz/kern.cu(29): error: identifier "__float2half_rn" is undefined\n\n/tmp/tmpuo9118zz/kern.cu(32): error: identifier "__float2half_rn" is undefined\n\n/tmp/tmpuo9118zz/kern.cu(35): error: identifier "__half2float" is undefined\n\n8 errors detected in the compilation of "/tmp/tmpxft_00001d30_00000000-6_kern.cpp1.ii".\n'

環境は下記の通りです。
Ubuntu16.04
cuda-9.0(cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb)
Driver Version: 384.90
GeForce GTX 1060

cudaかchainer周りとは推定しているのですが、どうも上手くいきません。
ご教授頂けましたらよろしくお願いします。

How to use train data and label

Hello.
Deep Learning I am a beginner.
I made it in mnist.py of deep-learning-from-scratch,
I have mnist.pkl, but label and image are packaged together.
In GUINNESS it seems that it is necessary to make DATA and LABEL separately,
I examined various things, but it will not work.
Please tell me how to divide DATA and LABEL successfully with mnist.

A Lightweight YOLOv2 - training

Hello @HirokiNakahara,

Thank you for the wonderful work.
I am trying to train the similar network as yours: https://dl.acm.org/citation.cfm?id=3174266.
But I only want to generate the training weights (no FPGA). As this repo is not the final version source of the paper above, can you please give me some instructions for training darknet19 like network.

Best,
Deepak

Typo in guinness.py

guinness.py
from line 227 to 232
fpgaboard = QtGui.QLabel('Target FPGA Board')
self.combo2 = QtGui.QComboBox()
self.combo2.addItem("zed")
self.combo2.addItem("zybo")
self.combo2.addItem("vc702")
self.combo2.addItem("zcu102")

In line 231, "self.combo2.addItem("vc702")" should be "self.combo2.addItem("zc702")"
It is a typo.

ERROR: application performed illegal memory access and is being terminated

Hello, I had an error from the step number 6 in your tutorial. The error comes from when I input
./guinness_tut1_v1.elf (my ip) 10050 into the terminal which connected with FPGA(ZCU102)
The complier error is as below.

load weights
b0_BNFb.txt
CANNOT OPEN
ERROR: application performed illegal memory access and is being terminated

Do you have any suggestion for solving this problem?

Problems in SDx build

No module named cv2

:~/workspace/GUINNESS-master$ sudo python gen_training_data.py --pathfile list.txt --dataset class3 --size 48 --keepaspect yes
Traceback (most recent call last):
File "gen_training_data.py", line 16, in
import cv2
ImportError: No module named cv2

How to train SVR in A Lightweight YOLOv2 paper.

Hello @HirokiNakahara,

I was following your paper A Lightweight YOLOv2 . But while training SVR I am facing some difficulties. Will you kindly clarify following things.

1. How did you prepare training dataset for SVR. Did you take binary output from each cell from the last conv layer? If yes, what was the target values in the cell where non of the GT object falls.
2. In some cell, if the center of the GT box lies, how did you take x, y, dx, dy, conf and probabilities for training 55 SVR?

Thanks.

Best,
Deepak

do not work with chainer-3

This work is very interesting.
Today do not work.
Can you realign the code with chainer-3?

Bug for code generator?

In the tutorial, the output of the integer convolution called on layer 0 is called as 128.

        int_conv2d_layer<bit_64, bit_128, 64, 128, 48, 48>
        ( in_img, fb_tmp, conv0W, b0_BNFb);

It should be 64 from the weight and bias memory width.

// weight memory -----------------------------------------------------------
ap_int<3> conv0W[64][33];
ap_int<64> conv1W[128][33];
ap_int<128> conv2W[128][3*3];
ap_int<1> fc0W[3][128];

// bias memory ------------------------------------------------------------
ap_int<20> b0_BNFb[64];
ap_int<16> b1_BNFb[128];
ap_int<16> b2_BNFb[128];
ap_int<16> b3_BNFb[3];