xilinx/qnn-mo-pynq

QNN-MO-PYNQ PIP INSTALL Package

This repo contains the pip install package for Quantized Neural Network (QNN) on PYNQ using a Multi-Layer Offload (MO) architecture. Two different overlays are here included, namely W1A2 (1 bit weights, 2 bit activations) and W1A3 (1 bit weights, 3 bit activations), executing in the Programmable Logic 1 Convolutional layer and 1 (optional) Max Pool layer.

Quick Start

In order to install it on your PYNQ board, connect to the board, open a terminal and type:

# (on PYNQ v2.3 and later versions, tested up to v2.5)
sudo pip3 install git+https://github.com/Xilinx/QNN-MO-PYNQ.git

NOTE: The board has to be connected to the internet.

The installation can take up to 10 minutes, since dependencies must be resolved and sources compiled.

This will install the QNN package to your board, and create a QNN directory in the Jupyter home area. You will find the Jupyter notebooks to test the QNN overlays in this directory.

In order to build the shared object during installation, the user should copy the include folder from VIVADO HLS on the PYNQ board (in windows in vivado-path/Vivado_HLS/201x.y/include, /vivado-path/Vidado_HLS/201x.y/include in unix) and set the environment variable VIVADOHLS_INCLUDE_PATH to the location in which the folder has been copied. If the env variable is not set, the precompiled version will be used instead.

Repo organization

The repo is organized as follows:

• qnn: contains the qnn class description as well as the classes for the test networks
  • src: contains the sources of the 2 overlays and the libraries to rebuild them:
    • library: FINN library for HLS QNN-MO descriptions, host code, script to rebuilt and drivers for the PYNQ (please refer to README for more details)
    • network: overlay topologies (W1A2 and w1A3) HLS top functions, host code and make script for HW and SW built (please refer to README for more details)
  • bitstreams: with the bitstream for the 2 overlays
  • libraries: pre-compiled shared objects for low-level driver of the 2 overlays
  • params: set of trained parameters for the 2 overlays:
    • A pruned version of DoReFa-Net network, trained on the ImageNet dataset with 1 bit weights and 2 bit activations
    • A modified version of Tiny Yolo, namely Tinier-Yolo, trained on PASCAL VOC dataset with 1 bit weights and 3 bit activations
• notebooks: lists a set of python notebooks examples, that during installation will be moved in /home/xilinx/jupyter_notebooks/qnn/ folder
• tests: contains test scripts and test images

Hardware design rebuilt

In order to rebuild the hardware designs, the repo should be cloned in a machine with installation of the Vivado Design Suite (tested with 2017.4). Following the step-by-step instructions:

1. Clone the repository on your linux machine: git clone https://github.com/Xilinx/QNN-MO-PYNQ.git;
2. Move to clone_path/QNN-MO-PYNQ/qnn/src/network/
3. Set the XILINX_QNN_ROOT environment variable to clone_path/QNN-MO-PYNQ/qnn/src/
4. Launch the shell script make-hw.sh with parameters the target network, target platform and mode, with the command ./make-hw.sh {network} {platform} {mode} where:
   • network can be W1A2 or W1A3;
   • platform can be pynqz1-z2 or ultra96;
   • mode can be h to launch Vivado HLS synthesis, b to launch the Vivado project (needs HLS synthesis results), a to launch both.
5. The results will be visible in clone_path/QNN-MO-PYNQ/qnn/src/network/output/ that is organized as follows:
   • bitstream: contains the generated bitstream(s);
   • hls-syn: contains the Vivado HLS generated RTL and IP (in the subfolder named as the target network);
   • report: contains the Vivado and Vivado HLS reports;
   • vivado: contains the Vivado project.
6. Copy the generated bitstream and tcl script on the PYNQ board pip_installation_path/qnn/bitstreams/

qnn-mo-pynq's Issues

Check FPGA hardware utilization of tinyYOLO

Hi,

Your framework is really good! It works great on my PYNQ board. Thanks for your job.

I am now trying to test the hardware utilization of the tiny YOLO QNN used in the examples on my FPGA.
I believe it should be done using Vivado.
Could you please provide a guide on doing this on Windows 10? I use Vivado 2018.3.

Thanks,
Aliaksei

Build synthesis error

Hi,
I ran ./make-hw.sh W1A2 pynq a and got error bellow.
My version of vivado is 2018.2.
Can you help me?
Thanks

`Phase 1.2 IO Placement/ Clock Placement/ Build Placer Device
ERROR: [Place 30-640] Place Check : This design requires more Slice LUTs cells than are available in the target device. This design requires 60002 of such cell types but only 53200 compatible sites are available in the target device. Please analyze your synthesis results and constraints to ensure the design is mapped to Xilinx primitives as expected. If so, please consider targeting a larger device. Please set tcl parameter "drc.disableLUTOverUtilError" to 1 to change this error to warning.
ERROR: [Place 30-640] Place Check : This design requires more LUT as Memory cells than are available in the target device. This design requires 30417 of such cell types but only 17400 compatible sites are available in the target device. Please analyze your synthesis results and constraints to ensure the design is mapped to Xilinx primitives as expected. If so, please consider targeting a larger device. Please set tcl parameter "drc.disableLUTOverUtilError" to 1 to change this error to warning.
ERROR: [Place 30-640] Place Check : This design requires more LUT as Distributed RAM cells than are available in the target device. This design requires 29892 of such cell types but only 17400 compatible sites are available in the target device. Please analyze your synthesis results and constraints to ensure the design is mapped to Xilinx primitives as expected. If so, please consider targeting a larger device. Please set tcl parameter "drc.disableLUTOverUtilError" to 1 to change this error to warning.
ERROR: [Place 30-640] Place Check : This design requires more RAMS64E cells than are available in the target device. This design requires 19968 of such cell types but only 17400 compatible sites are available in the target device. Please analyze your synthesis results and constraints to ensure the design is mapped to Xilinx primitives as expected. If so, please consider targeting a larger device.
INFO: [Timing 38-35] Done setting XDC timing constraints.
Phase 1.2 IO Placement/ Clock Placement/ Build Placer Device | Checksum: 132c4f8d9

Time (s): cpu = 00:00:20 ; elapsed = 00:00:12 . Memory (MB): peak = 3554.590 ; gain = 0.000 ; free physical = 3301 ; free virtual = 25181
Phase 1 Placer Initialization | Checksum: 132c4f8d9

Time (s): cpu = 00:00:20 ; elapsed = 00:00:12 . Memory (MB): peak = 3554.590 ; gain = 0.000 ; free physical = 3300 ; free virtual = 25180
ERROR: [Place 30-99] Placer failed with error: 'Implementation Feasibility check failed, Please see the previously displayed individual error or warning messages for more details.'
Please review all ERROR, CRITICAL WARNING, and WARNING messages during placement to understand the cause for failure.
Ending Placer Task | Checksum: 132c4f8d9

Time (s): cpu = 00:00:20 ; elapsed = 00:00:12 . Memory (MB): peak = 3554.590 ; gain = 0.000 ; free physical = 3305 ; free virtual = 25185
INFO: [Common 17-83] Releasing license: Implementation
56 Infos, 21 Warnings, 0 Critical Warnings and 6 Errors encountered.
place_design failed
ERROR: [Common 17-69] Command failed: Placer could not place all instances
INFO: [Common 17-206] Exiting Vivado at Fri Aug 24 14:17:56 2018...
[Fri Aug 24 14:17:56 2018] impl_1 finished
wait_on_run: Time (s): cpu = 00:07:18 ; elapsed = 00:10:54 . Memory (MB): peak = 1732.375 ; gain = 0.000 ; free physical = 3657 ; free virtual = 25537
INFO: [Common 17-206] Exiting Vivado at Fri Aug 24 14:17:56 2018...
cp: cannot stat '/home/anhnt/QNN-MO-PYNQ/qnn/src/network/output/vivado/W1A2-pynq/W1A2-pynq.runs/impl_1/procsys_wrapper.bit': No such file or directory
cat: /home/anhnt/QNN-MO-PYNQ/qnn/src/network/output/vivado/W1A2-pynq/W1A2-pynq.runs/impl_1/procsys_wrapper_timing_summary_routed.rpt: No such file or directory
cat: /home/anhnt/QNN-MO-PYNQ/qnn/src/network/output/vivado/W1A2-pynq/W1A2-pynq.runs/impl_1/procsys_wrapper_utilization_placed.rpt: No such file or directory
cat: /home/anhnt/QNN-MO-PYNQ/qnn/src/network/output/vivado/W1A2-pynq/W1A2-pynq.runs/impl_1/procsys_wrapper_utilization_placed.rpt: No such file or directory
cat: /home/anhnt/QNN-MO-PYNQ/qnn/src/network/output/vivado/W1A2-pynq/W1A2-pynq.runs/impl_1/procsys_wrapper_utilization_placed.rpt: No such file or directory
Bitstream copied to /home/anhnt/QNN-MO-PYNQ/qnn/src/network/output/bitstream/W1A2-overlay-pynq.bit
Done!
`

how to rebuild this for new platform like zcu104

Hi, now i am working on zcu104, could you tell me that how to rebuild this for new platform ?
thank you very much ,now i can ran the PYNQ on my zcu104 board.next ,what should i do ?

communication between pynq and arduino

Hi giuliogamba,

I am sorry to bother you again, but i have some trouble.
I want my pynq to communicate with arduino because there is some data on pynq that needs to be sent to arduino, and i want to send this data through the pmod pin, but the .bit file used in this project doesnt seem to support pmod. I am a pynq beginner, and i dont know how to use vivado rebuild hardware, so i am frustrated. I have to implement pynq communicate with arduino as quickly as possible, so I would like to ask what is the method to realize the communication with arduino on the basis of the existing .bit files of this project? like spi? or uart? Can you give me some advise? Looking forward to your early reply. Thank you in advance.

Kind regards,
Icarus.

No Output When Using QNN-MO-PYNQ default installation

Hi @shinya321 @giuliogamba

Ran with no luck from PYNQ-Z2.

1. Using QNN-MO-PYNQ default installation with py.
   run /opt/darknet/python/darknet.py
   1.1
   net = load_net(b"cfg/tiny-yolo.cfg", b'./tiny-yolov1.weights',0)#b"tiny-yolo.weights", 0)
   only output [ ]

1.2
net = load_net(b"cfg/yolov2-tiny-voc.cfg", b'weight/yolov2-tiny-voc.weights',0)
Segmentation fault

1.3
net = load_net(b"cfg/yolov2.cfg", b'./yolov2.weights',0)
Segmentation fault

2. Using QNN-MO-PYNQ default installation with ./darknet
   root@pynq:/opt/darknet# ./darknet detect yolov2-tiny.cfg yolov2-tiny.weights  data/dog.jpg
   Segmentation fault
   even change to yolov2-tiny-voc or yolov3-tiny also Segmentation fault.

Fail to detect objects in the image

Hi! That's really excellent work!I have tried to detect object in your work tiny-yolo-image, however, there is nothing detected.

Would you please tell me the reason?

Using webcam with QNN

Hey! Really thanks for the tutorial for QNN. I want to use my usb webcam to capture the image in real time and then use your model to identify it. Kidnly tell me where to change the code. I'm new to fpga domain.
Thanks.

hw rebuild

My board is PYNQ-Z2. When I implement the project, something wrong happens. How could I deal with it?

The follow is errors:
[Place 30-640] Place Check : This design requires more Slice LUTs cells than are available in the target device. This design requires 62315 of such cell types but only 53200 compatible sites are available in the target device. Please analyze your synthesis results and constraints to ensure the design is mapped to Xilinx primitives as expected. If so, please consider targeting a larger device. Please set tcl parameter "drc.disableLUTOverUtilError" to 1 to change this error to warning.

training model

Hello,
Can I get the training model of dorefanet-classifier? I want to train the plant classes.

How to get verification.bin and input.bin?

i want to use another image for testbench, but i don't know how to get the bin file like verification.bin and input.bin, can anyone help me?

Synthesis error

The provided ./make-hw.sh fails to fit the implemented design on the PYNQ 7z020.

How to train customized datasheet?

I have read dorefanet-imagenet-samples.ipynb,
In the second step, we loaded a trained classifier. What I want to ask is how to make a classifier. Is there a general template for reference? For example, in tensorflow, you can load it by reading ckpt. Model, if I want to do an image classification on PYNQ, how should I operate?

What's the format of json file of the network structure definition?

In the "QNN-MO-PYNQ/qnn/src/network/sw/main.cpp"
line 66 to 68:
extern "C" { void initParameters(unsigned int const batch, unsigned int const threads); void initAccelerator(char const *networkJson, char const *layerJson);
The json file of network need to be assigned.
But is the format of this json file the same with general deeplearning framework like theano?

Darknet

When do you plan to release the customized darknet for training quantized YOLO?

Using this project to implement binary(qnn) version of tiny yolo2

Hey,
I am planning to implement QNN version of tiny yolo2. Because I want to implement tiny yolo2 on fpga and I am planning to use bnn for that. Then I got to know about this project. So I would like to have your advice and ideas about changing this to work for tiny yolo2 on pync.
It would be a great help if you could give me your ideas.

Thanks in advance !

undefined symbol: forward_region_layer_pointer_nolayer

When I run tiny-yolo-image notebook on my pynq board I got this error.

.... /darknet/libdarknet.so: undefined symbol: forward_region_layer_pointer_nolayer

There is no function names forward_region_layer_pointer_nolayer, how to fix that?

Python3 to lib_hw.so connection problem on ZynqMP (aarch64) , "python3: unhandled alignment fault (7)"

I compile this library for ZCU102 board (ZynqMP Ultrascale+) and get working app_hw.elf binary.
When trying to use library from Python3.6 getting error (python crash) with:
In console, when python crash/get back to shell:
"Bus error"
In system logs:
"
python3: unhandled alignment fault (7) at 0x7f78695080, esr 0x92000061, in libc-2.27.so[7f8bed6000+140000]
CPU: 0 PID: 5065 Comm: python3 Not tainted 4.14.0-xilinx-v2018.2 #1
Hardware name: ZynqMP ZCU102 Rev1.0 (DT)
"

How to get the boxes, classes and probs, it seems only plot out

Hi, giuliogamba
After ran the tiny-yolo-image.iynb, only get the plotted image with bbox and a probs file.
How to get each boxes information like (meta.names[i], probs[j][i], (boxes[j].x, boxes[j].y, boxes[j].w, boxes[j].h) as /opt/darknet/python/darknet.py.

Thanks.

How to train on custom dataset?

Hi,

Could you please give us directions to train this to our own data set?
Or is it possible to train this network on original yolo implementation and covert the .weights files to the required format? Any script?

Open HLS project

Thanks for your work .
I'm trying to open the HLS project of this work by
: vivado_hls -f hls-syn.tcl
but I don't know what exactly I should type for argv defined in the .tcl
such as config_hwsrcdir, config_proj_part, config_clkperiod ...
Could you please give me some guide ? I will be very greatful.
Sorry that I'm a beginer and Vivado or HLS.

Kernel dies when trying to run tiny-yolo-image-loop

Hi,

I have a pynq board with latest (2.1) SD card image into it. I have downloaded this (https://github.com/Xilinx/QNN-MO-PYNQ) repository and trying to run "tiny-yolo-image-loop" jupyter nottebook. Sometimes it runs perfectly but many times kernel dies in-between.
Is there a way to debug or analyse the issue.

Hardware Rebuild Utilization exceeds 100%

Hello.
I am trying to rebuild hardware.
I notice that hardware utilization reports exceed 100% for some device parts.
For example, for the W1A2 ultra96:
== Utilization Estimates

• Summary:
  +-----------------+---------+-------+--------+-------+-----+
  | Name | BRAM_18K| DSP48E| FF | LUT | URAM|
  +-----------------+---------+-------+--------+-------+-----+
  |DSP | -| -| -| -| -|
  |Expression | -| -| 0| 47| -|
  |FIFO | -| -| -| -| -|
  |Instance | 150| 67| 26404| 72210| -|
  |Memory | 256| -| 4096| 384| -|
  |Multiplexer | -| -| -| 2445| -|
  |Register | -| -| 511| -| -|
  +-----------------+---------+-------+--------+-------+-----+
  |Total | 406| 67| 31011| 75086| 0|
  +-----------------+---------+-------+--------+-------+-----+
  |Available | 432| 360| 141120| 70560| 0|
  +-----------------+---------+-------+--------+-------+-----+
  |Utilization (%) | 93| 18| 21| 106| 0|
  +-----------------+---------+-------+--------+-------+-----+

LUT 106%, does this matter? I am using 2017.4 on ubuntu.
Actually I also tried on 2018.3 on another machine and exceeds as well.
What might be the reason for these? Are there any additional settings I need to make?
Thank you!

Is your network trained with swapped X / Y axis ??

Hello,
Analyzing your provided Python code I came to :
Image loaded array axes (shape) (3,h,w) = [Color, Y, X] , it's ok
then going img_copy = np.swapaxes(img_copy, 0,2) with result [X,Y,C] <- Not Ok for most image operations
then in utils.conv_layer transpose(0, 3, 1, 2) on [1,X,Y,C] with result [1,C,X,Y] <- here it not ok for Darknet too..
And then you get elements for convolution like from correctly positioned image ...
After buffer prepare, same data (with incorrect convolution direction is passed to Accelerator) ..

Could you comment on that ??

Also I tried loading "dog.jpg" image whith OpenCV , it give me better result.
But OpenCV have different colour space BGR (not RGB) , did you use RGB for training ?

Not able to run tiny-yolo-image-loop.ipynb

I was not able to run tiny-yolo-image-loop.ipynb out of the box, when trying to run the code in the box after "2. Launch Demo", I got this error

NameError                                 Traceback (most recent call last)`
<ipython-input-3-282cc81fa050> in <module>()
     59             print("class: {}\tprobability: {}".format(det[1], det[0]))
     60         res = Image.open(file_name_out.value.decode() + ".png")
---> 61         display(res)

NameError: name 'display' is not defined

I was able to fix this with adding the following line to the imports

from IPython.display import display

lib_sw_W1A3 .so

hi,
i have a problem about the files of lib_sw_W1A3.so and lib_sw_W1A3.so. i want to know the function of these two documents. Thank you！

Hello!I'm beginner in FPGA . And I just bought a PYNQ-Z1 borad but feel a little confused.

I want to achieve a real-time face recognition system on PYNQ with a camera and HDMI output. Could I use this network to train my own parameters? Or could you please give some suggestion to me? Thank you very much!

model

Excuse me
The original tiny-yolo model filter number is 16-32-64-128-256-512-1024-1024
Why is the filter number used here 16-64-64-128-256-512-512?
And more accurate than the tiny-yolo model

Error when i install the QNN package to my board

When i use
sudo pip3 install git+https://github.com/Xilinx/QNN-MO-PYNQ.git
to install the QNN package to my board PYNQ Z2, the following error exists:

The directory '/home/xilinx/.cache/pip/http' or its parent directory is not owned by the current user and the cache has been disabled. Please check the permissions and owner of that directory. If executi.
The directory '/home/xilinx/.cache/pip' or its parent directory is not owned by the current user and caching wheels has been disabled. check the permissions and owner of that directory. If executing pip .
Collecting git+https://github.com/Xilinx/QNN-MO-PYNQ.git
Cloning https://github.com/Xilinx/QNN-MO-PYNQ.git to /tmp/pip-cz7ywm9o-build
Installing collected packages: qnn-loopback
Running setup.py install for qnn-loopback ... error
Complete output from command /opt/python3.6/bin/python3.6 -u -c "import setuptools, tokenize;file='/tmp/pip-cz7ywm9o-build/setup.py';f=getattr(tokenize, 'open', open)(file);code=f.read().repl:
Running pre installation script...
Building hardware library...
Cloning into '/tmp/pip-cz7ywm9o-build/qnn/src/library/rapidjson'...
Note: checking out 'tags/v1.1.0'.

You are in 'detached HEAD' state. You can look around, make experimental
changes and commit them, and you can discard any commits you make in this
state without impacting any branches by performing another checkout.

If you want to create a new branch to retain commits you create, you may
do so (now or later) by using -b with the checkout command again. Example:

  git checkout -b <new-branch-name>

HEAD is now at f54b0e4... Merge pull request #714 from miloyip/version1.1.0
/usr/bin/ld: cannot find -lcma
collect2: error: ld returned 1 exit status
make[1]: *** [output/lib_hw.so] Error 1
make: *** [lib_hw] Error 2
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/tmp/pip-cz7ywm9o-build/setup.py", line 56, in <module>
    subprocess.check_output(["make", "-j2", "-C", my_path + "qnn/src/network/", "lib_hw"])
  File "/opt/python3.6/lib/python3.6/subprocess.py", line 336, in check_output
    **kwargs).stdout
  File "/opt/python3.6/lib/python3.6/subprocess.py", line 418, in run
    output=stdout, stderr=stderr)
subprocess.CalledProcessError: Command '['make', '-j2', '-C', '/tmp/pip-cz7ywm9o-build/qnn/src/network/', 'lib_hw']' returned non-zero exit status 2.

----------------------------------------

Command "/opt/python3.6/bin/python3.6 -u -c "import setuptools, tokenize;file='/tmp/pip-cz7ywm9o-build/setup.py';f=getattr(tokenize, 'open', open)(file);code=f.read().replace('\r\n', '\n');f.clos/

HW rebuild for other plaform

Hi, I am trying to rebuild this system for other platform, Could I change the device in make-hw.sh script to other device and redo the synthesis?

Tinier YOLO processing flow

Hi,

I am wondering if this Tinier YOLO design have the same flow with Tincy YOLO shown in Xilinx website?
https://forums.xilinx.com/t5/Xcell-Daily-Blog/Tincy-YOLO-a-real-time-low-latency-low-power-object-detection/ba-p/815840

ZCU104 Support

Is there a estimated time frame for zcu104 support?

Regards
Jack

Training on Custom Tinier-YOLO Dataset

I've created a custom dataset by following the original YOLO guidelines here:
https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

It is unclear to me how to train on the Tinier Yolo architecture to produce the necessary .npy files. Is this possible at all? If so what are some good resources to help me proceed?

Thanks!

Question on the output activation

Hello,

I don't understand why for the network W1A3, the output activation is 3-bit wide, the accumulated value (stored in the macRegisters) is compared with 5 (=3*2-1) thresholds read from the threshold memory. Can you explain me this case?

Many thanks!

test

This is a very powerful experiment.
Excuse me! in vivado hls , what should I do if I want to detect object for my model ?

Try to get the boxs location

I tried to modify detector. c to get the boxs location. After the modification, I deleted all . o files under obj and libdarknet. o file , and entered "make" in the terminal to try to recompile. But an error occurred under the adapter notebook like that:

OSError Traceback (most recent call last)
in ()
16 from qnn import utils
17 sys.path.append("/opt/darknet/python/")
---> 18 from darknet import *
19
20 from matplotlib import pyplot as plt

/opt/darknet/python/darknet.py in ()
37
38 #lib = CDLL("/home/pjreddie/documents/darknet/libdarknet.so", RTLD_GLOBAL)
---> 39 lib = CDLL("/opt/darknet/libdarknet.so", RTLD_GLOBAL)
40 lib.network_width.argtypes = [c_void_p]
41 lib.network_width.restype = c_int

/usr/lib/python3.6/ctypes/init.py in init(self, name, mode, handle, use_errno, use_last_error)
346
347 if handle is None:
--> 348 self._handle = _dlopen(self._name, mode)
349 else:
350 self._handle = handle

OSError: /opt/darknet/libdarknet.so: file too short
Can you help me? Thank you very much.

Implementation of real-time object detecion

Hi, giuliogamba,

Really thanks for your QNN tutorial ,recently, i want to achieve real-time object detection by QNN with pynq, i changed the tiny-yolo-image-loop.ipynq code, capture the image through usb webcam in the loop because i want to achieve real-time detection in this way, but my frame rate is only 1/4 FPS, that means it takes 4 seconds to load an image, it is too slow, no real-time, so i want to ask how can i improve my frame rate? I am new to pynq domain, hope get some advise from you.

Hope to hearing from you soon, and thanks in advance!
Icarus.

If I want to know the location of detection boxes, what should i do

C synthesis error

hi，
When i rebuild the HLS project,there are some errors: default template arguments for a function template are a C++11 extension [-Wc++11-extensions] and error: use of undeclared identifier 'constexpr'.
So i need your help. Thank you very much.

platform ultra96 is v1 or v2?

the supported platform ultra96 is v1 or v2? thanks

About the Tinier YOLO

Hi,
Currently, I want to train a new weights for the Tinier YOLO Topology. So I trained a network with the same topology as the network you provided in the example. I tried to match it with the work you done by using the BNN weight generator:

1. I try to using the Keras/Tensorflow to train a network and convert it to the form of the Keras/Theano form.
2. Using the BNNWeightReader to read and generate weights.
   But something still unclear for me:
3. The topology and the jupyter notebook showed that your network has 9 layer from 0 to 8. But in the params folder of Tiny Yolo, there're 10 set of weights from 1 to 10. Did you remove the layer 0 or just shift it from 0-9 to 1-10
4. As I read from other questions, your way to calculate the peCounts and simdCounts map have not been not public yet. Could you please provide me the peCounts and simdCounts map which you use for this topology? Because I try a lot of value and nothing can look alike like the original weights size.
5. If I want to add another Conv Layer, whether add it to the end of tinier-yolo-layers.json and put the binparams to the folder binparam-tinier-yolo-nopool is enough?
   Hope to hearing from you soon,
   Thanks,

HW rebuild on windows

Hello,

I am trying to use Vivado HLS 2017.4 command prompt on Windows 10 to rebuild the script but when I followed the instructions, it gave the following error when entering the following line "./make-hw.sh W1A2 pynqz1-z2 h" .

C:\Users\Sayf\Desktop\QNN\QNN-MO-PYNQ-master\QNN-MO-PYNQ-master\qnn\src\network>./make-hw.sh W1A2 pynqz1-z2 h
'.' is not recognized as an internal or external command,
operable program or batch file.

Kind regards,
Sayf Khan

pip3 install problem

My board is pynq-z2.
And, I tried to pip3 install git+https://github.com/Xilinx/QNN-MO-PYNQ.git,
but there was no any response after cloning.

Collecting git+https://github.com/Xilinx/QNN-MO-PYNQ.git
Cloning https://github.com/Xilinx/QNN-MO-PYNQ.git to /tmp/pip-bl5011_8-build

I had waited for about 20min.
By the way, I could ping github.com.

Question on "stride" of sliding window module

Hi,
Very impressive work! Thank you for sharing this project to us.
I have some questions regarding the behaviour of "stride" parameter in the sliding window.
Is it true that assigning 0 to "stride" parameter will result in a striding of 1 during convolution?
If so, does this module support a striding of 2? And what value should we assign to this stride parameter?
Best,
awai54st

NameError: name 'lib' is not defined

Hello,

I try tiny-yolo-image-loop.ipynb, then it has the problem as shown below

import sys
import os, platform
import json
import numpy as np
import cv2
from ctypes import c_char_p # i add import
from matplotlib import pyplot as plt
from PIL import Image
from datetime import datetime
from qnn import TinierYolo
from qnn import utils
sys.path.append("/opt/darknet/python/")
from darknet import *
from tests.test_qnn import test_tinier_yolo # i add import

%matplotlib inline
import IPython

1. Instantiate a Classifier

---

conv0_weights = np.load('/home/xilinx/jupyter_notebooks/QNN-MO-PYNQ/qnn/params/tinier-yolo-conv0-W.npy', encoding="latin1")
conv0_weights_correct = np.transpose(conv0_weights, axes=(3, 2, 1, 0))
conv8_weights = np.load('/home/xilinx/jupyter_notebooks/QNN-MO-PYNQ/qnn/params/tinier-yolo-conv8-W.npy', encoding="latin1")
conv8_weights_correct = np.transpose(conv8_weights, axes=(3, 2, 1, 0))
conv0_bias = np.load('/home/xilinx/jupyter_notebooks/QNN-MO-PYNQ/qnn/params/tinier-yolo-conv0-bias.npy', encoding="latin1")
conv0_bias_broadcast = np.broadcast_to(conv0_bias[:,np.newaxis], (net['conv1']['input'][0],net['conv1']['input'][1]net['conv1']['input'][1]))
conv8_bias = np.load('/home/xilinx/jupyter_notebooks/QNN-MO-PYNQ/qnn/params/tinier-yolo-conv8-bias.npy', encoding="latin1")
conv8_bias_broadcast = np.broadcast_to(conv8_bias[:,np.newaxis], (125,1313))

cfg = c_char_p("/home/xilinx/jupyter_notebooks/QNN-MO-PYNQ/qnn/params/tinier-yolo-bwn-3bit-relu-nomaxpool.cfg".encode())

net_darknet = lib.parse_network_cfg(file_name_cfg)

NameError Traceback (most recent call last)
in ()
14 cfg = c_char_p("/home/xilinx/jupyter_notebooks/QNN-MO-PYNQ/qnn/params/tinier-yolo-bwn-3bit-relu-nomaxpool.cfg".encode())
15
---> 16 net_darknet = lib.parse_network_cfg(file_name_cfg)

NameError: name 'lib' is not defined

How can I solve this problem? Thank you.

cannot find -lsds_lib

Hi
I use the pynq z2 with img v2.3. When I input the ''sudo pip3.6 install git+https://github.com/Xilinx/QNN-MO-PYNQ.git", the error is '/usr/bin/arm-linux-gnueabihf-ld : cannot find -lsds_lib'. Please help me, it troubled me for three days.

Hardware config vs accuracy

Hello,

Is it possible to change the parameter(s) in config.h file without affecting the inference result? As an example, I tried changing MAX_PE from 64 to 32 to generate a smaller accelerator (so that I could fit other IP blocks on the PYNQ). When I ran the Python test code, the inference could not accurately detect object in a given picture anymore. I thought modifying MAX_PE only affects the resource consumption and performance, yet I expected that the computational result would not change. Does changing configuration (in this case, MAX_PE) require re-training the network?

Install without the Internet?

Hi,

I don't have the Internet connection on my PYNQ board, but I can access it locally.
Is there a way to install the framework without the Internet?

Thanks.
Aliaksei

Weight download

hi~
If I want to download your model weights, how should I do it?

how does the function "draw_detection_python" work?

Hi,

I want to send the coordinates of the identified object to my arduino, so I have to understand how the recognition box is drawn, which is done by the function "draw_detection_python", When I want to understand it through the help command, I cannot learn more about it. so could you please help me?

kind regards,
Icarus.

If I want to know the coordinates of Bounding-box (four points), what should I do

This is what the author wrote here:(in the project of tiny-yolo-image.ipynb)

6. Draw detection boxes using Darknet
   The image postprocessing (drawing the bounding boxes) is performed in darknet using python bindings.

Then he used this function (lib.draw_detection_python()) to draw the bounding-box,if I want to know the location of the box. What should I do?