cnn_vhdl_generator - PLEASE CITE THE PAPER BELOW IF YOU USE THIS WORK
AUTOMATIC VHDL GENERATION FOR CNN MODELS
"Tool Name : Automatic VHDL Generation for Small Scale Convolutional Neural Networks"
- " Description: This tool generates VHDL code for hand-tuned and benchmarked models like LeNet"
- " System requirements:"
- " Max needed RAM : 2GB "
- " OS : Any operating system supports Java 8 or higher"
- " Java version : 8 or higher x64 distribution"
This version of the tool offers the following
- " This version supports the following:"
- " Layers : Up to 16 "
- " Feature-maps : Up to 1024 "
- " Filter size : Up to 10 "
- " Classes : Up to 200 "
- " Pooling : Max pooling "
- " Activation : ReLU "
- " Classifier : SoftMax"
Preparing the data:
1- Weights and input data must be formatted in a binary representation with the desried precision.
2- You need to specify the whole bits and decimal bits in your data by evaluating your fixed-point representaiton.
3- The output will be in the same representation you provided as the input (signed binary).
4- Example of formatting data:
- Assume the maximum number among classes value is 2.9423 and minimum number is -0.00324.
- The maximum and minimum numbers can be represented by 8bit fixed-point precision.
- For most accurate results, 3 bits will be used to represent the whole number and 5 bits for the decimal part.
- After converting the maximum number to binary this is what we get: 010.11110
- Now the . will no be included, but you'll just need to know what you've specified for precision to be able to parse the results.
- You can use any way to convert floats to fixed-point binary, but you'll need to convert the results back to floats in the same way.Get Started with the Generation Tool
1- In the main (first) window, select network type and target platform then click on "Customize and build model"
2- In the Configuration Block do the following:
* You can manually build your network, where you can specify all the network parameters from the drop lists in the table
* You can load a pre-configured model by clicking on "load configuration" button
* Once you are done with configuration, click on "Validate Model Entries" to check the validity of your model
* You can save your manually configured model by clicking on "Save Configuration" button
* Click on "load weights and biases" button to load weights and biases
3- In Weights and Biases Incorpration window, do the following:
* Choose representation, weight size, bias size, and whole(digits) and decimal point (fraction) for fixed-point representation. If weight precision is 8
then you could specify 3 whole bits and 5 fractional ones to represent a floating number as a fixed-point. Input (images) decimals must match weight decimals.
* Click on "Load Weights and Biases" button to load weights and biases,
you may use the sample random weights provided to you for the provided configuration file for testing purposes
* The tool will automatically check and verify the loaded file to see if it matches the configuration file as well as chosen
weight and bias sizes and representation
* If loaded file passes the validation and checking test, the "Generate VHDL files" button will become enabled automatically
and then Generate Testbench button will get enabled automatically after generating the VHDL files.
* Generated files can be found in the same directoy in a folder called "generated_files".Toturial
Publication
- IEEE VHDL Generator for a High Performance Convolutional Neural Network FPGA-base Accelerator
- RG. VHDL Generator for a High Performance Convolutional Neural Network FPGA-base Accelerator
Thesis
PLEASE CITE THIS PAPER IF YOU USE THIS WORK
@inproceedings{hamdan2017vhdl,
title={Vhdl generator for a high performance convolutional neural network fpga-based accelerator},
author={Hamdan, Muhammad K and Rover, Diane T},
booktitle={2017 International Conference on ReConFigurable Computing and FPGAs (ReConFig)},
pages={1--6},
year={2017},
organization={IEEE}
}