├── 技术报告.pdf ├── cifar-10-torch ├── figures │ └── code.png ├── cifar-source │ ├── res │ │ └── distilled_lenet5_best.pt │ ├── save_params.py │ ├── test.py │ ├── train.py │ ├── distill.py │ └── models.py └── README.md ├── quantification ├── distilled_lenet5_best.pt ├── quantification_img.py ├── input_pic2_lable8.txt └── quantification_para.py ├── CNN-FPGA-Vivado ├── CNN-FPGA-Vivado.srcs │ ├── sim_1 │ │ └── new │ │ │ └── lenet_TB.v │ └── sources_1 │ │ ├── new │ │ ├── Lenet.v │ │ ├── FindMax.v │ │ ├── MaxPoolMulti.v │ │ ├── UsingTheRelu16.v │ │ ├── max.v │ │ ├── UsingTheRelu.v │ │ ├── MaxPoolSingle.v │ │ └── IEEE162IEEE32.v │ │ └── imports │ │ └── Integration first part │ │ ├── layer.v │ │ ├── convUnit.v │ │ ├── floatAdd.v │ │ ├── floatMult.v │ │ ├── floatAdd16.v │ │ ├── floatMult16.v │ │ ├── weightMemory.v │ │ ├── convLayerMulti.v │ │ ├── convLayerSingle.v │ │ ├── processingElement.v │ │ ├── activationFunction.v │ │ ├── processingElement16.v │ │ ├── RFselector.v │ │ ├── softmax_tb.v │ │ ├── IntegrationConvPart.v │ │ └── ANNfull.v ├── README.md └── CNN-FPGA-Vivado.xpr ├── weight ├── layer1.txt ├── layer2.txt └── classifier.txt ├── LICENSE └── README.md /技术报告.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/技术报告.pdf -------------------------------------------------------------------------------- /cifar-10-torch/figures/code.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/cifar-10-torch/figures/code.png -------------------------------------------------------------------------------- /quantification/distilled_lenet5_best.pt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/quantification/distilled_lenet5_best.pt -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sim_1/new/lenet_TB.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sim_1/new/lenet_TB.v -------------------------------------------------------------------------------- /cifar-10-torch/cifar-source/res/distilled_lenet5_best.pt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/cifar-10-torch/cifar-source/res/distilled_lenet5_best.pt -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/Lenet.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/Lenet.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/FindMax.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/FindMax.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/MaxPoolMulti.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/MaxPoolMulti.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/UsingTheRelu16.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/UsingTheRelu16.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/layer.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/layer.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/convUnit.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/convUnit.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatAdd.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatAdd.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatMult.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatMult.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatAdd16.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatAdd16.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatMult16.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/floatMult16.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/weightMemory.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/weightMemory.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/convLayerMulti.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/convLayerMulti.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/convLayerSingle.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/convLayerSingle.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/processingElement.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/processingElement.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/activationFunction.v: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Robin-WZQ/CNN-FPGA/HEAD/CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/activationFunction.v -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/README.md: -------------------------------------------------------------------------------- 1 | # CNN-FPGA-Vivado 2 | 3 | ### 运行方法 4 | 5 | - 打开项目 6 | 7 | 打开CNN-FPGA-Vivado.xpr 8 | 9 | - 修改路径 10 | 11 | 1. leNet_TB.v 12 | 修改33行，41行，将路径改为权重所在绝对路径 13 | 14 | 2. ANNfull.v 15 | 修改34行，44行，54行，将路径改为权重所在绝对路径 16 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/max.v: -------------------------------------------------------------------------------- 1 | module max (n1,n2,n3,n4,max); 2 | 3 | parameter DATA_WIDTH = 16; 4 | 5 | input [DATA_WIDTH-1:0]n1; 6 | input [DATA_WIDTH-1:0]n2; 7 | input [DATA_WIDTH-1:0]n3; 8 | input [DATA_WIDTH-1:0]n4; 9 | output [DATA_WIDTH-1:0]max; 10 | wire [DATA_WIDTH-1:0]max12; 11 | wire [DATA_WIDTH-1:0]max34; 12 | 13 | assign max12=(n1>n2)?n1:n2; 14 | assign max34=(n3>n4)?n3:n4; 15 | assign max = (max12>=max34) ? max12 : max34; 16 | endmodule -------------------------------------------------------------------------------- /cifar-10-torch/cifar-source/save_params.py: -------------------------------------------------------------------------------- 1 | from models import Lenet5, NormalCNN 2 | import torch 3 | import numpy as np 4 | 5 | np.set_printoptions(threshold=np.inf, linewidth=100) 6 | 7 | model = Lenet5() 8 | model.load_state_dict(torch.load('./res/distilled_lenet5_best.pt')) 9 | 10 | f = open('lenet5_parameters_cifar.txt', 'w', encoding='utf-8') 11 | for name, params in model.named_parameters(): 12 | print(name, params.shape) 13 | f.write(name + ':' + str(params.shape) + '\n') 14 | f.write(str(params.detach().numpy()) + '\n') 15 | -------------------------------------------------------------------------------- /weight/layer1.txt: -------------------------------------------------------------------------------- 1 | 2461aafe2b65a524a71ca0932b6ba816a91d28ccb17935442a19b6592ff8b0d13882b4aab605358faeb536bdb14fb6883304 2 | 2c45246b2dccb066ad3db0ffb2a326b0300a2f3dab85b449af982e1333f8a4c7b2ecb4762b57333c309caf89b138b06f306c 3 | 32b82ce6afd7b172a95a2565b5b9b6282c432e45b543b52630e532eb26302257346334f3321ba5cf2d662e692d55ab92b11e 4 | 2d93b267aab62d6927a7aed93116ac9da5ad221231d4b1c4a8b2ab8e99d833383552ac3aafacb36eb431332335d22cf3aefc 5 | 2f62a89431da2b6d3460250cae9731072cf330d1b337b0f7ab3ab419b104b16d34ed14a733113403144b3194b02a2ec7b230 6 | ac952ee22cf62ed528c5a5b7b33c345dafb136042b19b3e8b5afb5e8b51c2ff332982d3a356fb1b0aed5301fb0d333f5a459 7 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/UsingTheRelu.v: -------------------------------------------------------------------------------- 1 | `timescale 1ns / 1ps 2 | ////////////////////////////////////////////////////////////////////////////////// 3 | // Company: 4 | // Engineer: 5 | // 6 | // Create Date: 2022/08/29 16:30:03 7 | // Design Name: 8 | // Module Name: UsingTheRelu 9 | // Project Name: 10 | // Target Devices: 11 | // Tool Versions: 12 | // Description: 13 | // 14 | // Dependencies: 15 | // 16 | // Revision: 17 | // Revision 0.01 - File Created 18 | // Additional Comments: 19 | // 20 | ////////////////////////////////////////////////////////////////////////////////// 21 | 22 | 23 | module UsingTheRelu( 24 | 25 | ); 26 | endmodule 27 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/processingElement16.v: -------------------------------------------------------------------------------- 1 | `timescale 100 ns / 10 ps 2 | 3 | module processingElement16(clk,reset,floatA,floatB,result); 4 | 5 | parameter DATA_WIDTH = 16; 6 | 7 | input clk, reset; 8 | input [DATA_WIDTH-1:0] floatA, floatB; 9 | output reg [DATA_WIDTH-1:0] result; 10 | 11 | wire [DATA_WIDTH-1:0] multResult; 12 | wire [DATA_WIDTH-1:0] addResult; 13 | 14 | floatMult16 FM (floatA,floatB,multResult); 15 | floatAdd16 FADD (multResult,result,addResult); 16 | 17 | always @ (posedge clk or posedge reset) begin 18 | if (reset == 1'b1) begin 19 | result = 0; 20 | end else begin 21 | result = addResult; 22 | end 23 | end 24 | 25 | endmodule 26 | -------------------------------------------------------------------------------- /cifar-10-torch/README.md: -------------------------------------------------------------------------------- 1 | # 文件说明 2 | 3 | **cifar-source：**源代码文件夹 4 | 5 | - cifar-10-python：数据集文件夹 6 | - res：模型存储文件夹 7 | - distilled_lenet5_best.pt：蒸馏得到的Lenet5模型参数文件 8 | - train.py：Teacher model训练代码 9 | - distill.py：模型蒸馏训练代码 10 | - test.py：验证评估代码 11 | - save_params.py：参数导出代码 12 | - models.py：模型结构代码 13 | - Lenet5_parameters_cifar.txt：导出模型参数文件 14 | 15 | # 实验环境 16 | 17 | Python==3.9 18 | 19 | torch==1.9.0+cu111 20 | 21 | torchvision==0.10.0+cu111 22 | 23 | numpy==1.20.3 24 | 25 | # 代码说明 26 |
27 | 28 |
29 | 30 | # 运行说明 31 | 32 | 以cifar-source目录为源根 33 | 34 | 训练Teacher model 35 | 36 | ``` 37 | python train.py 38 | ``` 39 | 40 | 模型蒸馏 41 | 42 | ``` 43 | python distill.py 44 | ``` 45 | 46 | 参数导出 47 | 48 | ``` 49 | python save_params.py 50 | ``` 51 | 52 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/MaxPoolSingle.v: -------------------------------------------------------------------------------- 1 | `timescale 1 ns / 10 ps 2 | 3 | module MaxPoolSingle(aPoolIn,aPoolOut); 4 | 5 | parameter DATA_WIDTH = 16; 6 | parameter InputH = 28; 7 | parameter InputW = 28; 8 | parameter Depth = 1; 9 | 10 | input [0:InputH*InputW*Depth*DATA_WIDTH-1] aPoolIn; 11 | output [0:(InputH/2)*(InputW/2)*Depth*DATA_WIDTH-1] aPoolOut; 12 | 13 | genvar i,j; 14 | 15 | generate 16 | for (i=0; i<(InputH); i=i+2) begin 17 | for (j=0; j<(InputW); j=j+2) begin 18 | max 19 | #( 20 | .DATA_WIDTH(DATA_WIDTH) 21 | ) max1 22 | ( 23 | .n1(aPoolIn[(i*InputH+j)*DATA_WIDTH+:DATA_WIDTH]), 24 | .n2(aPoolIn[(i*InputH+j+1)*DATA_WIDTH+:DATA_WIDTH]), 25 | .n3(aPoolIn[((i+1)*InputH+j)*DATA_WIDTH+:DATA_WIDTH]), 26 | .n4(aPoolIn[((i+1)*InputH+j+1)*DATA_WIDTH+:DATA_WIDTH]), 27 | .max(aPoolOut[(i/2*InputH/2+j/2)*DATA_WIDTH+:DATA_WIDTH]) 28 | ); 29 | end 30 | end 31 | endgenerate 32 | 33 | endmodule 34 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/new/IEEE162IEEE32.v: -------------------------------------------------------------------------------- 1 | module IEEE162IEEE32(clk,reset,input_fc,output_fc); 2 | 3 | parameter DATA_WIDTH_1 = 16; 4 | parameter DATA_WIDTH_2 = 32; 5 | parameter NODES = 400; 6 | 7 | input clk, reset; 8 | input [DATA_WIDTH_1*NODES-1:0] input_fc; 9 | output reg [DATA_WIDTH_2*NODES-1:0] output_fc; 10 | 11 | reg [7:0] temp; 12 | integer i; 13 | 14 | always @ (negedge clk or posedge reset) begin 15 | if (reset == 1'b1) begin 16 | output_fc = 0; 17 | end else begin 18 | temp = 8'b00000000; 19 | for (i = 0; i < NODES; i = i + 1) begin 20 | output_fc[DATA_WIDTH_2*(i+1)-1] = input_fc[DATA_WIDTH_1*(i+1)-1]; 21 | temp[0+:5] = input_fc[(DATA_WIDTH_1*i+10)+:5]; 22 | output_fc[(DATA_WIDTH_2*i+23)+:8] = temp + 8'b01110000; 23 | output_fc[(DATA_WIDTH_2*i+13)+:10] = input_fc[(DATA_WIDTH_1*i+0)+:10]; 24 | output_fc[(DATA_WIDTH_2*i)+:13] = 0; 25 | end 26 | 27 | end 28 | end 29 | 30 | endmodule 31 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2022 Robin-WZQ 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/RFselector.v: -------------------------------------------------------------------------------- 1 | `timescale 100 ns / 10 ps 2 | 3 | module RFselector(image,rowNumber, column,receptiveField); 4 | 5 | parameter DATA_WIDTH = 16; 6 | parameter D = 1; //Depth of the filter 7 | parameter H = 32; //Height of the image 8 | parameter W = 32; //Width of the image 9 | parameter F = 5; //Size of the filter 10 | 11 | input [0:D*H*W*DATA_WIDTH-1] image; 12 | input [5:0] rowNumber, column; 13 | output reg [0:(((W-F+1)/2)*D*F*F*DATA_WIDTH)-1] receptiveField; 14 | 15 | integer address, c, k, i; 16 | 17 | always @ (image or rowNumber or column) begin 18 | address = 0; 19 | if (column == 0) begin 20 | for (c = 0; c < (W-F+1)/2; c = c + 1) begin 21 | for (k = 0; k < D; k = k + 1) begin 22 | for (i = 0; i < F; i = i + 1) begin 23 | receptiveField[address*F*DATA_WIDTH+:F*DATA_WIDTH] = image[rowNumber*W*DATA_WIDTH+c*DATA_WIDTH+k*H*W*DATA_WIDTH+i*W*DATA_WIDTH+:F*DATA_WIDTH]; 24 | address = address + 1; 25 | end 26 | end 27 | end 28 | end else begin 29 | for (c = (W-F+1)/2; c < (W-F+1); c = c + 1) begin 30 | for (k = 0; k < D; k = k + 1) begin 31 | for (i = 0; i < F; i = i + 1) begin 32 | receptiveField[address*F*DATA_WIDTH+:F*DATA_WIDTH] = image[rowNumber*W*DATA_WIDTH+c*DATA_WIDTH+k*H*W*DATA_WIDTH+i*W*DATA_WIDTH+:F*DATA_WIDTH]; 33 | address = address + 1; 34 | end 35 | end 36 | end 37 | end 38 | 39 | end 40 | 41 | endmodule 42 | 43 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/softmax_tb.v: -------------------------------------------------------------------------------- 1 | `timescale 1ps/1ps 2 | module softmax_tb(); 3 | localparam DATA_WIDTH=32; 4 | localparam inputNum=10; 5 | reg [DATA_WIDTH*inputNum-1:0] inputs; 6 | reg clk; 7 | reg enable; 8 | wire [DATA_WIDTH*inputNum-1:0] outputs; 9 | wire ackSoft; 10 | softmax1 #(.DATA_WIDTH(DATA_WIDTH)) soft(inputs,clk,enable,outputs,ackSoft); 11 | 12 | localparam PERIOD = 100; 13 | integer count; 14 | always 15 | #(PERIOD/2) clk = ~clk; 16 | 17 | initial begin 18 | clk=1'b1; 19 | inputs=320'b00111110010011001100110011001101_10111110010011001100110011001101_00111111100110011001100110011010_00111111101001100110011001100110_10111111011001100110011001100110_00111110100110011001100110011010_01000000010001100110011001100110_10111100101000111101011100001010_00111111100011100001010001111011_00111110101001010110000001000010; 20 | //inputs are 0.2 -0.2 1.2 1.3 -0.9 0.3 3.1 -0.02 1.11 0.323 21 | count=1; 22 | enable=1'b0; 23 | #(PERIOD); 24 | enable=1'b1; 25 | 26 | while(ackSoft!=1'b1) begin 27 | count=count+1; 28 | #(PERIOD); 29 | end 30 | //outputs are 0.03255, 0.02182, 0.08847, 0.09776, 0.0108, 0.0359, 0.5687, 0.02612, 0.0808, 0.03681 31 | 32 | inputs=320'b00111111001100001010001111010111_10111110010011001100110011001101_00111111100110011001100110011010_00111111101001100110011001100110_10111111011001100110011001100110_00111110100110011001100110011010_01000000010001100110011001100110_10111100101000111101011100001010_00111111100011100001010001111011_00111110101001010110000001000010; 33 | //inputs are 0.69 -0.2 1.2 1.3 -0.9 0.3 3.1 -0.02 1.11 0.323 34 | count=1; 35 | enable=1'b0; 36 | #(PERIOD); 37 | enable=1'b1; 38 | while(ackSoft!=1'b1) begin 39 | count=count+1; 40 | #(PERIOD); 41 | 42 | end 43 | //outputs are 0.05207118 0.0213835 0.0866926 0.09579553 0.01062096 0.03525543 0.5572659 0.0256007 0.07923851 0.0360757 44 | 45 | end 46 | 47 | endmodule 48 | -------------------------------------------------------------------------------- /cifar-10-torch/cifar-source/test.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch.utils.data import DataLoader 3 | from models import NormalCNN, Lenet5 4 | from torchvision.models import resnet152 5 | from torchvision.datasets import CIFAR10 6 | from torchvision.transforms import transforms 7 | import torch.nn as nn 8 | 9 | device = 'cuda:0' if torch.cuda.is_available() else 'cpu' 10 | 11 | 12 | def evaluate(model, data_loader): 13 | model.eval() 14 | pred_tags = [] 15 | true_tags = [] 16 | with torch.no_grad(): 17 | for batch in data_loader: 18 | batch_data = batch['img'].to(device) 19 | batch_label = batch['label'].to(device) 20 | 21 | logits = model(batch_data) 22 | 23 | pred = torch.argmax(logits, dim=1).cpu().numpy() 24 | tags = batch_label.cpu().numpy() 25 | 26 | pred_tags.extend(pred) 27 | true_tags.extend(tags) 28 | 29 | assert len(pred_tags) == len(true_tags) 30 | correct_num = sum(int(x == y) for (x, y) in zip(pred_tags, true_tags)) 31 | accuracy = correct_num / len(pred_tags) 32 | 33 | return accuracy 34 | 35 | 36 | if __name__ == '__main__': 37 | transform = transforms.Compose([ 38 | transforms.Grayscale(), 39 | transforms.ToTensor(), 40 | transforms.Normalize((0.4734), (0.2507)) 41 | ]) 42 | test_dataloader = DataLoader(CIFAR10(root='./cifar-10-python/', train=False, transform=transform), 43 | batch_size=128) 44 | model = Lenet5() 45 | # model.conv1 = nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) 46 | # model.fc = nn.Linear(2048, 10) 47 | model.to(device) 48 | model.load_state_dict(torch.load('./res/distilled_lenet5_best.pt')) 49 | 50 | model.eval() 51 | pred_tags = [] 52 | true_tags = [] 53 | with torch.no_grad(): 54 | for batch in test_dataloader: 55 | batch_data = batch[0].to(device) 56 | batch_label = batch[1].to(device) 57 | 58 | logits = model(batch_data) 59 | 60 | pred = torch.argmax(logits, dim=1).cpu().numpy() 61 | tags = batch_label.cpu().numpy() 62 | 63 | pred_tags.extend(pred) 64 | true_tags.extend(tags) 65 | 66 | assert len(pred_tags) == len(true_tags) 67 | correct_num = sum(int(x == y) for (x, y) in zip(pred_tags, true_tags)) 68 | accuracy = correct_num / len(pred_tags) 69 | 70 | print(accuracy * 100) 71 | -------------------------------------------------------------------------------- /quantification/quantification_img.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import matplotlib.pyplot as plt 3 | from torchvision import transforms 4 | import torch 5 | 6 | def dec2bin(x): 7 | x -= int(x) 8 | bins = [] 9 | while x: 10 | x *= 2 11 | bins.append("1" if x>=1. else "0") 12 | x -= int(x) 13 | return "".join(bins) 14 | 15 | def float2IEEE16(x): 16 | ms = "0" if x > 0 else "1" 17 | x = abs(x) 18 | x0 = int(x) # 整数部分 19 | x1 = x - x0 # 小数部分 20 | x0 = bin(x0).replace("0b", "") 21 | x1 = dec2bin(x1) 22 | if x0[0] == "0": 23 | E = 15 - x1.find("1") - 1 24 | m = x1[x1.find("1"):] 25 | if E < 0: 26 | E = 15 27 | m = "00000000000" 28 | else: 29 | E = 15 + len(x0) - 1 30 | m = x0 + x1 31 | E = bin(E).replace("0b", "") 32 | if len(E)>5: 33 | E = E[:5] 34 | else: 35 | for i in range(5 - len(E)): 36 | E = "0" + E 37 | m = m[1:] 38 | if len(m)>10: 39 | m = m[:10] 40 | else: 41 | for i in range(10 - len(m)): 42 | m += "0" 43 | y = ms+E+m 44 | y1 = "" 45 | for i in range(len(y)//4): 46 | y1 += hex(int(y[4*i:4*(i+1)], 2)).replace("0x", "") 47 | return y1 48 | 49 | # print(halfpre2spre('0011110000000000')) # 0.1 50 | # print(uint82float16(255)) #0101101111111000/5bf8 51 | 52 | def unpickle(file): 53 | import pickle 54 | with open(file, 'rb') as fo: 55 | dict = pickle.load(fo, encoding='bytes') 56 | return dict 57 | 58 | if __name__ == '__main__': 59 | transform = transforms.Compose([ 60 | transforms.ToPILImage(), 61 | transforms.Grayscale(), 62 | transforms.ToTensor(), 63 | transforms.Normalize((0.4734), (0.2507)) 64 | ]) 65 | 66 | filename = "./cifar-10-python/cifar-10-batches-py/test_batch" 67 | dataset = unpickle(filename) 68 | data = dataset[b"data"] 69 | labels = dataset[b"labels"] 70 | 71 | id = 2 72 | 73 | one_data = np.zeros((3, 1024), dtype=np.float32) 74 | one_data[0], one_data[1], one_data[2] = data[id][:1024], data[id][1024:2048], data[id][2048:] 75 | one_data = one_data.reshape((3, 32, 32)) 76 | one_data = torch.tensor(one_data, dtype=torch.float32) 77 | one_data = transform(one_data) 78 | 79 | one_label = labels[id] 80 | print(f"图像的标签为 {one_label}") 81 | 82 | result = "" 83 | 84 | for i in range(1): 85 | for j in range(32): 86 | for k in range(32): 87 | result += float2IEEE16(one_data[i][j][k]) 88 | 89 | f = open(f"input_pic{id}_lable{one_label}.txt", 'w', encoding="utf-8") 90 | f.write(result) 91 | 92 | img = np.reshape(data[id], (3, 32, 32)) 93 | plt.imshow(img.transpose((1, 2, 0))) 94 | plt.show() -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/IntegrationConvPart.v: -------------------------------------------------------------------------------- 1 | module integrationConv (clk,reset,CNNinput,Conv1F,Conv2F,iConvOutput); 2 | 3 | parameter DATA_WIDTH = 16; 4 | parameter ImgInW = 32; 5 | parameter ImgInH = 32; 6 | parameter Conv1Out = 28; 7 | parameter MvgP1out = 14; 8 | parameter Conv2Out = 10; 9 | parameter Kernel = 5; 10 | parameter MvgP2out = 5; 11 | parameter DepthC1 = 6; 12 | parameter DepthC2 = 16; 13 | 14 | integer counter; 15 | 16 | input clk, reset; 17 | input [ImgInW*ImgInH*DATA_WIDTH-1:0] CNNinput; 18 | input [Kernel*Kernel*DepthC1*DATA_WIDTH-1:0] Conv1F; 19 | input [DepthC2*Kernel*Kernel*DepthC1*DATA_WIDTH-1:0] Conv2F; 20 | output [MvgP2out*MvgP2out*DepthC2*DATA_WIDTH-1:0] iConvOutput; 21 | 22 | reg C1rst,C2rst,MP1rst,MP2rst,Relu1Reset,Relu2Reset,enRelu; 23 | //wire Tanh1Flag,Tanh2Flag,Tanh3Flag; 24 | 25 | wire [Conv1Out*Conv1Out*DepthC1*DATA_WIDTH-1:0] C1out; 26 | wire [Conv1Out*Conv1Out*DepthC1*DATA_WIDTH-1:0] C1outRelu; 27 | 28 | wire [MvgP1out*MvgP1out*DepthC1*DATA_WIDTH-1:0] MP1out; 29 | 30 | wire [Conv2Out*Conv2Out*DepthC2*DATA_WIDTH-1:0] C2out; 31 | wire [Conv2Out*Conv2Out*DepthC2*DATA_WIDTH-1:0] C2outRelu; 32 | 33 | wire [MvgP2out*MvgP2out*DepthC2*DATA_WIDTH-1:0] MP2out; 34 | 35 | convLayerMulti C1 36 | ( 37 | .clk(clk), 38 | .reset(reset), 39 | .image(CNNinput), 40 | .filters(Conv1F), 41 | .outputConv(C1out) 42 | ); 43 | 44 | UsingTheRelu16 45 | #(.OUTPUT_NODES(Conv1Out*Conv1Out*DepthC1)) 46 | relu_1 47 | ( 48 | .clk(clk), 49 | .reset(Relu1Reset), 50 | .en(enRelu), 51 | .input_fc(C1out), 52 | .output_fc(C1outRelu) 53 | ); 54 | 55 | MaxPoolMulti MP1 56 | ( 57 | .clk(clk), 58 | .reset(MP1rst), 59 | .apInput(C1outRelu), 60 | .apOutput(MP1out) 61 | ); 62 | 63 | convLayerMulti 64 | #( 65 | .DATA_WIDTH(16), 66 | .D(6), 67 | .H(14), 68 | .W(14), 69 | .F(5), 70 | .K(16) 71 | ) C2 72 | ( 73 | .clk(clk), 74 | .reset(C2rst), 75 | .image(MP1out), 76 | .filters(Conv2F), 77 | .outputConv(C2out) 78 | ); 79 | 80 | UsingTheRelu16 81 | #(.OUTPUT_NODES(Conv2Out*Conv2Out*DepthC2)) 82 | relu_2 83 | ( 84 | .clk(clk), 85 | .reset(Relu2Reset), 86 | .en(enRelu), 87 | .input_fc(C2out), 88 | .output_fc(C2outRelu) 89 | ); 90 | 91 | MaxPoolMulti 92 | #( 93 | .D(16), 94 | .H(10), 95 | .W(10) 96 | ) MP2 97 | ( 98 | .clk(clk), 99 | .reset(MP2rst), 100 | .apInput(C2outRelu), 101 | .apOutput(iConvOutput) 102 | ); 103 | 104 | always @(posedge clk or posedge reset) begin 105 | if (reset == 1'b1) begin 106 | C1rst = 1'b1; 107 | C2rst = 1'b1; 108 | MP1rst = 1'b1; 109 | MP2rst = 1'b1; 110 | Relu1Reset = 1'b1; 111 | Relu2Reset = 1'b1; 112 | enRelu = 1'b1; 113 | counter = 0; 114 | end 115 | else begin 116 | counter = counter + 1; 117 | if (counter > 0 && counter < 7*1457) begin 118 | C1rst = 1'b0; 119 | end 120 | else if (counter > 7*1457 && counter < 7*1457+6*784*6) begin 121 | Relu1Reset = 1'b0; 122 | end 123 | else if (counter > 7*1457+6*784*6 && counter < 7*1457+6*784*6+8) begin 124 | MP1rst = 1'b0; 125 | end 126 | else if (counter > 7*1457+6*784*6+8 && counter < 7*1457+6*784*6+8+18*22*152) begin 127 | C2rst = 1'b0; 128 | end 129 | else if (counter > 7*1457+6*784*6+8+18*22*152 && counter < 7*1457+6*784*6+8+18*22*152 + 6*1600) begin 130 | Relu2Reset = 1'b0; 131 | end 132 | else if (counter > 7*1457+6*784*6+8+18*22*152 + 6*1600 && counter < 7*1457+6*784*6+8+18*22*152 + 6*1600 + 20) begin 133 | MP2rst = 1'b0; 134 | end 135 | end 136 | end 137 | 138 | endmodule -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.srcs/sources_1/imports/Integration first part/ANNfull.v: -------------------------------------------------------------------------------- 1 | module ANNfull(clk,reset,input_ANN,output_ANN); 2 | 3 | parameter DATA_WIDTH = 32; 4 | parameter INPUT_NODES_L1 = 400; 5 | parameter INPUT_NODES_L2 = 120; 6 | parameter INPUT_NODES_L3 = 84; 7 | parameter OUTPUT_NODES = 10; 8 | 9 | input clk, reset; 10 | input [DATA_WIDTH*INPUT_NODES_L1-1:0] input_ANN; 11 | output [3:0] output_ANN; 12 | 13 | reg rstLayer; 14 | reg rstRelu; 15 | reg enRelu; 16 | 17 | reg [8:0] address; 18 | 19 | wire [DATA_WIDTH*INPUT_NODES_L2-1:0] output_L1; 20 | wire [DATA_WIDTH*INPUT_NODES_L2-1:0] output_L1_relu; 21 | 22 | wire [DATA_WIDTH*INPUT_NODES_L3-1:0] output_L2; 23 | wire [DATA_WIDTH*INPUT_NODES_L3-1:0] output_L2_relu; 24 | 25 | wire [DATA_WIDTH*OUTPUT_NODES-1:0] output_L3; 26 | 27 | wire [DATA_WIDTH*INPUT_NODES_L2-1:0] WL1; 28 | wire [DATA_WIDTH*INPUT_NODES_L3-1:0] WL2; 29 | wire [DATA_WIDTH*OUTPUT_NODES-1:0] WL3; 30 | 31 | weightMemory 32 | #(.INPUT_NODES(INPUT_NODES_L1), 33 | .OUTPUT_NODES(INPUT_NODES_L2), 34 | .file("C:/Users/24857/Desktop/try/para/fc1.txt")) 35 | W1( 36 | .clk(clk), 37 | .address(address), 38 | .weights(WL1) 39 | ); 40 | 41 | weightMemory 42 | #(.INPUT_NODES(INPUT_NODES_L2), 43 | .OUTPUT_NODES(INPUT_NODES_L3), 44 | .file("C:/Users/24857/Desktop/try/para/fc2.txt")) 45 | W2( 46 | .clk(clk), 47 | .address(address), 48 | .weights(WL2) 49 | ); 50 | 51 | weightMemory 52 | #(.INPUT_NODES(INPUT_NODES_L3), 53 | .OUTPUT_NODES(OUTPUT_NODES), 54 | .file("C:/Users/24857/Desktop/try/para/classifier.txt")) 55 | W3( 56 | .clk(clk), 57 | .address(address), 58 | .weights(WL3) 59 | ); 60 | 61 | layer 62 | #(.INPUT_NODES(INPUT_NODES_L1), 63 | .OUTPUT_NODES(INPUT_NODES_L2)) 64 | L1( 65 | .clk(clk), 66 | .reset(rstLayer), 67 | .input_fc(input_ANN), 68 | .weights(WL1), 69 | .output_fc(output_L1) 70 | ); 71 | 72 | activationFunction #(.OUTPUT_NODES(INPUT_NODES_L2)) relu_1 73 | ( 74 | .clk(clk), 75 | .reset(rstRelu), 76 | .en(enRelu), 77 | .input_fc(output_L1), 78 | .output_fc(output_L1_relu) 79 | ); 80 | 81 | layer 82 | #(.INPUT_NODES(INPUT_NODES_L2), 83 | .OUTPUT_NODES(INPUT_NODES_L3)) 84 | L2( 85 | .clk(clk), 86 | .reset(rstLayer), 87 | .input_fc(output_L1_relu), 88 | .weights(WL2), 89 | .output_fc(output_L2) 90 | ); 91 | 92 | activationFunction #(.OUTPUT_NODES(INPUT_NODES_L3)) relu_2 93 | ( 94 | .clk(clk), 95 | .reset(rstRelu), 96 | .en(enRelu), 97 | .input_fc(output_L2), 98 | .output_fc(output_L2_relu) 99 | ); 100 | 101 | layer 102 | #(.INPUT_NODES(INPUT_NODES_L3), 103 | .OUTPUT_NODES(OUTPUT_NODES)) 104 | L3( 105 | .clk(clk), 106 | .reset(rstLayer), 107 | .input_fc(output_L2_relu), 108 | .weights(WL3), 109 | .output_fc(output_L3) 110 | ); 111 | 112 | FindMax findmax1 113 | ( 114 | .n(output_L3), 115 | .max(output_ANN) 116 | ); 117 | 118 | 119 | always @(posedge clk or posedge reset) begin 120 | if (reset == 1'b1) begin 121 | rstRelu = 1'b1; 122 | rstLayer = 1'b1; 123 | address = -1; 124 | enRelu = 1'b0; 125 | end 126 | else begin 127 | rstRelu = 1'b0; 128 | rstLayer = 1'b0; 129 | if (address == INPUT_NODES_L1+1) begin 130 | address = address + 1; 131 | enRelu = 1'b1; 132 | end 133 | else if (address == INPUT_NODES_L1+2) begin 134 | address = -1; 135 | enRelu = 1'b0; 136 | rstLayer = 1'b1; 137 | end 138 | else begin 139 | address = address + 1; 140 | end 141 | end 142 | end 143 | 144 | endmodule -------------------------------------------------------------------------------- /cifar-10-torch/cifar-source/train.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import numpy as np 4 | from torch.utils.data import DataLoader 5 | from models import Lenet5, NormalCNN 6 | from torchvision import transforms 7 | from torchvision.models import vgg19_bn, resnet152 8 | from torchvision.datasets import CIFAR10 9 | from torch.optim import AdamW, Adam, SGD 10 | import os 11 | 12 | device = 'cuda:0' if torch.cuda.is_available() else 'cpu' 13 | 14 | 15 | def train(EPOCHS, lr, batch_size): 16 | transform_train = transforms.Compose([ 17 | transforms.Grayscale(), 18 | transforms.Resize((224, 224)), 19 | transforms.RandomHorizontalFlip(0.5), 20 | transforms.ToTensor(), 21 | transforms.Normalize((0.449), (0.226)) 22 | ]) 23 | transform_test = transforms.Compose([ 24 | transforms.Grayscale(), 25 | transforms.Resize((224, 224)), 26 | transforms.ToTensor(), 27 | transforms.Normalize((0.449), (0.226)) 28 | ]) 29 | 30 | train_dataloader = DataLoader(CIFAR10(root='./cifar-10-python/', train=True, transform=transform_train), shuffle=True, batch_size=batch_size, num_workers=4) 31 | test_dataloader = DataLoader(CIFAR10(root='./cifar-10-python/', train=False, transform=transform_test), batch_size=batch_size, num_workers=4) 32 | 33 | model = resnet152(pretrained=True) 34 | model.conv1 = nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) 35 | model.fc = nn.Linear(2048, 10) 36 | # model = Lenet5() 37 | model.to(device) 38 | optimizer = Adam(model.parameters(), lr=lr) 39 | loss_fn = nn.CrossEntropyLoss() 40 | 41 | max_test_acc = 0. 42 | for epoch in range(EPOCHS): 43 | model.train() 44 | for step, batch in enumerate(train_dataloader): 45 | batch_train = batch[0].to(device) 46 | batch_label = batch[1].to(device) 47 | 48 | model.zero_grad() 49 | logits = model(batch_train) 50 | loss = loss_fn(logits, batch_label) 51 | 52 | loss.backward() 53 | # nn.utils.clip_grad_norm_(model.parameters(), max_norm=5.) 54 | optimizer.step() 55 | 56 | if step % (len(train_dataloader) // 9) == 0: 57 | print("epoch: {} step: {}/{}".format(epoch, step, len(train_dataloader))) 58 | 59 | torch.save(model.state_dict(), './res/lenet5_last_1c.pt') 60 | 61 | # train_acc = evaluate(model, train_dataloader) 62 | test_acc = evaluate(model, test_dataloader) 63 | if test_acc > max_test_acc: 64 | max_test_acc = test_acc 65 | torch.save(model.state_dict(), './res/lenet5_best_1c.pt') 66 | print("Best model saved!") 67 | 68 | print("epoch: {} test_acc: {:.2f}%".format(epoch, test_acc * 100)) 69 | 70 | 71 | def evaluate(model, data_loader): 72 | model.eval() 73 | pred_tags = [] 74 | true_tags = [] 75 | with torch.no_grad(): 76 | for batch in data_loader: 77 | batch_data = batch[0].to(device) 78 | batch_label = batch[1].to(device) 79 | 80 | logits = model(batch_data) 81 | 82 | pred = torch.argmax(logits, dim=1).cpu().numpy() 83 | tags = batch_label.cpu().numpy() 84 | 85 | pred_tags.extend(pred) 86 | true_tags.extend(tags) 87 | 88 | assert len(pred_tags) == len(true_tags) 89 | correct_num = sum(int(x == y) for (x, y) in zip(pred_tags, true_tags)) 90 | accuracy = correct_num / len(pred_tags) 91 | 92 | return accuracy 93 | 94 | 95 | if __name__ == '__main__': 96 | if not os.path.exists('./res'): 97 | os.mkdir('./res') 98 | train(50, 0.001, 128) 99 | -------------------------------------------------------------------------------- /quantification/input_pic2_lable8.txt: -------------------------------------------------------------------------------- 1 | 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 -------------------------------------------------------------------------------- /cifar-10-torch/cifar-source/distill.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from models import Lenet5 3 | from torchvision.models import resnet152 4 | from torch.optim import Adam 5 | from torch.utils.data import DataLoader 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | import torch 9 | from torchvision.transforms import transforms 10 | from torchvision.datasets import CIFAR10 11 | 12 | device = 'cuda:0' if torch.cuda.is_available() else 'cpu' 13 | 14 | 15 | def run(EPOCHS, lr, batch_size, alpha, T, resume=True): 16 | teacher_model = resnet152() 17 | teacher_model.conv1 = nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) 18 | teacher_model.fc = nn.Linear(2048, 10) 19 | teacher_model.to(device) 20 | teacher_model.load_state_dict(torch.load('./res/resnet152_best_1c.pt')) 21 | student_model = Lenet5().to(device) 22 | if resume: 23 | student_model.load_state_dict(torch.load('./res/distilled_lenet5_best.pt')) 24 | 25 | transform_teacher = transforms.Compose([ 26 | transforms.Grayscale(), 27 | transforms.Resize((224, 224)), 28 | transforms.ToTensor(), 29 | transforms.Normalize((0.449), (0.226)), 30 | ]) 31 | transform_student = transforms.Compose([ 32 | transforms.Grayscale(), 33 | transforms.RandomHorizontalFlip(0.5), 34 | transforms.ToTensor(), 35 | transforms.Normalize((0.4734), (0.2507)), 36 | ]) 37 | transform_test = transforms.Compose([ 38 | transforms.Grayscale(), 39 | transforms.ToTensor(), 40 | transforms.Normalize((0.4734), (0.2507)) 41 | ]) 42 | torch.manual_seed(0) 43 | g = torch.Generator() 44 | train_dataloader_teacher = DataLoader(CIFAR10(root='./cifar-10-python/', train=True, transform=transform_teacher), 45 | shuffle=True, generator=g, batch_size=batch_size, num_workers=4) 46 | torch.manual_seed(0) 47 | g = torch.Generator() 48 | train_dataloader_student = DataLoader(CIFAR10(root='./cifar-10-python/', train=True, transform=transform_student), 49 | shuffle=True, generator=g, batch_size=batch_size, num_workers=4) 50 | test_dataloader = DataLoader(CIFAR10(root='./cifar-10-python/', train=False, transform=transform_test), 51 | batch_size=batch_size, num_workers=4) 52 | 53 | optimizer = Adam(student_model.parameters(), lr=lr) 54 | student_loss = nn.CrossEntropyLoss() 55 | distill_loss = nn.KLDivLoss(reduction='batchmean') 56 | 57 | teacher_model.eval() 58 | max_test_acc = 0. 59 | for epoch in range(EPOCHS): 60 | student_model.train() 61 | for step, (batch_tea, batch_stu) in enumerate(zip(train_dataloader_teacher, train_dataloader_student)): 62 | tea_data = batch_tea[0].to(device) 63 | stu_data = batch_stu[0].to(device) 64 | stu_label = batch_stu[1].to(device) 65 | 66 | with torch.no_grad(): 67 | teacher_logits = teacher_model(tea_data) 68 | 69 | student_model.zero_grad() 70 | logits = student_model(stu_data) 71 | loss = alpha * student_loss(logits, stu_label) + (1 - alpha) * distill_loss( 72 | F.log_softmax(logits / T, dim=1), F.softmax(teacher_logits / T, dim=1)) 73 | 74 | loss.backward() 75 | optimizer.step() 76 | 77 | if step % (len(train_dataloader_student) // 9) == 0: 78 | print("epoch: {} step: {}/{}".format(epoch, step, len(train_dataloader_student))) 79 | 80 | torch.save(student_model.state_dict(), './res/distilled_lenet5_last.pt') 81 | 82 | test_acc = evaluate(student_model, test_dataloader) 83 | if test_acc > max_test_acc: 84 | max_test_acc = test_acc 85 | torch.save(student_model.state_dict(), './res/distilled_lenet5_best.pt') 86 | print("Best model saved!") 87 | 88 | print("epoch: {} test_acc: {:.2f}%".format(epoch, test_acc * 100)) 89 | 90 | 91 | def evaluate(model, data_loader): 92 | model.eval() 93 | pred_tags = [] 94 | true_tags = [] 95 | with torch.no_grad(): 96 | for batch in data_loader: 97 | batch_data = batch[0].to(device) 98 | batch_label = batch[1].to(device) 99 | 100 | logits = model(batch_data) 101 | 102 | pred = torch.argmax(logits, dim=1).cpu().numpy() 103 | tags = batch_label.cpu().numpy() 104 | 105 | pred_tags.extend(pred) 106 | true_tags.extend(tags) 107 | 108 | assert len(pred_tags) == len(true_tags) 109 | correct_num = sum(int(x == y) for (x, y) in zip(pred_tags, true_tags)) 110 | accuracy = correct_num / len(pred_tags) 111 | 112 | return accuracy 113 | 114 | 115 | if __name__ == '__main__': 116 | run(EPOCHS=100, lr=0.001, batch_size=128, alpha=0.3, T=7, resume=True) 117 | -------------------------------------------------------------------------------- /quantification/quantification_para.py: -------------------------------------------------------------------------------- 1 | from multiprocessing.spawn import import_main_path 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | import struct 6 | import ctypes 7 | 8 | 9 | class Lenet5(nn.Module): 10 | ''' 11 | 所用的卷积神经网络 12 | ''' 13 | 14 | def __init__(self): 15 | super(Lenet5, self).__init__() 16 | # 32, 32, 1 17 | self.layer1 = nn.Sequential( 18 | nn.Conv2d(1, 6, kernel_size=5, stride=1, padding=0), 19 | nn.ReLU(), 20 | nn.MaxPool2d(kernel_size=2, stride=2, padding=0) 21 | ) 22 | # 14, 14, 6 23 | self.layer2 = nn.Sequential( 24 | nn.Conv2d(6, 16, kernel_size=5, stride=1, padding=0), 25 | nn.ReLU(), 26 | nn.MaxPool2d(kernel_size=2, stride=2) 27 | ) 28 | # 5, 5, 16 29 | self.fc1 = nn.Sequential( 30 | nn.Linear(400, 120), 31 | nn.ReLU() 32 | ) 33 | self.fc2 = nn.Sequential( 34 | nn.Linear(120, 84), 35 | nn.ReLU() 36 | ) 37 | self.classifier = nn.Linear(84, 10) 38 | 39 | def forward(self, input): 40 | x = self.layer1(input) 41 | x = self.layer2(x) 42 | x = self.fc1(x.view(x.shape[0], -1)) 43 | x = self.fc2(x) 44 | logits = self.classifier(x) 45 | 46 | return logits 47 | 48 | 49 | def dec2bin(x): 50 | ''' 51 | 将十进制小数x转为对应的二进制小数 52 | ''' 53 | x -= int(x) 54 | bins = [] 55 | while x: 56 | x *= 2 57 | bins.append("1" if x >= 1. else "0") 58 | x -= int(x) 59 | return "".join(bins) 60 | 61 | 62 | def float2IEEE16(x): 63 | ''' 64 | float转IEEE754的半精度浮点数 65 | ''' 66 | ms = "0" if x > 0 else "1" 67 | x = abs(x) 68 | x0 = int(x) # 整数部分 69 | x1 = x - x0 # 小数部分 70 | x0 = bin(x0).replace("0b", "") 71 | x1 = dec2bin(x1) 72 | if x0[0] == "0": 73 | E = 15 - x1.find("1") - 1 74 | m = x1[x1.find("1"):] 75 | if E < 0: 76 | E = 15 77 | m = "00000000000" 78 | else: 79 | E = 15 + len(x0) - 1 80 | m = x0 + x1 81 | E = bin(E).replace("0b", "") 82 | if len(E) > 5: 83 | E = E[:5] 84 | else: 85 | for i in range(5 - len(E)): 86 | E = "0" + E 87 | m = m[1:] 88 | if len(m) > 10: 89 | m = m[:10] 90 | else: 91 | for i in range(10 - len(m)): 92 | m += "0" 93 | y = ms+E+m 94 | y1 = "" 95 | for i in range(len(y)//4): 96 | y1 += hex(int(y[4*i:4*(i+1)], 2)).replace("0x", "") 97 | return y1 98 | 99 | 100 | def float2IEEE32(x): 101 | ''' 102 | float转IEEE754的单精度浮点数 103 | ''' 104 | ms = "0" if x > 0 else "1" 105 | x = abs(x) 106 | x0 = int(x) # 整数部分 107 | x1 = x - x0 # 小数部分 108 | x0 = bin(x0).replace("0b", "") 109 | x1 = dec2bin(x1) 110 | if x0[0] == "0": 111 | E = 127 - x1.find("1") - 1 112 | m = x1[x1.find("1"):] 113 | if E < 0: 114 | E = 127 115 | m = "000000000000000000000000" 116 | else: 117 | E = 127 + len(x0) - 1 118 | m = x0 + x1 119 | E = bin(E).replace("0b", "") 120 | if len(E) > 8: 121 | E = E[:8] 122 | else: 123 | for i in range(8 - len(E)): 124 | E = "0" + E 125 | m = m[1:] 126 | if len(m) > 23: 127 | m = m[:23] 128 | else: 129 | for i in range(23 - len(m)): 130 | m += "0" 131 | y = ms+E+m 132 | y1 = "" 133 | for i in range(len(y)//4): 134 | y1 += hex(int(y[4*i:4*(i+1)], 2)).replace("0x", "") 135 | return y1 136 | 137 | 138 | if __name__ == '__main__': 139 | model = Lenet5() 140 | model.load_state_dict(torch.load("distilled_lenet5_best.pt")) 141 | for name in model.state_dict(): 142 | # print(name, '\t', model.state_dict()[name].shape) 143 | # print(model.state_dict()[name]) 144 | 145 | # 卷积层权重量化 146 | if name in ["layer1.0.weight", "layer2.0.weight"]: 147 | fname = name.split(".")[0] 148 | Tensor = model.state_dict()[name] 149 | s1, s2, s3, s4 = Tensor.shape 150 | with open("parameters/"+fname+".txt", "w", encoding="utf-8") as f: 151 | for i in range(s1): 152 | for j in range(s2): 153 | for k in range(s3): 154 | for t in range(s4): 155 | f.write(float2IEEE16(Tensor[i][j][k][t])) 156 | f.write("\n") 157 | # 全连接层权重量化 158 | if name in ["fc1.0.weight", "fc2.0.weight", "classifier.weight"]: 159 | fname = name.split(".")[0] 160 | Matrix = model.state_dict()[name].T 161 | with open("parameters/"+fname+".txt", "w", encoding="utf-8") as f: 162 | for i in range(Matrix.shape[0]): 163 | for j in range(Matrix.shape[1]): 164 | f.write(float2IEEE32(Matrix[i][j])+"\n") 165 | -------------------------------------------------------------------------------- /cifar-10-torch/cifar-source/models.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | 4 | class Lenet5(nn.Module): 5 | def __init__(self): 6 | super(Lenet5, self).__init__() 7 | # 32, 32, 1 8 | self.layer1 = nn.Sequential( 9 | nn.Conv2d(1, 6, kernel_size=5, stride=1, padding=0), 10 | nn.ReLU(), 11 | nn.MaxPool2d(kernel_size=2, stride=2, padding=0) 12 | ) 13 | # 14, 14, 6 14 | self.layer2 = nn.Sequential( 15 | nn.Conv2d(6, 16, kernel_size=5, stride=1, padding=0), 16 | nn.ReLU(), 17 | nn.MaxPool2d(kernel_size=2, stride=2, padding=0) 18 | ) 19 | # 5, 5, 16 20 | self.fc1 = nn.Sequential( 21 | nn.Linear(400, 120), 22 | nn.ReLU(), 23 | nn.Dropout(0.2) 24 | ) 25 | self.fc2 = nn.Sequential( 26 | nn.Linear(120, 84), 27 | nn.ReLU(), 28 | nn.Dropout(0.2) 29 | ) 30 | self.classifier = nn.Linear(84, 10) 31 | 32 | def forward(self, input): 33 | x = self.layer1(input) 34 | x = self.layer2(x) 35 | x = self.fc1(x.view(x.shape[0], -1)) 36 | x = self.fc2(x) 37 | logits = self.classifier(x) 38 | 39 | return logits 40 | 41 | 42 | class NormalCNN(nn.Module): 43 | def __init__(self): 44 | super(NormalCNN, self).__init__() 45 | # (bs, 32, 32, 1) 46 | self.layer1 = nn.Sequential( 47 | nn.Conv2d(1, 32, kernel_size=5, stride=1, padding=0), 48 | nn.BatchNorm2d(32), 49 | nn.ReLU() 50 | ) 51 | # (bs, 28, 28, 32) 52 | self.layer2 = nn.Sequential( 53 | nn.Conv2d(32, 32, kernel_size=5, stride=1, padding=0), 54 | nn.BatchNorm2d(32), 55 | nn.ReLU(), 56 | nn.MaxPool2d(kernel_size=2, stride=2), 57 | nn.Dropout(0.2) 58 | ) 59 | # (bs, 12, 12, 32) 60 | self.layer3 = nn.Sequential( 61 | nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=0), 62 | nn.BatchNorm2d(64), 63 | nn.ReLU(), 64 | ) 65 | # (bs, 10, 10, 64) 66 | self.layer4 = nn.Sequential( 67 | nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=0), 68 | nn.BatchNorm2d(64), 69 | nn.ReLU(), 70 | nn.MaxPool2d(kernel_size=2, stride=2), 71 | nn.Dropout(0.2) 72 | ) 73 | # bs, 4, 4, 64 -> bs, 1024 74 | self.layer5 = nn.Sequential( 75 | nn.Linear(1024, 256), 76 | nn.ReLU() 77 | ) 78 | self.classifier = nn.Linear(256, 10) 79 | 80 | def forward(self, input): 81 | x = self.layer1(input) 82 | x = self.layer2(x) 83 | x = self.layer3(x) 84 | x = self.layer4(x) 85 | x = x.view(x.shape[0], -1) 86 | x = self.layer5(x) 87 | logits = self.classifier(x) 88 | 89 | return logits 90 | 91 | 92 | class NormalCNN2(nn.Module): 93 | def __init__(self): 94 | super(NormalCNN2, self).__init__() 95 | # (bs, 32, 32, 1) 96 | self.layer1 = nn.Sequential( 97 | nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1), 98 | nn.ReLU(), 99 | # nn.BatchNorm2d(32), 100 | nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1), 101 | nn.ReLU(), 102 | nn.MaxPool2d(kernel_size=2, stride=2) 103 | ) 104 | # (bs, 28, 28, 32) 105 | self.layer2 = nn.Sequential( 106 | nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1), 107 | # nn.BatchNorm2d(32), 108 | nn.ReLU(), 109 | nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1), 110 | nn.ReLU(), 111 | nn.MaxPool2d(kernel_size=2, stride=2) 112 | # nn.Dropout(0.2) 113 | ) 114 | # (bs, 12, 12, 32) 115 | self.layer3 = nn.Sequential( 116 | nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1), 117 | # nn.BatchNorm2d(64), 118 | nn.ReLU(), 119 | nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1), 120 | nn.ReLU(), 121 | nn.MaxPool2d(kernel_size=2, stride=2) 122 | ) 123 | # (bs, 10, 10, 64) 124 | # self.layer4 = nn.Sequential( 125 | # nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=0), 126 | # nn.BatchNorm2d(64), 127 | # nn.ReLU(), 128 | # nn.MaxPool2d(kernel_size=2, stride=2), 129 | # nn.Dropout(0.2) 130 | # ) 131 | # bs, 4, 4, 64 -> bs, 1024 132 | # self.layer5 = nn.Sequential( 133 | # nn.Linear(1024, 256), 134 | # nn.ReLU() 135 | # ) 136 | self.classifier = nn.Linear(4096, 10) 137 | 138 | def forward(self, input): 139 | x = self.layer1(input) 140 | x = self.layer2(x) 141 | x = self.layer3(x) 142 | # x = self.layer4(x) 143 | x = x.view(x.shape[0], -1) 144 | # x = self.layer5(x) 145 | logits = self.classifier(x) 146 | 147 | return logits 148 | -------------------------------------------------------------------------------- /weight/layer2.txt: -------------------------------------------------------------------------------- 1 | 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 2 | 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 3 | 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 4 | 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 5 | a93d2d5b2ed42c16a2272d7f28b22ddb2cf12d86ab911b542641254628852ef82a7396eca37e2d4fa438b149b17fb340ad2caff0aa5eb024af7fb290b12cae37ac5b26caa52aacfdacb32015ac172ed4b13020f62a252e9f34072f30326b31c0345b319621c8a8ca259d2cc4a4531d832e7a30592a742e7a2e79305c30d92d012dc632992f2132c12d64a6da291730a92e1832c035362a4d2ec81d8da845ad3b30609f07296598fbb004307c2ec42d64afc8b055306eadc9b065b238b184aee0b1d7b054a48cafa42fe22cfb2c7caacaa81f28ef315d2d5a2c9ab20c305b31232f192d54b2d131ff307d2d220fd4b47e2890243fad9ca7a8b37ab0b3a64029f09805a596aaf6a9ccac202b79a0eaae65ad16a54b2836ab9c27ed2e54001eb05bb4a3a5071da4add9af16b0d1 6 | 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 7 | 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 8 | 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 9 | 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 10 | a76426db23812fb11c572abcb4112e292c252c9ea40fab6a304b2be92a4daeb934d4a1652a10242432a932cf2f6829e32bf3283bad34a37b2ceb2a802947a407b03127c7a550a6662d2db01facf72885ab3ea4b03065afb6a9cba43c9e052b13ab70b0832cb0b179aaa8ac79ae06aec9ab322c8ea879ae0131b92632b53fa9222c552ae9b255aef730843054ad632224347029c81539b10c27f72e0922013128b05a29502b3d2b202dddaea2ab272a5e2d622acb2c8a2ecab3d62b0aa4e322b2297728d92d0730e0aa5d2250a475ace6a7eeb094301ba823b1bb2b09b0a92a02ac18a8222eafac1bb28d2e46305c3058a99bac05275c26202f2d2d7fa4742dc5b1aaab5f30deaf10a77fa003ad7930e2aea5b5582904a44129b52fdbb2f1b119aed92c112679ac55b241b07e 11 | 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 12 | 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 13 | 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 14 | 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 15 | 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 16 | 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 17 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # CNN-FPGA 2 | 3 | 基于FPGA的cifar-10二维卷积识别任务 4 | 5 | ## Background - 背景 6 | 7 | > 本项目为基于FPGA的二维卷积cifar-10识别任务。 8 | > 9 | > 参考项目链接：https://github.com/omarelhedaby/CNN-FPGA. 我们主要的工作是针对这个开源项目进行补充和修改，并适配到我们的任务上（如果是MNIST也没问题）。 10 | 11 | ## Introduction - 项目简介 12 | 13 | ### 技术路线 14 | 15 |
16 | 17 |
18 | 19 | 如上图所示，我们的技术路线从工作内容来说，一共分为五部分：需求定义、数据准备、模型构建、verilog编写、工程部署。 20 | 21 | 1. 在**需求定义**阶段，我们的主要任务就是确定整个小组的任务目标：针对输入为彩色图像的多对象分类任务仿真。 22 | 2. 基于我们的任务需求，在**数据准备**阶段，我们小组要做的事情就是确定要使用的数据集，并对数据集进行清洗、量化等一系列预处理工作，使数据集能为我们的模型所用。 23 | 3. 准备好数据集后就是**模型构建**阶段，这一阶段主要是针对分类任务的算法层面实现，因此我们工作的发力点在于训练一个合适的神经网络模型，并通过蒸馏、剪枝、权重量化等操作使模型转变为更适合部署在芯片上的状态。 24 | 4. 随后就是模型的**Verilog仿真**实现，在这一阶段我们的主要任务就是学习Verilog的使用，并在确定合适的开发方式后将神经网络模型仿真实现。 25 | 5. 最后一步就是在得到具体芯片的情况下，尝试将我们的仿真结果**上板烧录**，转化为可以实际应用的工程成果。 26 | 27 | ### 网络结构 28 | 29 | 下图展示了我们模型的结构，针对cifar-10数据集中的输入，进行多轮操作，得到最终的分类结果。 30 | 31 |
32 | 33 |
34 | 35 | 更具体的参数设置如下所示（激活函数使用ReLu）： 36 | 37 |
38 | 39 |
40 | 41 | ### 识别结果 42 | 43 | | 模型名称 | 精度 | 44 | | :---------------------: | :--------: | 45 | | teacher模型（resnet） | 0.947 | 46 | | student模型（lenet） | 0.674 | 47 | | 非蒸馏lenet | 0.526 | 48 | | 量化后lenet（仿真结果） | 0.55 | 49 | 50 | 最终的结果如上表所示，由于初学的原因，所以没有尝试使用verilog搭建更复杂的网络。因此，我们选择基于resnet模型进行蒸馏，以此来让lenet-5学到更好的结果。那么由于数据量化会损失一定精度，最终基于FPGA的CNN识别cifar-10的准确率到达55%。 51 | 52 | ## Our Work - 我们的工作 53 | 54 | 首先说明为什么参考[这个项目](https://github.com/omarelhedaby/CNN-FPGA)： 55 | 56 | 1. 源码采用层次化、模块化设计，方便阅读和理解，对小白及其友好； 57 | 2. 提供了很多写好的代码模块； 58 | 3. B站up主推荐。 59 | 60 | 然后说说我们做了哪些工作： 61 | 62 | 1. 原项目目标为手写数字识别，我们进一步拓展，研究其在cifar-10数据集上的效果； 63 | 2. 据此，我们基于pytorch设计了网络，并使用蒸馏得到更好的结果； 64 | 3. 删除第三层卷积； 65 | 4. 增加了一层全连接； 66 | 5. 在卷积层使用了relu替代tanh（自编代码）； 67 | 6. 使用maxpool替代avgpool（自编代码）； 68 | 7. 修改了源码中的一些错误（如reset命名错误，数据位宽错误等）； 69 | 8. 改变了全连接层的输入维度、输出维度； 70 | 9. 编写了卷积层的testbench，并通过了仿真； 71 | 10. 自编16转32位转换器以及对应tesetbench代码； 72 | 11. 原项目独立编写了卷积和全连接，我们将二者合到了一起； 73 | 12. 上板进行测试； 74 | 13. 编写了中文注释，方便阅读; 75 | 14. 删除了softmax； 76 | 15. 提供了pytroch的训练源码、模型以及量化代码。 77 | 78 | ## Module Description - 模块说明 79 | 80 | > *更多有关其中的计算说明详见**技术报告.pdf*** 81 | 82 | ### integrationConv 83 | 84 | **说明：** 85 | 86 |   卷积模块，对模型中的卷积部分（包含激活、池化）进行仿真，对应代码中的integrationConvPart.v。 87 | 88 | **可配置参数：** 89 | 90 | | 名称 | 说明 | 默认值 | 91 | | :--------: | :--------------: | :----: | 92 | | DATA_WIDTH | 数据位宽 | 16 | 93 | | ImgInW | 输入图像的宽度 | 32 | 94 | | ImgInH | 输入图像的高度 | 32 | 95 | | Conv1Out | 第一层卷积输出 | 28 | 96 | | MvgP1out | 最大池化输出 | 14 | 97 | | Conv2Out | 第二层卷积输出 | 10 | 98 | | MvgP2out | 最大池化输出 | 5 | 99 | | Kernel | 卷积核的大小 | 5 | 100 | | DepthC1 | 第一层卷积核数量 | 6 | 101 | | DepthC2 | 第二层卷积核数量 | 16 | 102 | 103 | **输入输出：** 104 | 105 | | 名称 | 类型 | 说明 | 长度 | 106 | | :---------: | :----: | :----------------------------------------------------------: | :----------------------------------------------: | 107 | | CNNinput | input | 输入的图像，数据从左上至右下排列，每一个像素值用半精度浮点数表示 | ImgInW × ImgInH × DATA_WIDTH | 108 | | Conv1F | input | 第一层卷积核权值，从第一个卷积核左上开始，到最后一个卷积核右下，每一个值用半精度浮点数表示 | Kernel × Kernel × DepthC1× DATA_WIDTH | 109 | | Conv2F | input | 第二层卷积核权值，从第一个卷积核左上开始，到最后一个卷积核右下，每一个值用半精度浮点数表示 | DepthC2 × Kernel × Kernel × DepthC1 × DATA_WIDTH | 110 | | iConvOutput | output | 输出的特征图 | MvgP2out × MvgP2out × DepthC2 × DATA_WIDTH | 111 | 112 | ### Relu 113 | 114 | **说明：** 115 | 116 |   激活函数单元1，relu激活函数。对应代码中的UsingTheRelu16.v和activationFunction.v（二者数据位宽不同，其实有冗余，懒得改了） 117 | 118 | **可配置参数：** 119 | 120 | | 名称 | 说明 | 默认值 | 121 | | :----------: | :------: | :----: | 122 | | DATA_WIDTH | 数据位宽 | 16 | 123 | | OUTPUT_NODES | 输出位宽 | 32 | 124 | 125 | **输入输出：**（每层不同，这里仅举例） 126 | 127 | | 名称 | 类型 | 说明 | 长度 | 128 | | :-------: | :----: | :------: | :---------------------------: | 129 | | input_fc | input | 输入特征 | Conv1Out × Conv1Out × DepthC1 | 130 | | output_fc | output | 输出特征 | Conv1Out × Conv1Out × DepthC1 | 131 | 132 | ### MaxPoolMulti 133 | 134 | **说明：** 135 | 136 |   最大池化模块，可以对输入进行最大池化运算。 137 | 138 | **可配置参数：** 139 | 140 | | 名称 | 说明 | 默认值 | 141 | | :--------: | :----------: | :----: | 142 | | DATA_WIDTH | 数据位宽 | 16 | 143 | | D | 数据通道 | 32 | 144 | | H | 输入特征高度 | 28 | 145 | | W | 输出特征宽度 | 28 | 146 | 147 | **输入输出：**（每层不同，这里仅举例） 148 | 149 | | 名称 | 类型 | 说明 | 长度 | 150 | | :------: | :----: | :------: | :----------------------------: | 151 | | apInput | input | 输入特征 | H × W × D × DATA_WIDTH | 152 | | apOutput | output | 输出特征 | (H/2) × (W/2) × D × DATA_WIDTH | 153 | 154 | ### ANNfull 155 | 156 | **说明：** 157 | 158 |   全连接模块，对模型中的全连接部分（包含激活）进行仿真，对应代码中的ANNfull.v。 159 | 160 | **可配置参数：** 161 | 162 | | 名称 | 说明 | 默认值 | 163 | | :------------: | :--------------: | :----: | 164 | | DATA_WIDTH | 数据位宽 | 32 | 165 | | INPUT_NODES_L1 | 第一层输入节点数 | 400 | 166 | | INPUT_NODES_L2 | 第二层输入节点数 | 120 | 167 | | INPUT_NODES_L3 | 第三层输入节点数 | 84 | 168 | | OUTPUT_NODES | 输出节点数 | 10 | 169 | 170 | **输入输出：** 171 | 172 | | 名称 | 类型 | 说明 | 长度 | 173 | | :--------: | :----: | :-------------------------------------------------: | :-------------------------: | 174 | | input_ANN | input | 全连接层的输入，用单精度浮点数表示 | DATA_WIDTH × INPUT_NODES_L1 | 175 | | output_ANN | output | 预测的标签值，cifar-10为10分类，需要用4位二进制表示 | 4 | 176 | 177 | ### IEEE162IEEE32 178 | 179 | **说明：** 180 | 181 |   精度转换模块，将16位宽浮点数转换为32位浮点数，对应代码中的IEEE162IEEE32.v。 182 | 183 | **可配置参数：** 184 | 185 | | 名称 | 说明 | 默认值 | 186 | | :----------: | :----------: | :----: | 187 | | DATA_WIDTH_1 | 输入数据位宽 | 16 | 188 | | DATA_WIDTH_2 | 输出数据位宽 | 32 | 189 | | NODES | 输出节点数 | 400 | 190 | 191 | **输入输出：** 192 | 193 | | 名称 | 类型 | 说明 | 长度 | 194 | | :-------: | :----: | :----------------------------------------: | :------------------: | 195 | | input_fc | input | 精度转换模块的输入，数据用半精度浮点数表示 | DATA_WIDTH_1 × NODES | 196 | | output_fc | output | 精度转换模块的输出，数据用单精度浮点数表示 | DATA_WIDTH_2 × NODES | 197 | 198 | ### LeNet 199 | 200 | **说明：** 201 | 202 |   整个网络模块，包含两层卷积和三层全连接，对应源码中的Lenet.v. 203 | 204 | **可配置参数：** 205 | 206 | | 名称 | 说明 | 默认值 | 207 | | :----------: | :--------------: | :----: | 208 | | DATA_WIDTH_1 | 卷积层数据位宽 | 16 | 209 | | DATA_WIDTH_2 | 全连接层数据位宽 | 32 | 210 | | ImgInW | 输入图像的宽度 | 32 | 211 | | ImgInH | 输入图像的高度 | 32 | 212 | | Kernel | 卷积核的大小 | 5 | 213 | | MvgP2out | 最大池化输出 | 5 | 214 | | DepthC1 | 第一层卷积核数量 | 6 | 215 | | DepthC2 | 第二层卷积核数量 | 16 | 216 | 217 | **输入输出：** 218 | 219 | | 名称 | 类型 | 说明 | 长度 | 220 | | :---------: | :----: | :----------------------------------------------------------: | :------------------------------------------------: | 221 | | CNNinput | input | 输入的图像，数据从左上至右下排列，每一个像素值用半精度浮点数表示 | ImgInW × ImgInH × DATA_WIDTH_1 | 222 | | Conv1F | input | 第一层卷积核权值，从第一个卷积核左上开始，到最后一个卷积核右下，每一个值用半精度浮点数表示 | Kernel × Kernel × DepthC1× DATA_WIDTH_1 | 223 | | Conv2F | input | 第二层卷积核权值，从第一个卷积核左上开始，到最后一个卷积核右下，每一个值用半精度浮点数表示 | DepthC2 × Kernel × Kernel × DepthC1 × DATA_WIDTH_1 | 224 | | LeNetoutput | output | 输出的特征图 | 3 | 225 | 226 | ------ 227 | 228 | ## Requirements - 必要条件 229 | 230 | - Windows 231 | - python3.7 or up 232 | - pytorch 233 | - vivado 234 | 235 | ## Usage - 使用方法 236 | 237 | 1. 下载本仓库 238 | 239 | ``` 240 | mkdir CNN-FPGA 241 | cd ./CNN-FPGA 242 | git clone https://github.com/Robin-WZQ/CNN-FPGA.git 243 | ``` 244 | 245 | 2. 训练模型 246 | 247 | > 已提供训练好的模型。 248 | 249 | ``` 250 | cd ./cifar_source_torch 251 | python train.py 252 | python distill.py 253 | python save_params.py 254 | ``` 255 | 256 | 3. 量化并保存 257 | 258 | > 如果用自己训练好的模型，就将pt文件移动到quantification文件夹下。 259 | > 260 | > 修改ANNfull.v和Lenet_tb.v中的权重路径，我们提供了我们模型训练好的权重文件，详见weight文件夹中。 261 | 262 | ``` 263 | # 权重量化 264 | python quantification_para.py 265 | # 输入图像量化 266 | python quantification_img.py 267 | ``` 268 | 269 | 4. 运行仿真 270 | 271 | 打开CNN-FPGA-Vivado，在 vivado 里，运行LeNet_tb的simulation，即可得到结果。 272 | 273 | 一开始为9，之后运行约1~2分钟之后变为预测的数字。 274 | 275 |
276 | 277 |
278 | 279 | 280 | ## Code Tree - 文件代码树 281 | 282 | > 文件及对应说明如下所示 283 | 284 | |cifar_source_torch 285 | 286 | ----|cifar-source # 源代码文件夹 287 | 288 | --------|cifar-10-python # 数据集文件夹 289 | 290 | --------|res # 模型存储文件夹 291 | 292 | --------|distilled_lenet5_best.pt # 蒸馏得到的Lenet5模型参数文件 293 | 294 | --------|train.py # Teacher model训练代码 295 | 296 | --------|distill.py # 模型蒸馏训练代码 297 | 298 | --------|models.py # 模型结构代码 299 | 300 | --------|save_params.py # 参数导出代码 301 | 302 | --------|test.py # 测试代码 303 | 304 | --------|Lenet5_parameters_cifar.txt # 导出模型参数文件 305 | 306 | ----|README.md # 说明文件 307 | 308 | |CNN-FPGA-Vivado # 包含testbench和源码，以及整个工程文件 309 | 310 | ----|CNN-FPGA-Vivado.src # 项目源码 311 | 312 | --------|sim_1/new # testbench源码 313 | 314 | --------|sources_1/new # 工程源码 315 | 316 | --------|CNN-FPGA-Vivado.xpr # 项目工程 317 | 318 | |quantification 319 | 320 | ----|cifar-10-python # 数据集 321 | 322 | ----|distilled_lenet5_best.pt # 训练好的模型 323 | 324 | ----|input_pic2_label8.txt # 数据转换例子 325 | 326 | ----|quantification_img.py # 输入图像量化代码 327 | 328 | ----|quantification_para.py # 模型权重量化代码 329 | 330 | |weight 331 | 332 | ----|classifier.txt # 全连接第3层权重 333 | 334 | ----|fc1.txt # 全连接第1层权重 335 | 336 | ----|fc2.txt # 全连接第2层权重 337 | 338 | ----|layer1.txt # 卷积第1层权重 339 | 340 | ----|layer2.txt # 卷积第2层权重 341 | 342 | 343 | ## LICENSE 344 | MIT 345 | 346 | -------------------------------------------------------------------------------- /weight/classifier.txt: -------------------------------------------------------------------------------- 1 | be4b4c87 2 | 3ceefe79 3 | 3df8e9db 4 | bd516a79 5 | be7c7673 6 | be152151 7 | be1fd5dc 8 | 3dc65057 9 | 3e7ef06a 10 | 3e0511a1 11 | 3de60a14 12 | 3cb9cbe3 13 | 3da85904 14 | bd1fdd59 15 | bc6086cc 16 | 3de90085 17 | bbbc85b3 18 | 3d7e6d9b 19 | bd428e8a 20 | bec021ff 21 | be696c34 22 | bdd42c08 23 | 3da58f4d 24 | 3d01d170 25 | bdb3b2e8 26 | 3dd2bdab 27 | bcd0b617 28 | 3e499ede 29 | be2f5533 30 | 3e063bbe 31 | bb923bd1 32 | 3f12aa47 33 | bd943697 34 | 3dc6f077 35 | 3e8366f0 36 | bccf3a32 37 | bec2e4a9 38 | bde4fd24 39 | bec128cd 40 | bdeea901 41 | 3e39a185 42 | be1b9092 43 | 3dfbd8bd 44 | be1fefde 45 | be614541 46 | be4667ac 47 | be84c283 48 | befa1750 49 | be15fd73 50 | 3f287d43 51 | 3e3ddcab 52 | 3e3d35de 53 | be445548 54 | 3cfb6877 55 | 3dc37e97 56 | bd29e38d 57 | be942f6f 58 | 3e1b3cd1 59 | bda955d9 60 | 3e0e0536 61 | beb2bdc0 62 | 3e8f994e 63 | 3d94386a 64 | bdc27877 65 | bd1e6e31 66 | be51c476 67 | 3e209b51 68 | 3e184ea3 69 | 3d9694ad 70 | be2a3c10 71 | 3d887726 72 | 3e23fc64 73 | bdf36936 74 | be7cfc6b 75 | 3dff1e27 76 | bd00db19 77 | 3e8af1d0 78 | be101818 79 | be9cad79 80 | 3dcc0c18 81 | bea2cfab 82 | be9b5479 83 | 3e1a1da7 84 | 3d34f0d8 85 | be010c8b 86 | 3e8178e1 87 | bdf3afc6 88 | 3d357e41 89 | beeaaf20 90 | 3e076c98 91 | be463c95 92 | 3e0b3eae 93 | bd1365a3 94 | 3e172c3e 95 | be1e9071 96 | 3ddbbe6b 97 | bc42e3cf 98 | 3e213d16 99 | be15b1d2 100 | 3e0bf2a7 101 | be7987c6 102 | bef632f0 103 | 3dbb129a 104 | 3de9d764 105 | 3d442dfb 106 | 3e5610e2 107 | be5ff38e 108 | be47426f 109 | 3d9da125 110 | 3e6f0e43 111 | 3cb3be23 112 | 3c78f19a 113 | bf0ed667 114 | 3def653e 115 | 3e9fde29 116 | 3c1a8b87 117 | be4a1648 118 | be715a24 119 | 3d11f26c 120 | 3ed57b71 121 | be0700d5 122 | bed43dd6 123 | 3e02fd20 124 | 3dc0c7c3 125 | be9a4225 126 | bdbf4c57 127 | 3e2529be 128 | bd390ca8 129 | 3e5fffc4 130 | 3e4da289 131 | bededc6f 132 | 3eb2c667 133 | 3ed8901c 134 | beb0cc27 135 | 3de257e4 136 | bea91c15 137 | bcbf7298 138 | bd4ad9ab 139 | 3f29a3c3 140 | be17b258 141 | 3ddb49d5 142 | 3d9b97f2 143 | 3e8d7fd4 144 | be6608f8 145 | beebb0f6 146 | bb9d5783 147 | 3e0254fd 148 | bb97b60d 149 | 3db2105b 150 | be2ddc83 151 | 3db5d7ac 152 | bf022625 153 | 3dc6fa45 154 | 3e07d6b1 155 | 3d6f63ee 156 | be2c7eb7 157 | beb83a3f 158 | 3dc3cf10 159 | b8dad2c0 160 | 3d5cc908 161 | 3d00f10c 162 | 3ddce94e 163 | bdab0f37 164 | be27023c 165 | be3a13c3 166 | 3dee60d1 167 | 3da511ee 168 | be38da03 169 | 3d64197f 170 | 3e229db8 171 | 3cc2713f 172 | bf24c0a0 173 | 3c2cc42e 174 | beb2bdf6 175 | 3dcb6672 176 | becbd7f9 177 | beee7517 178 | bf17f1b3 179 | 3ea41113 180 | 3cd11f7b 181 | 3daaf53a 182 | be68df39 183 | bd685ae9 184 | bdb6200f 185 | 3e36f91b 186 | 3e18ae03 187 | 3df5ddc9 188 | 3e4011af 189 | bdeec2ed 190 | beacf0cd 191 | 3d73b97b 192 | be84213f 193 | be37391b 194 | bdad8007 195 | 3e534556 196 | be910a89 197 | 3da5e232 198 | 3df2b739 199 | bee20776 200 | 3e3105db 201 | becd7c03 202 | be04f3d0 203 | be7c8833 204 | 3e346be8 205 | 3dfca8de 206 | 3e6c044c 207 | 3dabf6e6 208 | bea13ba6 209 | 3e84ea7a 210 | 3e25b1b6 211 | 3e631f2c 212 | be05668d 213 | bea00656 214 | 3cb90908 215 | bc26cf0c 216 | bd758caa 217 | 3db27476 218 | 3e371835 219 | be84ace3 220 | 3e3fb6e6 221 | 3e96401e 222 | 3e7ca2f3 223 | beb96dcb 224 | bdd33281 225 | be9aa7cb 226 | 3e564dc9 227 | be8507de 228 | 3e14228a 229 | be0b4ff8 230 | be9a50ba 231 | bd022c3a 232 | be3d4245 233 | 3e49874a 234 | be8639d2 235 | 3d947469 236 | 3e6aa7f3 237 | be63a515 238 | 3ddb556f 239 | 3d55fe41 240 | bee033f0 241 | bd295b2d 242 | bd7a1934 243 | bd292918 244 | 3cb1a640 245 | bd9a7e1b 246 | 3c1f44ac 247 | 3d6fc863 248 | bb8e48fd 249 | 3d7e8b9d 250 | bd32fb2e 251 | 3ec9b0b6 252 | 3d1c0562 253 | 3ef564e0 254 | bd9f0a3a 255 | bf017410 256 | be7b615c 257 | 3cb118e0 258 | be2a541d 259 | be8cbcb7 260 | 3e4bdfba 261 | be868cf8 262 | 3c8b82b3 263 | be6829d6 264 | 3dd14a87 265 | 3ddd8634 266 | 3bc7b5f8 267 | 3d813b0f 268 | 3db10327 269 | 3e241e70 270 | be89dede 271 | 3e0fc28a 272 | 3e0c10ee 273 | 3d9daaa2 274 | bd58100f 275 | be372728 276 | 3e0456e5 277 | 3c7c28e2 278 | be92c5a5 279 | 3e30bfad 280 | 3deebed5 281 | bec19155 282 | be8fb393 283 | bd63b4ef 284 | 3d4c7f5c 285 | 3e6a85f1 286 | 3ddfb626 287 | 3e2f3b7d 288 | 3b452693 289 | 3e3a6b43 290 | be4794bb 291 | 3c974dac 292 | 3e75c7b3 293 | bec124d1 294 | ba3cd6c7 295 | bdc6be8b 296 | be757c79 297 | 3d1487aa 298 | bd7b960c 299 | b9ce92bf 300 | 3e3512af 301 | bdba53e5 302 | 3e1dd6df 303 | bf082324 304 | bde256eb 305 | be50d430 306 | 3e2b8aaf 307 | 3e64f3d9 308 | be0efc9a 309 | 3e86f412 310 | b91f4290 311 | bec413de 312 | 3e842537 313 | 3de4e35d 314 | be2ee674 315 | 3e6f5d89 316 | bafc3a83 317 | 3dccea55 318 | 3d01fee0 319 | 3cb744e0 320 | be4aace8 321 | 3e62696a 322 | beffa57f 323 | be278da9 324 | 3e13a2cb 325 | bdd2d745 326 | bd8df180 327 | 3d6113fc 328 | bda807ae 329 | 3e81e2d7 330 | 3df02bd6 331 | 3daca07e 332 | 3dc02455 333 | bb9bb816 334 | 3cec78d7 335 | bdffcb9d 336 | be297c67 337 | bd456930 338 | be45293a 339 | 3dc2ba39 340 | 3da920b7 341 | bf123b3d 342 | 3ecb4f51 343 | 3e38ffdd 344 | bcf4321a 345 | bd9e04bb 346 | bd76bdbe 347 | be54fe8d 348 | be8c2110 349 | be5d8de4 350 | 3e627536 351 | 3db1ffa3 352 | bdc79259 353 | bd94a564 354 | be531384 355 | 3dc975c9 356 | be19272d 357 | beabe813 358 | 3d8048a6 359 | 3e653772 360 | 3dd9fb6e 361 | be53ae36 362 | 3e18fc95 363 | be9194f5 364 | 3d836c3b 365 | 3e19f681 366 | 3c1057df 367 | be0530a6 368 | 3d2139fc 369 | be9afa4c 370 | 3e252262 371 | be5ab526 372 | bd7c6756 373 | bd852e64 374 | bdf6fb51 375 | be489f5c 376 | 3dba380b 377 | be1d80e8 378 | 3e4e2ca4 379 | bda10ebe 380 | 3e511e4c 381 | 3e2d41d8 382 | bf15df98 383 | 3dcd14bf 384 | 3d811548 385 | 3e114d15 386 | 3dd62ecf 387 | bec7b3ea 388 | 3da8dc7d 389 | bea6d31e 390 | be4c5667 391 | bf075800 392 | bf2216b5 393 | 3d0c6ee5 394 | bf01c228 395 | 3f3de054 396 | beeebbfb 397 | bef02729 398 | 3dd22729 399 | 3e33d67a 400 | 3d4e790c 401 | bd37c3f3 402 | 3c781451 403 | 3d88df1c 404 | bcc66f80 405 | be0b2d2d 406 | 3d56915a 407 | bd7a7871 408 | 3d917d6a 409 | bba9f4f3 410 | bd8e512c 411 | 3c43c976 412 | be896b18 413 | 3d130684 414 | bd58c972 415 | bd8fc565 416 | bed598a5 417 | bd1964e0 418 | 3ea9b437 419 | 3e02397f 420 | 3d0d0a2f 421 | bda64064 422 | 3d124be7 423 | bdf5bc2a 424 | bba43eb5 425 | bdb57cdf 426 | 3d089adf 427 | bcf86830 428 | bdac9296 429 | be078e1c 430 | bded709e 431 | 3db38127 432 | befd1644 433 | 3e89326d 434 | be6f71db 435 | 3e801a82 436 | bdd28715 437 | 3e4b8607 438 | bd8c71ba 439 | 3c07a443 440 | bf0aa91a 441 | beafb748 442 | be303398 443 | bf0c3347 444 | bf03d512 445 | be96f413 446 | 3d992d1d 447 | be7b32b1 448 | 3f6ac1d0 449 | bdce6105 450 | beae3e35 451 | bdb182d3 452 | be0147c9 453 | 3e0ff084 454 | 3e024e99 455 | bb4478ee 456 | 3e05934e 457 | bee1e00d 458 | 3e2404aa 459 | 3e8e944d 460 | be97abc6 461 | 3eb396f0 462 | 3e2a4de7 463 | be46e320 464 | 3e2ec41c 465 | bec7f8b1 466 | b9ad51d8 467 | be641b66 468 | 3db9dd07 469 | be79156a 470 | 3e306831 471 | 3c5e7144 472 | 3cd0523c 473 | 3daa4266 474 | 3e408f95 475 | be9a4b73 476 | beac3908 477 | 3e230a6a 478 | be6ce08d 479 | 3e2e01e4 480 | bc8b58fa 481 | 3e9272c6 482 | be13870e 483 | bf417bc2 484 | bd72f224 485 | 3e37c4b7 486 | 3e95a655 487 | bdd1472f 488 | be75e993 489 | 3e8bdcfb 490 | be179790 491 | be98c6c0 492 | 3e913774 493 | be28c58a 494 | 3da79a04 495 | bd6b28ea 496 | 3e595881 497 | be6adfa5 498 | 3dff4102 499 | 3cf5eab0 500 | bed3c62c 501 | bda1368c 502 | 3e33f94a 503 | 3e13618b 504 | bde4b8dd 505 | bd31ed46 506 | be23c22a 507 | 3d9890b0 508 | 3c55bdd5 509 | 3e354ab9 510 | 3e513dea 511 | 3e683c1f 512 | be3a9377 513 | be1e2ef2 514 | 3dbb5490 515 | 3e3ec1e4 516 | bd8e1b73 517 | bd8051aa 518 | be874d92 519 | 3e526bd9 520 | bec7f2e4 521 | 3f0ac195 522 | 3cd00bf1 523 | be723578 524 | be6d8891 525 | be39fd6d 526 | 3eaed145 527 | 3d8d290b 528 | be3aa2b9 529 | bf0910f5 530 | 3de292c5 531 | bd86f9fa 532 | 3da2d50f 533 | 3c6e107e 534 | 3e2a2f85 535 | 3d45a7fc 536 | be09a22c 537 | 3d557a17 538 | bd207464 539 | beb131d0 540 | 3d6a6cde 541 | be02bc20 542 | be840e79 543 | 3de47749 544 | 3dec05c8 545 | be1ddd41 546 | 3d7c7ba0 547 | bd16a2f6 548 | 3df58cfb 549 | 3d872d1e 550 | be772587 551 | be7e761e 552 | 3d8072c8 553 | be12f833 554 | bd9c8ddc 555 | 3e0d47b2 556 | 3de284cb 557 | be0c68c8 558 | bd991fab 559 | be2e8990 560 | 3e5fad86 561 | 3dfd38f0 562 | 3e5716f6 563 | 3dd6a5aa 564 | bd40cfc5 565 | be81b90a 566 | bd0c04c9 567 | beafd518 568 | be8fc398 569 | 3d030383 570 | bbd5d3a6 571 | 3e5d82bf 572 | 3d75bc4c 573 | 3d8e0492 574 | 3ccbd89f 575 | bec219c0 576 | 3dd65b11 577 | be8bdf5d 578 | 3cdd6662 579 | bdfe181e 580 | 3de9ef79 581 | 3e8bce8a 582 | 3f4620d9 583 | 3ecc825e 584 | be4e4500 585 | bec855b7 586 | be717dc0 587 | 3dff3742 588 | bf00e734 589 | be203a40 590 | 3ef00c92 591 | 3e500c9a 592 | 3e0b6dcb 593 | 3e300e61 594 | be120d4c 595 | 3ca17410 596 | be6df469 597 | 3a0232b6 598 | 3dff5fd2 599 | be1d1768 600 | 3d56c1f8 601 | bda5d7f3 602 | be915235 603 | 3e26758c 604 | 3d6c2061 605 | bc220df9 606 | 3d970f43 607 | bd1121c9 608 | 3d6fcde5 609 | bef7b796 610 | be0f0f76 611 | be373c40 612 | be5923b7 613 | 3e046f2c 614 | 3e84aa2b 615 | bd510266 616 | bdc342ca 617 | 3eb8c21d 618 | 3cc62193 619 | bf2b60c7 620 | be154a82 621 | be784795 622 | bee79fd5 623 | 3e005680 624 | 3deccaaa 625 | 3e6ef988 626 | 3d53a765 627 | bba5c484 628 | 3d8827eb 629 | be20d5ed 630 | bd0fc377 631 | 3da02278 632 | 3e51c5b5 633 | be184a16 634 | be938639 635 | be9f1e8c 636 | 3b01cd9e 637 | 3d7604f4 638 | be028a83 639 | be31f12a 640 | 3e36bc36 641 | beb4bcfd 642 | 3e131049 643 | 3e793c56 644 | 3e61eb48 645 | 3de59df2 646 | 3db20a1e 647 | be6feeaf 648 | be59c316 649 | be298e65 650 | bd251738 651 | 3cf8911e 652 | be7745b2 653 | be317190 654 | 3df12dbb 655 | be8d2474 656 | 3e2f8e5b 657 | bdfc4a22 658 | bae25d72 659 | 3ee8aac5 660 | 3ba0ace6 661 | bdb619f2 662 | bca4e24d 663 | 3cc4f189 664 | bd4c6617 665 | bc36a791 666 | bd96eba2 667 | bdcbe048 668 | 3cb2f8a9 669 | be02fb45 670 | 3d78ed93 671 | 3e2ba221 672 | 3e154be9 673 | 3e26899b 674 | 3c1578d9 675 | 3e07adcb 676 | bea6d642 677 | be3b146c 678 | bcac56fc 679 | 3d27e08a 680 | 3e10be2d 681 | bdb35f49 682 | 3e0aea64 683 | be6e3719 684 | 3dadd721 685 | 3d939025 686 | bde3522e 687 | 3de81125 688 | 3cec8dd2 689 | 3d96360f 690 | 3d9b22ee 691 | 3cf7dcf4 692 | bd9d3632 693 | bd710481 694 | bd1fffa7 695 | 3d418a85 696 | bdb1bddf 697 | 3d491f9a 698 | bdaf92c8 699 | 3d8c79b9 700 | bd9581a6 701 | 3ded73be 702 | bdbc119e 703 | bd928435 704 | 3e160b82 705 | be815055 706 | 3e39de7f 707 | 3edd01fa 708 | be9018ea 709 | bebe0e94 710 | be9d4756 711 | 3dd50256 712 | be2936ca 713 | 3e23747f 714 | 3c9000f3 715 | 3e58981c 716 | 3e206f03 717 | 3d213430 718 | be7008ea 719 | be67bd26 720 | be7c531d 721 | 3e6c7818 722 | 3d9d8956 723 | 3cf9b7b5 724 | be18dd45 725 | 3d4aa977 726 | 3db78579 727 | 3ea64643 728 | bdbea10e 729 | bee9ff3e 730 | be1943f0 731 | befc54ee 732 | 3dd269c5 733 | 3de7b8e5 734 | 3d95a514 735 | be61a55e 736 | 3e91a20d 737 | 3ed6eba3 738 | be214629 739 | be86d795 740 | be98adf7 741 | 3e4abc2d 742 | be9797e7 743 | 3d916d0c 744 | bc82ee4b 745 | bb7d6c48 746 | be1b5874 747 | 3d9c3f24 748 | 3e84f4b0 749 | be0e2f3b 750 | be1a3d39 751 | be23e32a 752 | 3d15090c 753 | 3e26cffe 754 | be0f3a2c 755 | 3d05c102 756 | bc40f081 757 | be3362a4 758 | be468222 759 | 3e1bfedc 760 | 3e014683 761 | 3c653949 762 | bd2a6cf5 763 | 3cf667f3 764 | bd4674df 765 | bdbf0234 766 | bbd7abf7 767 | 3d883bbd 768 | 3cbce24b 769 | 3d9b8e64 770 | be0594b2 771 | 3ed16a98 772 | bec4c752 773 | bdfccee3 774 | be52242a 775 | 3dec26f8 776 | be31bfdc 777 | 3e11fcbd 778 | 3de43c91 779 | bdddb3b7 780 | bd501b7a 781 | 3d658c68 782 | 3d39f1bd 783 | bda6ea6d 784 | 3e479229 785 | be14be18 786 | 3e058743 787 | 3e53f0a1 788 | bebb7016 789 | be14982a 790 | bcbe6125 791 | be2eb4d3 792 | be895188 793 | 3e8fcc5d 794 | be15cb99 795 | 3eb71a23 796 | 3cae6d87 797 | 3e5fab30 798 | be66b481 799 | be6312c5 800 | bedcc7db 801 | be07fbc6 802 | 3e173ce1 803 | 3dfc2c41 804 | 3e0e8938 805 | 3dcd4759 806 | bc2918f1 807 | bd895c7d 808 | bd39970b 809 | be48db34 810 | be748729 811 | 3dc04952 812 | 3dcd7d19 813 | bde91033 814 | 3e22dfd8 815 | be8436e4 816 | 3cda482a 817 | 3dca802e 818 | 3de91d60 819 | 3e07ddd1 820 | be170b5f 821 | 3c82551e 822 | 3e6919b0 823 | be584904 824 | bea5f53e 825 | bda35886 826 | be26dd2c 827 | 3f16a0ad 828 | be15e805 829 | be2cfbf3 830 | be3b0061 831 | 3d34edcf 832 | be02dbf3 833 | be9f1feb 834 | bea491a9 835 | 3e42a5c5 836 | 3e156671 837 | be3fdcae 838 | be266cb6 839 | bdd56ef6 840 | 3e06b589 841 | -------------------------------------------------------------------------------- /CNN-FPGA-Vivado/CNN-FPGA-Vivado.xpr: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 | 64 | 65 | 66 | 67 | 68 | 69 | 70 | 71 | 72 | 73 | 74 | 75 | 76 | 77 | 78 | 79 | 80 | 81 | 82 | 83 | 84 | 85 | 86 | 87 | 88 | 89 | 90 | 91 | 92 | 93 | 94 | 95 | 96 | 97 | 98 | 99 | 100 | 101 | 102 | 103 | 104 | 105 | 106 | 107 | 108 | 109 | 110 | 111 | 112 | 113 | 114 | 115 | 116 | 117 | 118 | 119 | 120 | 121 | 122 | 123 | 124 | 125 | 126 | 127 | 128 | 129 | 130 | 131 | 132 | 133 | 134 | 135 | 136 | 137 | 138 | 139 | 140 | 141 | 142 | 143 | 144 | 145 | 146 | 147 | 148 | 149 | 150 | 151 | 152 | 153 | 154 | 155 | 156 | 157 | 158 | 159 | 160 | 161 | 162 | 163 | 164 | 165 | 166 | 167 | 168 | 169 | 170 | 171 | 172 | 173 | 174 | 175 | 176 | 177 | 178 | 179 | 180 | 181 | 182 | 183 | 184 | 185 | 186 | 187 | 188 | 189 | 190 | 191 | 192 | 193 | 194 | 195 | 196 | 197 | 198 | 199 | 200 | 201 | 202 | 203 | 204 | 205 | 206 | 207 | 208 | 209 | 210 | 211 | 212 | 213 | 214 | 215 | 216 | 217 | 218 | 219 | 220 | 221 | 222 | 223 | 224 | 225 | 226 | 227 | 228 | 229 | 230 | 231 | 232 | 233 | 234 | 235 | 236 | 237 | 238 | 239 | 240 | 241 | 242 | 243 | 244 | 245 | 246 | 247 | 248 | 249 | 250 | 251 | 252 | 253 | 254 | 255 | 256 | 257 | 258 | 259 | 260 | 265 | 266 | 267 | 268 | 269 | 271 | 272 | 273 | 274 | 275 | 276 | 277 | 278 | 279 | 280 | 281 | 282 | 291 | 292 | 293 | 294 | 295 | 297 | 298 | 299 | 300 | 301 | 304 | 305 | 307 | 308 | 310 | 311 | 313 | 314 | 316 | 317 | 318 | 319 | 320 | 321 | 322 | 323 | 324 | 325 | 326 | 327 | 328 | 329 | 330 | 331 | 332 | 333 | 334 | 335 | 336 | 337 | 338 | 339 | 340 | 341 | 342 | 343 | 344 | 345 | 346 | 347 | 348 | 349 | 350 | 351 | 352 | 353 | 354 | 355 | 356 | 357 | 358 | 359 | 360 | 361 | 362 | 363 | 364 | 365 | 366 | 367 | 368 | 369 | 370 | 371 | 372 | default_dashboard 373 | 374 | 375 | 376 | --------------------------------------------------------------------------------