From f7fcff31286da48aca1d1b816c5d1d57326b240e Mon Sep 17 00:00:00 2001
From: peter94135 <84705600+peter94135@users.noreply.github.com>
Date: Sat, 15 Apr 2023 13:01:23 +0800
Subject: [PATCH 1/3] Update README.md

---
 Improvement/README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Improvement/README.md b/Improvement/README.md
index 2452420..d3da2c9 100644
--- a/Improvement/README.md
+++ b/Improvement/README.md
@@ -7,7 +7,7 @@ Since the user must convert the codeword into binary, it can be sent back to the
    
    Binary codeword = Quantized(codeword*2-1)
 
-2. tanh：- 1 < codeword < 1
+2. tanh：-1 < codeword < 1
    
    Binary codeword = Quantized(codeword)
 
@@ -34,6 +34,6 @@ The following are the simulation results of two different activation functions a
 From the above results, we can see that in most scenarios, using tanh for the activation function can bring better results. I think that the corresponding domain of tanh is larger than that of sigmoid, which means that the gradient is also stronger.
 
 ## Reference
-[1] Huang, Tzu-Hao. "5G downlink channel feedback technology based on deep learning"
+[1] T.-H. Huang, "5G downlink channel feedback technology based on deep learning," M.S. thesis, Comm. Eng., National Sun Yat-sen Univ., Kaohsiung, ＴTaiwan, 2023.
 
 [2] MathWorks CSI Feedback with Autoencoders. https://www.mathworks.com/help/5g/ug/csi-feedback-with-autoencoders.html

From cba32d1c1a53731e3094f96e3ad0a27f20337764 Mon Sep 17 00:00:00 2001
From: peter94135 <84705600+peter94135@users.noreply.github.com>
Date: Sat, 15 Apr 2023 13:03:16 +0800
Subject: [PATCH 2/3] Update README.md

---
 Improvement/README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Improvement/README.md b/Improvement/README.md
index d3da2c9..06aa4d6 100644
--- a/Improvement/README.md
+++ b/Improvement/README.md
@@ -34,6 +34,6 @@ The following are the simulation results of two different activation functions a
 From the above results, we can see that in most scenarios, using tanh for the activation function can bring better results. I think that the corresponding domain of tanh is larger than that of sigmoid, which means that the gradient is also stronger.
 
 ## Reference
-[1] T.-H. Huang, "5G downlink channel feedback technology based on deep learning," M.S. thesis, Comm. Eng., National Sun Yat-sen Univ., Kaohsiung, ＴTaiwan, 2023.
+[1] T.-H. Huang, "5G downlink channel feedback technology based on deep learning," M.S. thesis, Comm. Eng., National Sun Yat-sen Univ., Kaohsiung, Taiwan, 2023.
 
 [2] MathWorks CSI Feedback with Autoencoders. https://www.mathworks.com/help/5g/ug/csi-feedback-with-autoencoders.html

From 9959a6e42ac27cf4336e855c8dd88f4080cd5864 Mon Sep 17 00:00:00 2001
From: peter94135 <84705600+peter94135@users.noreply.github.com>
Date: Sun, 16 Apr 2023 11:46:51 +0800
Subject: [PATCH 3/3] Add files via upload

---
 Improvement/plotNetwork.m              | 24 ++++++++++++++++++
 Improvement/plot_Quantization_Result.m | 35 ++++++++++++++++++++++++++
 2 files changed, 59 insertions(+)
 create mode 100644 Improvement/plotNetwork.m
 create mode 100644 Improvement/plot_Quantization_Result.m

diff --git a/Improvement/plotNetwork.m b/Improvement/plotNetwork.m
new file mode 100644
index 0000000..e25e7e4
--- /dev/null
+++ b/Improvement/plotNetwork.m
@@ -0,0 +1,24 @@
+function plotNetwork(net,encNet,decNet)
+%plotNetwork Plot autoencoder network
+%   plotNetwork(NET,ENC,DEC) plots the full autoencoder network together
+%   with encoder and decoder networks.
+fig = figure;
+t1 = tiledlayout(1,2,'TileSpacing','Compact');
+t2 = tiledlayout(t1,1,1,'TileSpacing','Tight');
+t3 = tiledlayout(t1,2,1,'TileSpacing','Tight');
+t3.Layout.Tile = 2;
+nexttile(t2)
+plot(net)
+title("Autoencoder")
+nexttile(t3)
+plot(encNet)
+title("Encoder")
+nexttile(t3)
+plot(decNet)
+title("Decoder")
+pos = fig.Position;
+pos(3) = pos(3) + 200;
+pos(4) = pos(4) + 300;
+pos(2) = pos(2) - 300;
+fig.Position = pos;
+end
\ No newline at end of file
diff --git a/Improvement/plot_Quantization_Result.m b/Improvement/plot_Quantization_Result.m
new file mode 100644
index 0000000..6395207
--- /dev/null
+++ b/Improvement/plot_Quantization_Result.m
@@ -0,0 +1,35 @@
+clc
+clear
+close all
+
+envir = 'Indoor'; %'Indoor' or 'Outdoor'
+encoded_dim = 32;  %compress rate=1/4->dim.=512, compress rate=1/16->dim.=128, compress rate=1/32->dim.=64, compress rate=1/64->dim.=32
+model_name = "model_CsiNet_"+envir+"_dim"+num2str(encoded_dim);
+
+load("Sigmoid result\"+model_name+"_Quantization_Result.mat")
+rho_all_sigmoid=rho_all;
+NMSE_all_sigmoid=NMSE_all;
+load("Tanh result\"+model_name+"_Quantization_Result.mat")
+rho_all_tanh=rho_all;
+NMSE_all_tanh=NMSE_all;
+
+figure
+tiledlayout(2,1)
+nexttile
+plot(nBitsVec,rho_all_sigmoid,'*-','LineWidth',1.5)
+hold on
+plot(nBitsVec,rho_all_tanh,'*-','LineWidth',1.5)
+legend("Sigmoid","Tanh")
+title("Correlation ("+envir+" Codeword-" + encoded_dim + ")")
+xlabel("Number of Quantization Bits"); ylabel("\rho")
+grid on
+nexttile
+plot(nBitsVec,NMSE_all_sigmoid,'*-','LineWidth',1.5)
+hold on
+plot(nBitsVec,NMSE_all_tanh,'*-','LineWidth',1.5)
+legend("Sigmoid","Tanh")
+title("NMSE ("+envir+" Codeword-" + encoded_dim + ")")
+xlabel("Number of Quantization Bits"); ylabel("NMSE (dB)")
+grid on
+savefig(gcf,"Compare result\"+envir+"_"+num2str(encoded_dim)+"_Quantized.fig")
+saveas(gcf,"Compare result\"+envir+"_"+num2str(encoded_dim)+"_Quantized.jpg")
\ No newline at end of file