From 53590937fd6bb3a83d47e53a0423c791b0a71717 Mon Sep 17 00:00:00 2001
From: GreenGarnets <kmu9842@sogang.ac.kr>
Date: Fri, 24 Jan 2020 17:23:29 +0900
Subject: [PATCH] dataset cache & add STFT

---
 .gitignore                           |   4 +++
 __pycache__/dataset.cpython-36.pyc   | Bin 4221 -> 4957 bytes
 dataset.py                           |  41 +++++++++++++++++++++++++--
 net/__pycache__/model.cpython-36.pyc | Bin 1826 -> 1743 bytes
 net/model.py                         |  17 ++++-------
 train.py                             |  13 ++++-----
 6 files changed, 54 insertions(+), 21 deletions(-)
 create mode 100644 .gitignore

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..ee9a45a
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,4 @@
+/cache
+/data
+/data_test
+/model
diff --git a/__pycache__/dataset.cpython-36.pyc b/__pycache__/dataset.cpython-36.pyc
index edefce004de28f859f2fc0ffca6e543be7d0552f..e82a8262cfb54ef4e2497429c2f993c4eb284ebf 100644
GIT binary patch
delta 2176
zcmZt{OKclObY|9TuQ#^iq-p+Yt(K<VHnp9`91007YM_)N1S+CpOKWXBYiC`5wzCsj
zn+-m=;#O4HBjN-X5aQAk969p6qLq+9aEZi)Yb6Blt&=7y8Ec>4yx+Vx?=%18eqA`e
zoKEZCe|qZXFGmUagX~)h;1}TYe){lu@2z+<+sc|*N;u((=2&ao9FK6Vd8jpEPDD6u
z9yTX05uW7wJ;L<?GpFonK5~nR?`*v>!&3ul&e~IrIeT_ML4#f;Mmk`0RgKBpycMj_
zlXNBE^yw5&-6N*P(;$!uKB3pO5pd*lA~GMbALw}SBRe{m21*#0mZ;k@q>jkIeh5Y%
zCX_x+j{#wuFhUF}4h2E%rJJ)rflc%m7Za#(7(kd1w%fF-wwOdx3Q3C7mWC?u5HSV7
zOmsG-yD67;k+WhNdHYTc&Z*5l9EV4ei|~0UAxVS!WI(s5WDUiqi1`>ewc?ZzJ-4&E
zi#bo^R&z~fU35IF(Ck<|Uvrvv+iKa!TP>HDN+n}Yp^^jfwp-}gppXZD^W}0ML{60o
zrLu7b380<t8YhOl$1G!?k)eqVPg>Pm`Rm=qZsmrd@0~GZ3)|bUTssdw=FcoG=Fpxe
zis;U%at_QrR_J#@Lv}=Uqfir_R;9h!a=YNdo`fq<+bTpB_Z%@|A#1x~48x8x6Y6^g
z3>Ip;R}s?g{!51&4)xZi=Ts}Gn!1Yj9tS{Ddhl0#t`Bz{EgfJIz9=yZKz%^(9+3o7
zvj+H$EF_3F7>DFc5zgRKc<erVn{BZzMJkP$RC#tlr7Gh-!+$<Dpp&3IjmeFHM8ao&
ztPZ>_20P7Xd|ax1U*U;45aHScsj&f=tCKDD?)x&ylRj(cJUK%KY>nIyTX9eu0Vh&m
zC_V!X#7B6Fr|-wWWa0o`W{0O8z|(hl{iLrp(jbuW6Tapr>&k#OMhD;!CAUAgMk0<1
z2dIE$`6x#3s{;ny)HUKWDAYr*_QnfEZb_>mZBG`xPP^_Ed&y$EQ@dN})a%aLm)KV!
z^F;mKiLWmG-uvvlh`NuHKJ>@}T)V(LIDbNB9HIgE8uo?~s_jaxCPRIr<5r+p>vAL1
z?>Oxp+z~K?dRn1awW=F-u?{96Zy#THjBNi4%NP?0pofW_))7V6i2{<;ULTu7EFvZ9
z>-ezLvE`!*j9NZaBRo{BbuWy05<H4qu4}h>s9Hi;U6?bSrsGLqTJP9~J}gJ1h$S?A
z7C{++$-0%0TG4C>S$8Syv+h|SE<ahb*<sB_fmK@8Et`Y3HxGa)<8+oDqZu}@#ONfQ
z#~(YQ=xl=Z=B{`fYmf=gmW3xRRB`U$!uf{t5=;HzT;ez7G)^gKB33B<Z)ySOk-mbV
zR68yAx%<JwNUcAV7cT?zAO&#+7!Tm{@W?pnsq3~SSEHQw^aw8(%V}L;Ja2?lKBB8V
z1_JQJkRD6x+be)Xb@7P$W~>BjuX6GP)Wi{xFj=XzJ%ggh5Ilw8TyQeA(8oa?YQna7
z(`nmbtmd@2IE{Qvb*R5-H8<@Sgy;zIJW?_D!;(l?UX)VtERxWVQ0Y1D@HHsG)|icQ
z8pfn8T2WF%S~In3XR|FU-W?Mv#|nN<eL<&#Ub-;5?*TwtN$IrgtI}$@?*-qczoE%s
zD|4cs#j~P7&LFmo;4%UW!Jdf-8$}w9D|~l6^jxD_TuW#LZA_buIH6X7qJciC9Mqb|
z{BVM*(e#C|gZk*j7heZMrt;e5R|Jkn@eFQ9cV8Rsc<=&tqee2*Ue$5C1!zncUIhW|
y3Y`?uL^Ak$bY>Zy6YEHZp7SJkM3~&2&)zVwCu0iEXA9XBg&)&crcWG=>hwP**#>j~

delta 1492
zcmah}OOG2x5bkcj#<T0kzF$c;VcCEQu{WFLnJfgcQ4)bxD2s9!AZFIxw#Pd!dfFm3
zUMyj*NT4}z;DkU(NWdjGeghXSNVLQOaawWa5-EbJF_0A?F;ah1^;Orj`>QJdboTz~
zwTp#<dHUTmx9=|#@(Vfkq`|%dE7BjW^@r3gTV+mo)}8XEt!YZY&bc$*tTmh1)Sa{D
zuMuwW{QHFGN6cCfi`+b5@&{4uEb$3G$qOH_*O|3!t%!xr30@S-os*m_kr9J0&^=Nq
z9pUl26{?L}wsh=$S5%nJ?bynW%XBTiL?5PG@tgFCG;E%XKc%l3CM0}GWFbzo@90$g
zJzHIw1WIR>En7lq>c|54V<_I8NDXPB48##(gjA@Ujz3Oczr6$uILRw3XE0$FK<A_g
zUArOV98wcVr5s;XG7=Aw^8i*Z=&G=*F8pO|$_0!aJ2l?UxI<Jk2kS*xkqe+mhYsN0
zoS`a*^pGjmNvSl?ji^ei%z)yX8zVYTMsyL@VG4F+2Xv5XG4LpOxdG*Ql^!sjS%Doq
zJ4>2u1Yc7|Cy!q^qytpe9PT<hU*RP+i)8Da*>A4>-2eP}nMlFm*T{8%7A!(O{N&2t
z#?kNaR7{<7oqZWZwobc&%{B8wo#_i;<T=O~E0b+U%5EH>Xi?h^MWUlNn=<g~{;=zX
zy-HdeO~)0!?FmhL@;u~PtXDrS7Es#sx{=eU<57FyM<_O-)AS))WGAS}`g1QuZS0Cj
zsOp2g4Jc{yKqnI{R-*NLDFQ0r-8Fwqox`F)N3rV0zlsErd*)@=iBu!-!sp!hX0h2C
z@0ZU)@V^q|B?tttB0M=u`k8%UtL;S9zM0rtwXK3FS0MbVrs^Ku?lahM3j97Rm`9hv
zkn8bkY5vqBXr2VH*v7^YwxT?Pu!c~J_ey7n*oE4V!sf2y3!QE{K9`SS9IK$sTejO3
zSEUT3&e~xpAc```O^oL3$d=OX$%{yBA(iSo;ketM1h=v*pF%dRg!GcyXxg-lM$q+D
zJ$l>9#lMxlrVH_B<?8aW?#7h7ay+r8Y%hEx{<ZuqHR3O()`uks$wh?65uQP~iSQsh
zv5Ls$0dR04oM?j@StDmu4Kv}jQLlTp<Jar`O0~wJ@U<xLTTx90fr`SmkfQF|k@5p2
zYJSkXQw>@z=jyoOlbiG({og!sr(HdW+Lcq|zRmd9v2ruMF}*W<0e+K}dgaE;G7Fa6
zK)3YgNL0BK!JkAAbbE_?Rv~IQVXxY3st|@j;>hbv!bt<|(zstk88k!`38|SECYO}h
n>hc1@(+KAgP%fRBAJ5Z2Ws6s5s%3*t(n)5p;*eZ7U`GA`^Ppo~

diff --git a/dataset.py b/dataset.py
index ab61879..f6baf31 100644
--- a/dataset.py
+++ b/dataset.py
@@ -14,7 +14,7 @@
 
 import librosa
 import librosa.display
-import tqdm
+from tqdm import tqdm
 from scipy.fftpack import fft
 from scipy import signal
 
@@ -81,13 +81,46 @@ def targetToTensor(self, target):
     
 
     def music_load(filename) :
+        '''
         y, sr = librosa.load(filename, sr=44100)
         y_  = np.zeros(int(44100*0.04)-int(len(y)%int(44100*0.04)))
         y = np.hstack([y,y_])
 
         y = np.reshape(y, (-1, int(sr/100)*4 ))
+
         y = torch.from_numpy(y)
 
+        print(a.shape)
+        '''
+        y = np.load(filename)
+        y = torch.from_numpy(y)
+
+        return y
+
+    def music_cache_make(filelist) :
+        #f = codecs.open("music_cache_data","w")
+
+        for filename in tqdm(filelist):
+            y, sr = librosa.load("./data_test/songs/" + filename+"/nofx.ogg", sr=44100)
+            y_  = np.zeros(int(44100*0.04)-int(len(y)%int(44100*0.04)))
+            y = np.hstack([y,y_])
+
+            y = np.reshape(y, (-1, int(sr/100)*4 ))
+            #print(y.shape)
+
+            a = []
+            for i in range(0, y.shape[0]) :
+                y1 = np.abs(librosa.stft(y[i], n_fft = 1764, hop_length=2048, win_length = 441))
+                y2 = np.abs(librosa.stft(y[i], n_fft = 1764, hop_length=2048, win_length = 882))
+                y3 = np.abs(librosa.stft(y[i], n_fft = 1764, hop_length=2048, win_length = 1764))
+
+                a.append(np.array([y1,y2,y3]).tolist())
+                #if i == 200 : print(np.array([y1,y2,y3]))
+
+            a = np.array(a)
+            #print(a.shape)
+            np.save("./cache/"+filename,a)
+
         return y
 
     def timeStamp(filename, term) :
@@ -220,6 +253,8 @@ def timeStamp(filename, term) :
         return return_timestamp
 
 if __name__ == "__main__":
-    y, sr = librosa.load("./data/songs/rootsphere_lastnote/nofx.ogg", sr=44100)
-    KshDataset.timeStamp("./data/songs/rootsphere_lastnote/exh.ksh", y.shape[0])
+    filenames = os.listdir("./data_test/songs/")
+    KshDataset.music_cache_make(filenames)
+    #y, sr = librosa.load("./data/songs/rootsphere_lastnote/nofx.ogg", sr=44100)
+    #KshDataset.timeStamp("./data/songs/rootsphere_lastnote/exh.ksh", y.shape[0])
     #print(y.shape[0]//441)
\ No newline at end of file
diff --git a/net/__pycache__/model.cpython-36.pyc b/net/__pycache__/model.cpython-36.pyc
index 8870c40aff645d22e081341bc11f3a5568158db2..6536f1042fc5ebb48c9318c2152a3304a1cde076 100644
GIT binary patch
delta 739
zcma)3&ubGw6rMLT`=iZn8jDydjR~k|EY^b%{0}T11nJ#z-5G7-wy854Yb7?qo-363
z2SmZMc=lfq?{oE(zd-QRHwh#j3Uy)L%zW?rzW3hj=lZX7{eGG*KlpOz>9|J7Z}QI<
zoUR|GwU<y)AQE;+CprD3Z+rwoN>ajzRJ2bk0;FVpjw_HLDJ~|N?ZVAqc6tVDEkHCl
zp%otyu{>wfIU6|)IrbZ@ji)NfPXgqHGOPgisPZ+I&}HG1%zcnGtNI_R^di;x++y{M
zWap|@P}RWZqKe|+mn6X_j2}-JKJny(5BQR#!(8o^()+3qruR<SKeQQ5xKFIwEp}{w
zu;;LCTl|qVd~@{}6C>!8ioC}k(1#J&@BHX;=-sFn)m&s=50yfvZg73uMOvBN>_9p0
z-Svj&Fx|w>!CS`$!&0~KG7D`FM=u8YFw?S&7jp#zftcE_!ISYVpH79yi_Dl@6@?HB
zaR>A@%<Hxfenj7s@q@%P*_t9Q7Gf{!?F->}K9reb`N2?o%EX{Tdj&S<0e#PhDbR31
rInep8vqiX5zm2#HrFzrLmA-3#hRw#J|LLi1|AwulhcR4*geEIzg5jq=

delta 690
zcmaJ-&ubGw6n<}JcDLE?Ce?;Uf|?qC;3Bj_K?D!I2%@M^y}J!N+cqIIWj6iE`UB?N
zL!E!1c<Hs!f1qbCp1kBBqF(Y}_$Dby6mekQyqWL&-Z$@k2|pF9OJV5$$fGw;0)U?|
zvBXhzKlIm8GX#mRA(}oqR=3U(pbK5hL9-UqfY9YFLEDiindB7FJiyz|Fk8h+jnk`h
z$h62I_D5_sViSiG$DCs2Lbb{MR0SiIE8Po^zd7N*IqxUkixV7gSVxLOhhAOjX{6KB
zwx0rF<3E>1Do_4TH6y1q4!XQVip#Pfea$M=^sEFszL$NP9rQogz|9$3tTurGtPf}Y
zeYpvO5ZVxxO`p9hd5=|RtFzthIZ0=8=b>5S3BEF)`LmCrfJIXl_sX_d4lT%0<6A}<
z=FF^!{b^gKPu0oSd=occsH>FsJe6h}jjgmNvgSdmu3Ma|nuTz$$H`0Gpvy4S2a2_h
z+HENLh%VV21z<j!r_RcqyA}@PIB7SsEJ@q#I39~5RP%(Zx;b~sxMaRM^WiU-+?cXI
pQ}<2at`;L(!j<VxGfCBgIdHGdjO_<m-JH0!!t*J-f&mLkzX5W!n0EjG

diff --git a/net/model.py b/net/model.py
index 34b19a8..2b65862 100644
--- a/net/model.py
+++ b/net/model.py
@@ -10,7 +10,7 @@ def __init__(self):
         super(voltexNet, self).__init__()        
 
         self.conv1 = nn.Sequential(
-            nn.Conv1d(1, 128, kernel_size=3, stride=1, padding=1),
+            nn.Conv1d(3, 128, kernel_size=3, stride=1, padding=1),
             nn.BatchNorm1d(128),
             nn.ReLU())
         self.conv2 = nn.Sequential(
@@ -39,17 +39,12 @@ def __init__(self):
             nn.ReLU(),
             nn.MaxPool1d(3,stride=3))
         self.conv7 = nn.Sequential(
-            nn.Conv1d(256, 512, kernel_size=3, stride=1, padding=1),
-            nn.BatchNorm1d(512),
+            nn.Conv1d(256, 256, kernel_size=3, stride=1, padding=1),
+            nn.BatchNorm1d(256),
             nn.ReLU(),
             nn.MaxPool1d(3,stride=3))
-        self.conv8 = nn.Sequential(
-            nn.Conv1d(512, 512, kernel_size=3, stride=1, padding=1),
-            nn.BatchNorm1d(512),
-            nn.ReLU(),
-            nn.Dropout(0.5))
 
-        self.fc = nn.Linear(1024, 4)
+        self.fc = nn.Linear(256, 4)
         
         #self.LSTM = nn.LSTM(input_size = 4, hidden_size = 2, bidirectional=True)
         #self.fc3 = nn.Linear(hidden_size*2,output_size)
@@ -57,6 +52,7 @@ def __init__(self):
 
     
     def forward(self, x):
+        x = x.squeeze()
 
         out = self.conv1(x)
         out = self.conv2(out)
@@ -65,7 +61,6 @@ def forward(self, x):
         out = self.conv5(out)
         out = self.conv6(out)
         out = self.conv7(out)
-        out = self.conv8(out)
         
         out = out.reshape(out.size(0), out.size(1)* out.size(2))
         #print(out.shape)
@@ -83,7 +78,7 @@ def forward(self, x):
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # PyTorch v0.4.0
     model = voltexNet().to(device)
 
-    summary(model, [(1, 1764)])
+    summary(model, [(3, 883, 1)])
 
 '''
 class VoltexLSTM
diff --git a/train.py b/train.py
index 28e41a8..d127251 100644
--- a/train.py
+++ b/train.py
@@ -71,7 +71,7 @@ def infer(model, device, batch, filename, savename) :
 def main():
     
     model = voltexNet()
-    model.load_state_dict(torch.load("./model_99_.pth"))
+    #model.load_state_dict(torch.load("./model_99_.pth"))
     #print ("load model")
     
     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
@@ -97,7 +97,7 @@ def main():
     filenames = os.listdir(dirname)
     valnames = os.listdir(valname)
 
-    batch = 256
+    batch = 128
     song_index = 0
     best_Acc = 0
 
@@ -114,9 +114,9 @@ def main():
             full_filename = os.path.join(dirname, filename)
             #print(full_filename)
 
-            input = KshDataset.music_load(full_filename + '/nofx.ogg')
+            input = KshDataset.music_load("./cache/" + filename+'.npy')
             #print(input.shape)
-            input = input.reshape(input.shape[0], 1, -1)       
+            #input = input.reshape(input.shape[0], 1, -1)       
             target = KshDataset.timeStamp(full_filename + '/exh.ksh', input.shape[0]) 
             try : 
                 if target == None:
@@ -148,7 +148,7 @@ def main():
                     optimizer.zero_grad()
                     tmp_batch = input.shape[0] - i
                     if tmp_batch > 1 : 
-                        pred = model(input[i:i+tmp_batch-1],tmp_batch-1)
+                        pred = model(input[i:i+tmp_batch-1])
                         
                         loss = criterion(pred.squeeze(), target[i:i+tmp_batch-1].squeeze())
                         loss.backward()
@@ -165,8 +165,7 @@ def main():
         #model.to(torch.device("cpu"))
         for filename in tqdm(valnames):
             full_filename = os.path.join(valname, filename)
-            input = KshDataset.music_load(full_filename + '/nofx.ogg')
-            input = input.reshape(input.shape[0], 1, -1)    
+            input = KshDataset.music_load("./cache/" + filename+'.npy')
             input = input.to(device, dtype=torch.float)   
             target = KshDataset.timeStamp(full_filename + '/exh.ksh', input.shape[0])