dataset cache & add STFT

.gitignore

@@ -0,0 +1,4 @@
+/cache
+/data
+/data_test
+/model

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)

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(
@@ -39,24 +39,20 @@ 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)
         #self.output = F.softmax(linear(output),1)
 
 
     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

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])