simple vae style model

data.py

@@ -22,15 +22,9 @@ def process_file(file_path):
         end = start_pos + size
         if end <= sample_len:
             chunk = y[start_pos:end]
-            chunk = librosa.feature.melspectrogram(y=chunk, sr=SAMPLE_RATE)
-            chunk = ((librosa.amplitude_to_db(chunk,ref=np.max)+80)/80)
-
-            #chunk = librosa.feature.melspectrogram(y=chunk,sr=SAMPLE_RATE)
-            #chunk = ((librosa.amplitude_to_db(chunk,ref=np.max)+40)/40)
             file_chunks.append(chunk)
     return file_chunks
 
-#@mlflow.trace
 def load():
     """
     Load 10 second chunks of songs.
@@ -44,9 +38,6 @@ def load():
         audio.extend(l)
     return audio
 
-
-
-##DEP
 def audio_split(audio):
     """
     Split 10 seconds of audio to 2 5 second clips

dataInit.py

@@ -1,14 +1,14 @@
 import data
 import numpy as np
 
-x = data.load()
+x,y = data.dataset(data.load())
 size=len(x)
 print(size)
 
 x_np = np.stack(x)
 x_np = np.expand_dims(x_np, axis=1)
 
-#y_np = np.stack(y)
+y_np = np.stack(y)
-#y_np = np.expand_dims(y_np, axis=1)
+y_np = np.expand_dims(y_np, axis=1)
 
-np.savez_compressed("data",x_np)
+np.savez_compressed("data",x_np,y_np)

model.py

@@ -1,40 +1,39 @@
 from tinygrad import Tensor, nn
 
 class gen:
-    def __init__(self, input_channels=1, height=128, width=431, latent_dim=64):
+    def __init__(self, input_channels=1, height=128, width=216, latent_dim=1024):
         self.height = height
         self.width = width
         self.latent_dim = latent_dim
 
-        self.w = width // 4
+        self.w = width // 8
-        self.h = height // 4
+        self.h = height // 8
-        self.h = 32  # Output height after 2 strides
+        self.flattened_size = 256 * self.h * self.w
-        self.w = 108 # Output width after 2 strides
-        self.flattened_size = 128 * self.h * self.w
 
 
         self.e1 = nn.Conv2d(input_channels, 64, kernel_size=3, stride=2, padding=1)
         self.e2 = nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1)
+        self.e3 = nn.Conv2d(128,256, kernel_size=3,stride=2,padding=1)
+
 
         self.el = nn.Linear(self.flattened_size, self.latent_dim)
 
-        self.q = nn.Linear(self.latent_dim,self.latent_dim)
-        self.k = nn.Linear(self.latent_dim,self.latent_dim)
-        self.v = nn.Linear(self.latent_dim,self.latent_dim)
 
         self.dl = nn.Linear(self.latent_dim, self.flattened_size)
 
-        self.d1 = nn.ConvTranspose2d(128, 64, kernel_size=3, stride=2, padding=1, output_padding=1)
+        self.d1 = nn.ConvTranspose2d(256,128,kernel_size=3,stride=2,padding=1,output_padding=1)
-        self.d2 = nn.ConvTranspose2d(64, input_channels, kernel_size=3, stride=2, padding=1, output_padding=1)
+        self.d2 = nn.ConvTranspose2d(128, 64, kernel_size=3, stride=2, padding=1, output_padding=1)
+        self.d3 = nn.ConvTranspose2d(64, input_channels, kernel_size=3, stride=2, padding=1, output_padding=1)
 
     def __call__(self, x: Tensor) -> Tensor:
         y, shape = self.encode(x)
-        z = self.atten(y)
+        z = y#self.atten(y)
         return self.decode(z, shape)
 
     def encode(self, x: Tensor):
         x = self.e1(x).leakyrelu()
         x = self.e2(x).leakyrelu()
+        x = self.e3(x).leakyrelu()
         b, c, h, w = x.shape
 
         flattened_size = c * h * w
@@ -52,9 +51,10 @@ class gen:
 
     def decode(self, z: Tensor, shape):
         x = self.dl(z).leakyrelu()
-        x = x.reshape(shape=(-1, 128, self.h, self.w))
+        x = x.reshape(shape=(-1, 256, self.h, self.w))
         x = self.d1(x).leakyrelu()
-        x = self.d2(x).sigmoid()
+        x = self.d2(x).leakyrelu()
+        x = self.d3(x).sigmoid()
 
         # Crop or pad to match input size
         out_h, out_w = x.shape[2], x.shape[3]

train.py

@@ -8,27 +8,25 @@ from model import gen
 
 BATCH_SIZE = 16
 EPOCHS = 100
-LEARNING_RATE = 1e-5
+LEARNING_RATE = 3e-4
 print(Device.DEFAULT)
 mdl = gen()
 opt = nn.optim.AdamW(nn.state.get_parameters(mdl), lr=LEARNING_RATE)
-volume = 0.1
 
 def spec_loss(pred, target, eps=1e-6):
     # spectral convergence
     sc = ((target - pred).square().sum()) ** 0.5 / ((target.square().sum()) ** 0.5 + eps)
     # log magnitude difference
     log_mag = ((target.abs() + eps).log() - (pred.abs() + eps).log()).abs().mean()
-    return sc + log_mag
+    return 0.1*sc + 1.0*log_mag + 0.1*(pred - target).abs().mean()
 
 
 @TinyJit
-def step_gen(x):
+def step_gen(x,y):
     Tensor.training = True
-    noise = Tensor.rand_like(x).tanh()
+    z = mdl(x)
-    y = x+(noise*volume)
+    loss = spec_loss(z,y)
-    y = y.clamp(0,1)
+    #loss = (y - z).abs().mean()
-    loss = spec_loss(mdl(y),x)
     opt.zero_grad()
     loss.backward()
     opt.step()
@@ -36,8 +34,8 @@ def step_gen(x):
 
 print("loading")
 x = np.load("data.npz")["arr_0"]
-#x= x[0:64]
+y = np.load("data.npz")["arr_1"]
-run_name = f"tinygrad_autoencoder_{int(time.time())}"
+run_name = f"vae_{int(time.time())}"
 mlflow.set_tracking_uri("http://127.0.0.1:5000")
 mlflow.start_run()
 mlflow.log_params({"batch_size": BATCH_SIZE, "epochs": EPOCHS, "lr": LEARNING_RATE, "data size":len(x)})
@@ -52,20 +50,18 @@ for epoch in range(0,EPOCHS):
     loss=0
     for i in range(0,len(x),BATCH_SIZE):
         tx=Tensor(x[i:i+BATCH_SIZE])
+        ty=Tensor(y[i:i+BATCH_SIZE])
         if(tx.shape != eshape):
             continue
-        loss += step_gen(tx)
+        loss += step_gen(tx,ty)
 
     loss /= (len(x)/BATCH_SIZE)
     if epoch%5==0:
-        noise = Tensor.rand_like(Tensor(x[0:1])).tanh()
+        show.logSpec(mdl(Tensor(x[0:1])).numpy()[0][0],epoch)
-        y = Tensor(x[0:1]) + (noise*volume)
+    if epoch%15==0:
-        show.logSpec(mdl(y).numpy()[0][0],epoch)
+        state_dict = get_state_dict(mdl)
-        if(pl - loss < 0.03 and epoch > 25):
+        safe_save(state_dict, f"model_{epoch}.safetensors")
-            show.logSpec(y.numpy()[0][0],f"volume_{volume}")
+        show.logSpec(mdl(mdl(mdl(Tensor(y[0:1])))).numpy()[0][0],f"deep_{epoch}")
-            volume *= 2
-    pl = loss
 
-    mlflow.log_metric("volume", volume, step=epoch)
     mlflow.log_metric("loss", loss, step=epoch)
     print(f"loss of {loss}")