ignore safetensors

.gitignore

@@ -3,3 +3,4 @@
 /data/
 /music.safetensors
 /data.npz
+*.safetensors

data.py

@@ -6,7 +6,15 @@ import mlflow
 
 SAMPLE_RATE = 22050
 
-#@mlflow.trace
+def spec_to_audio(spec):
+    """
+    Convert a normalized mel-spectrogram back to audio.
+    """
+    spec = (spec * 80) - 80
+    spec = librosa.db_to_amplitude(spec)*80
+    audio = librosa.feature.inverse.mel_to_audio(spec,sr=SAMPLE_RATE)
+    return audio
+
 def process_file(file_path):
     """
     Load 10 second chunks single song.
@@ -22,15 +30,9 @@ 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 +46,6 @@ 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

@@ -1,31 +1,41 @@
 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.num_tokens = 16
+        self.dim_per_token = self.latent_dim // self.num_tokens
 
 
         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.q = nn.Linear(self.dim_per_token,self.dim_per_token)
-        self.k = nn.Linear(self.latent_dim,self.latent_dim)
+        self.k = nn.Linear(self.dim_per_token,self.dim_per_token)
-        self.v = nn.Linear(self.latent_dim,self.latent_dim)
+        self.v = nn.Linear(self.dim_per_token,self.dim_per_token)
+        self.norm1 = nn.LayerNorm(self.dim_per_token)
+
+        ffn_dim = self.dim_per_token * 4
+        self.ffn1 = nn.Linear(self.dim_per_token, ffn_dim)
+        self.ffn2 = nn.Linear(ffn_dim, self.dim_per_token)
+        self.norm2 = nn.LayerNorm(self.dim_per_token)
 
         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)
@@ -35,26 +45,37 @@ class gen:
     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
-
-
         x = x.reshape(shape=(b, flattened_size))
         z = self.el(x)
+
+        # reshape to multi-token: (batch, num_tokens, dim_per_token)
+        z = z.reshape(shape=(b, self.num_tokens, self.dim_per_token))
         return z, (c, h, w)
 
     def atten(self, x: Tensor):
-        q = self.q(x).relu()
+        q = self.q(x)
-        k = self.k(x).relu()
+        k = self.k(x)
-        v = self.v(x).relu()
+        v = self.v(x)
-        return q.scaled_dot_product_attention(k,v)
+        attn = q.scaled_dot_product_attention(k, v)
+        x = self.norm1(x+attn)
+
+        ffn = self.ffn1(x).relu()
+        ffn = self.ffn2(ffn)
+        x = self.norm2(x+ffn)
+
+        return x
 
     def decode(self, z: Tensor, shape):
+        z = z.reshape(shape=(z.shape[0], -1))
         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]

run.py

@@ -0,0 +1,43 @@
+import numpy as np
+import random
+import time
+from tinygrad import Tensor, nn
+from tinygrad.nn.state import safe_load, load_state_dict
+import librosa
+import sounddevice as sd
+from model import gen
+from data import spec_to_audio
+
+SAMPLE_RATE = 22050
+
+def load_model(filepath="model.safetensors"):
+    """Loads the model structure and weights."""
+    model = gen()
+    state_dict = safe_load(filepath)
+    load_state_dict(model, state_dict)
+    return model
+
+def load_data(filepath="data.npz"):
+    """Loads the pre-processed spectrogram data."""
+    print(f"Loading data from {filepath}...")
+    data = np.load(filepath)
+    x = data["arr_0"]
+    return x
+
+def play_spec(spec,i):
+    """Converts a spectrogram numpy array to audio and plays it."""
+    audio = spec_to_audio(spec)
+    sd.wait()
+    print(f"chunk:{i}")
+    sd.play(audio, samplerate=SAMPLE_RATE)
+
+def run_prediction_loop(model, data_x):
+    current_spect = data_x[0:1]
+    for i in range(10):
+        play_spec(current_spect[0][0],i)
+        current_spect = model(Tensor(current_spect)).numpy()
+
+if __name__ == "__main__":
+    model = load_model()
+    data_x = load_data()
+    run_prediction_loop(model, data_x)

shell.nix

@@ -0,0 +1,8 @@
+{pkgs ? import <nixpkgs> {}}:
+with pkgs;
+  mkShell rec {
+    packages = [python3 jupyter-all python3Packages.librosa python3Packages.tinygrad python3Packages.numpy python3Packages.mlflow python3Packages.tqdm python3Packages.sounddevice];
+    nativeBuildInputs = [];
+    buildInputs = [];
+    LD_LIBRARY_PATH = lib.makeLibraryPath buildInputs;
+  }

train.py

@@ -1,34 +1,34 @@
 import mlflow
 import numpy as np
 from tinygrad import Device,Tensor,nn,TinyJit
+from tinygrad.nn.state import safe_save, get_state_dict
 import matplotlib.pyplot as plt
 import time
 import show
 from model import gen
+from tqdm import tqdm
 
 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 +36,8 @@ 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)})
@@ -45,27 +45,28 @@ mlflow.log_params({"batch_size": BATCH_SIZE, "epochs": EPOCHS, "lr": LEARNING_RA
 show.logSpec(Tensor(x[0:1]).numpy()[0][0],"default")
 
 print("training")
-pl = 0
+eshape = (BATCH_SIZE, 1, 128, 216)
-eshape = (BATCH_SIZE, 1, 128, 431)
 for epoch in range(0,EPOCHS):
     print(f"\n--- Starting Epoch {epoch} ---\n")
     loss=0
-    for i in range(0,len(x),BATCH_SIZE):
+    for i in tqdm(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}")
+
+show.logSpec(mdl(Tensor(x[0:1])).numpy()[0][0],EPOCHS)
+state_dict = get_state_dict(mdl)
+safe_save(state_dict, "model.safetensors")