Modify input of GAN

src/constants.py

@@ -20,6 +20,7 @@
 
 SAMPLE_STEPS = 1
 CKPT_STEPS = 100
+INPUT_FEATURES = 3
 
 MIN_FREQ_NOTE = 27.5
 MAX_FREQ_NOTE = 4186.0
@@ -35,6 +36,7 @@
 EPOCHS = 100
 BATCH_SIZE = 5
 MAX_POLYPHONY = 12
+NORMALIZE_FREQ = False
 SAMPLE_TIMES = 50
 
 # Generator
@@ -43,7 +45,7 @@
 L2_G = 0.25
 HIDDEN_DIM_G = 30
 BIDIRECTIONAL_G = False
-PRETRAIN_G = 0
+PRETRAIN_G = 10
 TYPE_G = 'LSTM'
 LAYERS_G = 1
 

src/dataset/MusicDataset.py

@@ -5,7 +5,7 @@
 import torch
 from torch.utils.data import Dataset, DataLoader
 
-from constants import DATASET_PATH, BATCH_SIZE, MAX_POLYPHONY
+from constants import DATASET_PATH, BATCH_SIZE, INPUT_FEATURES
 from utils.tensors import use_cuda
 from utils.typings import File, FloatTensor
 
@@ -46,7 +46,7 @@ def _apply_padding(self) -> List[FloatTensor]:
         """
         padded_songs = []
         for song in self.songs:
-            padded_song = torch.zeros((self.longest_song, MAX_POLYPHONY), dtype=torch.float)
+            padded_song = torch.zeros((self.longest_song, INPUT_FEATURES), dtype=torch.float)
             padded_song[:len(song), :] = torch.tensor(song)
             padded_songs.append(padded_song)
         return padded_songs

src/dataset/preprocessing/parser.py

@@ -6,7 +6,7 @@
 from py_midicsv import midi_to_csv
 from tqdm import tqdm
 
-from constants import RAW_DATASET_PATH, DATASET_PATH, MAX_POLYPHONY
+from constants import RAW_DATASET_PATH, DATASET_PATH, MAX_POLYPHONY, NORMALIZE_FREQ
 from dataset.Music import Song, Track, NoteData
 
 
@@ -66,7 +66,6 @@ def csv_to_series(song: Song) -> List[Tuple[float, int, int]]:
     """
     ts_data = []
     last_start = None
-    note_max_start = song.max_time
 
     # Index of current treated element for each track as those are already sorted.
     track_time_indices: Union[int, None] = [0] * song.number_tracks
@@ -75,6 +74,7 @@ def csv_to_series(song: Song) -> List[Tuple[float, int, int]]:
 
     while True:
         # Get first note to be played
+        note_max_start = song.max_time
         first_note_idx = None
         for track_idx, note in enumerate(current_notes):
             if note is not None and note.note_start < note_max_start:
@@ -87,7 +87,8 @@ def csv_to_series(song: Song) -> List[Tuple[float, int, int]]:
 
         # Add id to the time series data
         time_since_last = first_note.note_start - last_start if last_start is not None else 0
-        ts_data.append((first_note.norm_freq, first_note.duration, time_since_last))
+        ts_data.append((first_note.norm_freq if NORMALIZE_FREQ else first_note.freq,
+                        first_note.duration, time_since_last))
         last_start = first_note.note_start
 
         # Update that note and discard finished tracks

src/dataset/preprocessing/reconstructor.py

@@ -44,7 +44,10 @@ def parse_data(notes_data: NDArray) -> str:
     csv_data_tracks = [[f'{idx + 2}, 0, Start_track'] for idx in range(MAX_POLYPHONY)]
 
     for note_data in notes_data:
-        note = freq_to_note(float(note_data[0]) * MAX_FREQ_NOTE)
+        note = freq_to_note(float(note_data[0]) * (MAX_FREQ_NOTE if MAX_FREQ_NOTE else 1))
+        # if note == 0:
+        #    continue
+
         duration = int(note_data[1])
         time_since_previous = int(note_data[2])
 

src/model/Trainer.py

@@ -63,7 +63,7 @@ def train(self):
             metric.print_metrics()
 
             if epoch % SAMPLE_STEPS == 0:
-                sample = self.generate_sample(length=500)
+                sample = self.generate_sample(length=5000)
                 reconstruct_midi(title=f'Sample_{epoch}', raw_data=sample)
                 # TODO: Visdom doesn't accept MIDI files.
                 #  Should convert to WAV or find an alternative for visualization.
@@ -86,7 +86,7 @@ def generate_sample(self, length: int) -> NDArray:
             noise_data = GANGenerator.noise((1, length))
             sample_data = self.model.generator(noise_data)
             # sample_notes = sample_data.argmax(2)
-            return sample_data.view(-1, MAX_POLYPHONY).cpu().numpy()
+            return sample_data.view(-1, 3).cpu().numpy()
 
     def _pretrain_epoch(self, epoch: int) -> EpochMetric:
         """

src/model/gan/GANDiscriminator.py

@@ -2,7 +2,7 @@
 from torch import nn
 from torch.nn import functional as F
 
-from constants import LAYERS_D, HIDDEN_DIM_D, BIDIRECTIONAL_D, TYPE_D, MAX_POLYPHONY
+from constants import LAYERS_D, HIDDEN_DIM_D, BIDIRECTIONAL_D, TYPE_D, INPUT_FEATURES
 from model.gan.RNN import RNN
 from utils.tensors import device
 
@@ -17,7 +17,7 @@ def __init__(self):
         super(GANDiscriminator, self).__init__()
 
         self.rnn = RNN(architecture=TYPE_D,
-                       inp_dim=MAX_POLYPHONY,
+                       inp_dim=INPUT_FEATURES,
                        hid_dim=HIDDEN_DIM_D,
                        layers=LAYERS_D,
                        bidirectional=BIDIRECTIONAL_D).to(device)

src/model/gan/GANGenerator.py

@@ -4,7 +4,7 @@
 from torch import nn
 from torch.nn import functional as F
 
-from constants import HIDDEN_DIM_G, LAYERS_G, BIDIRECTIONAL_G, TYPE_G, MAX_POLYPHONY
+from constants import HIDDEN_DIM_G, LAYERS_G, BIDIRECTIONAL_G, TYPE_G, INPUT_FEATURES
 from model.gan.RNN import RNN
 from utils.tensors import device
 
@@ -25,8 +25,8 @@ def __init__(self):
                        bidirectional=BIDIRECTIONAL_G).to(device)
 
         dense_input_features = (2 if BIDIRECTIONAL_G else 1) * HIDDEN_DIM_G
-        self.dense_1 = nn.Linear(in_features=dense_input_features, out_features=2 * MAX_POLYPHONY)
-        self.dense_2 = nn.Linear(in_features=2 * MAX_POLYPHONY, out_features=MAX_POLYPHONY)
+        self.dense_1 = nn.Linear(in_features=dense_input_features, out_features=2 * INPUT_FEATURES)
+        self.dense_2 = nn.Linear(in_features=2 * INPUT_FEATURES, out_features=INPUT_FEATURES)
 
     def forward(self, x):
         x, _ = self.rnn(x, )