Spaces:
Running
on
Zero
Running
on
Zero
File size: 9,277 Bytes
05d6e12 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 |
import torch
import torch.nn.functional as F
from torch import nn
from onsets_and_frames.constants import MAX_MIDI, MIN_MIDI, N_KEYS
from .lstm import BiLSTM
from .mel import melspectrogram
class ConvStack(nn.Module):
def __init__(self, input_features, output_features):
super().__init__()
# input is batch_size * 1 channel * frames * input_features
self.cnn = nn.Sequential(
# layer 0
nn.Conv2d(1, output_features // 16, (3, 3), padding=1),
nn.BatchNorm2d(output_features // 16),
nn.ReLU(),
# layer 1
nn.Conv2d(output_features // 16, output_features // 16, (3, 3), padding=1),
nn.BatchNorm2d(output_features // 16),
nn.ReLU(),
# layer 2
nn.MaxPool2d((1, 2)),
nn.Dropout(0.25),
nn.Conv2d(output_features // 16, output_features // 8, (3, 3), padding=1),
nn.BatchNorm2d(output_features // 8),
nn.ReLU(),
# layer 3
nn.MaxPool2d((1, 2)),
nn.Dropout(0.25),
)
self.fc = nn.Sequential(
nn.Linear((output_features // 8) * (input_features // 4), output_features),
nn.Dropout(0.5),
)
def forward(self, mel):
x = mel.view(mel.size(0), 1, mel.size(1), mel.size(2))
x = self.cnn(x)
x = x.transpose(1, 2).flatten(-2)
x = self.fc(x)
return x
class OnsetsAndFrames(nn.Module):
def __init__(
self,
input_features,
output_features,
model_complexity=48,
onset_complexity=1,
n_instruments=13,
):
nn.Module.__init__(self)
model_size = model_complexity * 16
sequence_model = lambda input_size, output_size: BiLSTM(
input_size, output_size // 2
)
onset_model_size = int(onset_complexity * model_size)
self.onset_stack = nn.Sequential(
ConvStack(input_features, onset_model_size),
sequence_model(onset_model_size, onset_model_size),
nn.Linear(onset_model_size, output_features * n_instruments),
nn.Sigmoid(),
)
self.offset_stack = nn.Sequential(
ConvStack(input_features, model_size),
sequence_model(model_size, model_size),
nn.Linear(model_size, output_features),
nn.Sigmoid(),
)
self.frame_stack = nn.Sequential(
ConvStack(input_features, model_size),
nn.Linear(model_size, output_features),
nn.Sigmoid(),
)
self.combined_stack = nn.Sequential(
sequence_model(output_features * 3, model_size),
nn.Linear(model_size, output_features),
nn.Sigmoid(),
)
self.velocity_stack = nn.Sequential(
ConvStack(input_features, model_size),
nn.Linear(model_size, output_features * n_instruments),
)
def forward(self, mel):
onset_pred = self.onset_stack(mel)
offset_pred = self.offset_stack(mel)
activation_pred = self.frame_stack(mel)
onset_detached = onset_pred.detach()
shape = onset_detached.shape
keys = MAX_MIDI - MIN_MIDI + 1
new_shape = shape[:-1] + (shape[-1] // keys, keys)
onset_detached = onset_detached.reshape(new_shape)
onset_detached, _ = onset_detached.max(axis=-2)
offset_detached = offset_pred.detach()
combined_pred = torch.cat(
[onset_detached, offset_detached, activation_pred], dim=-1
)
frame_pred = self.combined_stack(combined_pred)
velocity_pred = self.velocity_stack(mel)
return onset_pred, offset_pred, activation_pred, frame_pred, velocity_pred
def run_on_batch(
self,
batch,
parallel_model=None,
positive_weight=2.0,
inv_positive_weight=2.0,
with_onset_mask=False,
):
audio_label = batch["audio"]
onset_label = batch["onset"]
offset_label = batch["offset"]
frame_label = batch["frame"]
if "velocity" in batch:
velocity_label = batch["velocity"]
mel = melspectrogram(
audio_label.reshape(-1, audio_label.shape[-1])[:, :-1]
).transpose(-1, -2)
if not parallel_model:
onset_pred, offset_pred, _, frame_pred, velocity_pred = self(mel)
else:
onset_pred, offset_pred, _, frame_pred, velocity_pred = parallel_model(mel)
predictions = {
"onset": onset_pred.reshape(*onset_label.shape),
"offset": offset_pred.reshape(*offset_label.shape),
"frame": frame_pred.reshape(*frame_label.shape),
# 'velocity': velocity_pred.reshape(*velocity_label.shape)
}
if "velocity" in batch:
predictions["velocity"] = velocity_pred.reshape(*velocity_label.shape)
losses = {
"loss/onset": F.binary_cross_entropy(
predictions["onset"], onset_label, reduction="none"
),
"loss/offset": F.binary_cross_entropy(
predictions["offset"], offset_label, reduction="none"
),
"loss/frame": F.binary_cross_entropy(
predictions["frame"], frame_label, reduction="none"
),
# 'loss/velocity': self.velocity_loss(predictions['velocity'], velocity_label, onset_label)
}
if "velocity" in batch:
losses["loss/velocity"] = self.velocity_loss(
predictions["velocity"], velocity_label, onset_label
)
onset_mask = 1.0 * onset_label
onset_mask[..., :-N_KEYS] *= positive_weight - 1
onset_mask[..., -N_KEYS:] *= inv_positive_weight - 1
onset_mask += 1
if with_onset_mask:
if "onset_mask" in batch:
onset_mask = onset_mask * batch["onset_mask"]
# if 'onset_mask' in batch:
# onset_mask += batch['onset_mask']
offset_mask = 1.0 * offset_label
offset_positive_weight = 2.0
offset_mask *= offset_positive_weight - 1
offset_mask += 1.0
frame_mask = 1.0 * frame_label
frame_positive_weight = 2.0
frame_mask *= frame_positive_weight - 1
frame_mask += 1.0
for loss_key, mask in zip(
["onset", "offset", "frame"], [onset_mask, offset_mask, frame_mask]
):
losses["loss/" + loss_key] = (mask * losses["loss/" + loss_key]).mean()
return predictions, losses
def velocity_loss(self, velocity_pred, velocity_label, onset_label):
denominator = onset_label.sum()
if denominator.item() == 0:
return denominator
else:
return (
onset_label * (velocity_label - velocity_pred) ** 2
).sum() / denominator
# same implementation as OnsetsAndFrames, but with only onset stack
class OnsetsNoFrames(nn.Module):
def __init__(
self,
input_features,
output_features,
model_complexity=48,
onset_complexity=1,
n_instruments=13,
):
nn.Module.__init__(self)
model_size = model_complexity * 16
sequence_model = lambda input_size, output_size: BiLSTM(
input_size, output_size // 2
)
onset_model_size = int(onset_complexity * model_size)
self.onset_stack = nn.Sequential(
ConvStack(input_features, onset_model_size),
sequence_model(onset_model_size, onset_model_size),
nn.Linear(onset_model_size, output_features * n_instruments),
nn.Sigmoid(),
)
def forward(self, mel):
onset_pred = self.onset_stack(mel)
onset_detached = onset_pred.detach()
shape = onset_detached.shape
keys = MAX_MIDI - MIN_MIDI + 1
new_shape = shape[:-1] + (shape[-1] // keys, keys)
onset_detached = onset_detached.reshape(new_shape)
onset_detached, _ = onset_detached.max(axis=-2)
return onset_pred
def run_on_batch(
self,
batch,
parallel_model=None,
positive_weight=2.0,
inv_positive_weight=2.0,
with_onset_mask=False,
):
audio_label = batch["audio"]
onset_label = batch["onset"]
mel = melspectrogram(
audio_label.reshape(-1, audio_label.shape[-1])[:, :-1]
).transpose(-1, -2)
if not parallel_model:
onset_pred = self(mel)
else:
onset_pred = parallel_model(mel)
predictions = {
"onset": onset_pred,
}
losses = {
"loss/onset": F.binary_cross_entropy(
predictions["onset"], onset_label, reduction="none"
),
}
onset_mask = 1.0 * onset_label
onset_mask[..., :-N_KEYS] *= positive_weight - 1
onset_mask[..., -N_KEYS:] *= inv_positive_weight - 1
onset_mask += 1
if with_onset_mask:
if "onset_mask" in batch:
onset_mask = onset_mask * batch["onset_mask"]
losses["loss/onset"] = (onset_mask * losses["loss/onset"]).mean()
return predictions, losses
|