File size: 10,360 Bytes
2c6366f | 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 | from torch import optim, nn, mean
import torch
import lightning as L
from typing import Any
from pytorch_lightning.utilities.types import STEP_OUTPUT, OptimizerLRScheduler
from src.config import NUM_FRETS, NUM_STRINGS, PITCH_CLASSES, PLAYABILITY_THRESHOLD
import src.config as C
from torchmetrics import Accuracy, Precision, Recall, F1Score
from src.data.dataset import TorchDataset
from src.model.losses import pitch_class_loss, completeness_metric, open_chord_metric, playability_metric
import src.model.losses as Lo
import music21 as m21
class FingeringPredictor(L.LightningModule):
def __init__(self, span_loss_coeff: float = 0.0, with_mute: bool = False,
learning_rate: float = 1e-3,
*args: Any, **kwargs: Any) -> None:
super().__init__(*args, **kwargs)
self.save_hyperparameters()
self.with_mute = with_mute
self.num_frets = NUM_FRETS + 1 if with_mute else NUM_FRETS
self.model = nn.Linear((self.num_frets*NUM_STRINGS) + 2*PITCH_CLASSES,
out_features=(NUM_STRINGS*self.num_frets))
self.activation = nn.Sigmoid()
self.pc_loss = pitch_class_loss
self.loss = nn.BCELoss()
self.span_loss = Lo.hand_span_loss
self.oc_metric = open_chord_metric
self.acc = Accuracy(task="binary")
self.prec = Precision(task="binary")
self.rec = Recall(task="binary")
self.sf_prec = Lo.stringfret_precision
self.sf_rec = Lo.stringfret_recall
self.sf_f1 = Lo.stringfret_f1
self.lr = learning_rate
self.f1 = F1Score(task="binary")
self.sf_exactness = Lo.stringwise_exactness
self.playability = playability_metric
self.completeness = completeness_metric
self.span_loss_coeff = span_loss_coeff
def forward(self, x) -> Any:
z = self.model(x)
y_hat = self.activation(z) # [B, NUM_FRETS*NUM_STRINGS]
if not self.with_mute:
return y_hat
if x.ndim == 2:
y_hat = nn.functional.normalize(y_hat.view((x.size()[0], NUM_STRINGS, self.num_frets)),
p=1, dim=2).flatten(start_dim=1)
else:
y_hat = nn.functional.normalize(y_hat.view((NUM_STRINGS, self.num_frets)),
p=1, dim=1).flatten()
return y_hat
def training_step(self, batch, batch_idx, *args: Any, **kwargs: Any) -> STEP_OUTPUT:
x, y = batch
y_hat = self(x)
loss = self.loss(y_hat, y)
if self.span_loss_coeff > 0 :
span_loss = self.span_loss(torch.unflatten(y_hat, dim=1, sizes=(NUM_STRINGS, self.num_frets)),
with_mute=self.with_mute)
loss = loss + self.span_loss_coeff*span_loss
self.log("Train/span-loss", span_loss, prog_bar=True)
self.log("Train/loss", loss, prog_bar=True)
return loss
def validation_step(self, batch, batch_idx, *args: Any, **kwargs: Any) -> STEP_OUTPUT:
x, y = batch
y_hat = self(x)
loss = self.loss(y_hat, y)
if self.span_loss_coeff > 0 :
span_loss = self.span_loss(torch.unflatten(y_hat, dim=1, sizes=(NUM_STRINGS, self.num_frets)),
with_mute=self.with_mute)
loss = loss + self.span_loss_coeff*span_loss
self.log("Val/span-loss", span_loss, prog_bar=True)
self.log("Val/loss", loss, prog_bar=True)
self.log("val_loss", loss, prog_bar=False)
return loss
def binarize(self, tensor) -> torch.Tensor:
if self.with_mute:
max_tensor = torch.max(torch.unflatten(tensor, 1,
(-1, self.num_frets)), dim=-1).values
pred_bin = torch.where(torch.unflatten(tensor, 1,
(-1, self.num_frets)) == max_tensor[:, :, None], 1, 0)
else:
pred_bin = tensor > 0.5
pred_bin = torch.unflatten(pred_bin, 1, (-1, self.num_frets))
return pred_bin
def test_step(self, batch, batch_idx, *args: Any, **kwargs: Any) -> STEP_OUTPUT:
x, y = batch
batch_size = len(y)
y_hat = self(x)
loss = self.loss(y_hat, y)
acc = self.acc(y_hat, y)
rec = self.rec(y_hat, y)
prec = self.prec(y_hat, y)
f1 = self.f1(y_hat, y)
pred_bin = self.binarize(y_hat)
fingerings = pred_bin.view((y_hat.size()[0], NUM_STRINGS, self.num_frets))
expected = y.view((y_hat.size()[0], NUM_STRINGS, self.num_frets))
expected_notes = Lo._midi_notes_from_fingering_batchwise(expected[:, :,:-1])
expected_notes = [[m21.note.Note(a).name for a in expected_notes[i] if a !=0] for i in range(len(expected_notes))]
pred_notes = Lo._midi_notes_from_fingering_batchwise(pred_bin[:, :, :-1])
pred_notes = [[m21.note.Note(a).name for a in pred_notes[i] if a !=0] for i in range(len(pred_notes))]
pc_prec = torch.Tensor([Lo.pc_precision(expected_notes[i], pred_notes[i]) for i in range(len(expected_notes))])
pc_rec = torch.Tensor([Lo.pc_recall(expected_notes[i], pred_notes[i]) for i in range(len(expected_notes))])
oc = self.oc_metric(fingerings, expected)
sw_exact = self.sf_exactness(fingerings, expected)
sf_prec = self.sf_prec(fingerings, expected, with_mute=self.with_mute)
sf_rec = self.sf_rec(fingerings, expected, with_mute=self.with_mute)
sf_f1 = self.sf_f1(fingerings, expected, with_mute=self.with_mute)
pc_f1 = torch.div(2*pc_prec*pc_rec, pc_prec+pc_rec)
pc_f1 = torch.nan_to_num(pc_f1, nan=0)
diagrams = [TorchDataset.fingering_from_target_tensor(pred_bin.view((batch_size, NUM_STRINGS, self.num_frets))[i], with_mute=self.with_mute) for i in range(batch_size)]
diagrams_previous = [TorchDataset.fingering_from_source_tensor(x[i], with_mute=self.with_mute) for i in range(batch_size)]
playability = torch.Tensor([Lo.anatomical_score(diagrams[i])[0] for i in range(len(diagrams))])
unplayable = playability < PLAYABILITY_THRESHOLD
best_fingerings = [Lo.anatomical_score(diagrams[i])[1] for i in range(len(diagrams))]
best_fingerings_previous = [Lo.anatomical_score(diagrams_previous[i])[1] for i in range(batch_size)]
transition_costs = torch.Tensor([Lo.transition_cost(best_fingerings_previous[i],
best_fingerings[i]) for i in range(batch_size)])
c_r_muted = [Lo.ratio_muted_strings(current_diag) for current_diag in diagrams_previous]
c_r_open = [Lo.ratio_open_strings(current_diag) for current_diag in diagrams_previous]
c_num_strings = [Lo.num_strings_played(current_diag) for current_diag in diagrams_previous]
c_centroid = [Lo.string_centroid(current_diag) for current_diag in diagrams_previous]
c_r_unique = [Lo.ratio_unique_notes(current_diag) for current_diag in diagrams_previous]
n_r_muted = [Lo.ratio_muted_strings(current_diag) for current_diag in diagrams]
n_r_open = [Lo.ratio_open_strings(current_diag) for current_diag in diagrams]
n_num_strings = [Lo.num_strings_played(current_diag) for current_diag in diagrams]
n_centroid = [Lo.string_centroid(current_diag) for current_diag in diagrams]
n_r_unique = [Lo.ratio_unique_notes(current_diag) for current_diag in diagrams]
r_muted_diff = [abs(c_r_muted[i] - n_r_muted[i]) for i in range(batch_size)]
r_open_diff = [abs(c_r_open[i] - n_r_open[i]) for i in range(batch_size)]
num_strings_diff = [abs(c_num_strings[i] - n_num_strings[i]) for i in range(batch_size)]
centroid_diff = [abs(c_centroid[i] - n_centroid[i]) for i in range(batch_size)]
r_unique_diff = [abs(c_r_unique[i] - n_r_unique[i]) for i in range(batch_size)]
self.log("Test/loss", loss, prog_bar=False)
self.log("Test/acc", acc)
self.log("Test/rec", rec)
self.log("Test/prec", prec)
self.log("Test/SF rec", sf_rec.mean())
self.log("Test/SF prec", sf_prec.mean())
self.log("Test/SF F1", sf_f1.mean())
self.log("Test/PC F1", pc_f1.mean())
self.log("Test/F1", f1)
self.log("Test/exactness", self.exact_acc(pred_bin, y))
self.log("Test/StringExactness", sw_exact.mean())
self.log("Test/PC-Recall", pc_rec.mean())
self.log("Test/playability", playability.mean())
self.log("Test/unplayable", unplayable.mean(dtype=torch.float32))
self.log("Test/open-closed", oc.mean(dtype=torch.float32))
self.log("Test/PC-Precision", pc_prec.mean())
self.log('Test/transition_cost', transition_costs.mean())
self.log('Test/ratio_muted', torch.Tensor(n_r_muted).mean())
self.log('Test/ratio_open', torch.Tensor(n_r_open).mean())
self.log('Test/num_strings', torch.Tensor(n_num_strings).mean())
self.log('Test/centroid', torch.Tensor(n_centroid).mean())
self.log('Test/ratio_unique_notes', torch.Tensor(n_r_unique).mean())
self.log('Test/ratio_muted_diff', torch.Tensor(r_muted_diff).mean())
self.log('Test/ratio_open_diff', torch.Tensor(r_open_diff).mean())
self.log('Test/num_strings_diff', torch.Tensor(num_strings_diff).mean())
self.log('Test/centroid_diff', torch.Tensor(centroid_diff).mean())
self.log('Test/ratio_unique_notes_diff', torch.Tensor(r_unique_diff).mean())
return loss
def exact_acc(self, pred, target, threshold: float = 0.5):
tmp = torch.eq(torch.flatten(pred,start_dim=1), target)
return torch.mean(torch.all(tmp, dim=-1), dtype=torch.float16)
def configure_optimizers(self) -> OptimizerLRScheduler:
optimizer = optim.Adam(self.parameters(), lr=self.lr)
return optimizer
class FingeringPredictorBaseline(FingeringPredictor):
def __init__(self, learning_rate: float = 0.001, *args: Any, **kwargs: Any) -> None:
super().__init__(learning_rate=learning_rate, **kwargs)
self.model = nn.Linear(2*PITCH_CLASSES, self.num_frets*NUM_STRINGS)
def forward(self, x) -> Any:
# drop chord fingering info (first dim is batch size)
if x.dim() > 1:
x = x[:, -2*PITCH_CLASSES:]
else:
x = x[-2*PITCH_CLASSES:]
z = self.model(x)
y_hat = self.activation(z)
return y_hat
|