nodes/SayanoAI-Comfy-RVC
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

new loss functions to stabilize training

Browse Source
This commit is contained in:
SayanoAI
2024-09-02 04:16:55 -04:00
parent 5ecd6a84f2
commit 4807a8caf8
4 changed files with 219 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files
custom_nodes/rvc_nodes.py
+1 -1
View File
@@ -408,7 +408,7 @@ class RVCTrainModelNode:
retrain=("BOOLEAN",{"default": False}),
save_best_model=("BOOLEAN",{"default": True}),
best_model_threshold=("INT",dict(default=50,min=20,max=80)),
log_every_epoch=("FLOAT",dict(default=1.,min=0.,max=2.,step=.5)),
log_every_epoch=("FLOAT",dict(default=1.,min=0.,max=2.,step=.1)),
num_workers=("INT",dict(default=1,min=1,max=16))
)
}
lib/infer_pack/commons.py
+13 -2
View File
@@ -1,9 +1,20 @@
import math
import numpy as np
import torch
from torch import nn
from torch.nn import functional as F
def median_pool1d(signal, kernel_size, dim=-1,stride=1):
# Pad the input to handle edge cases
padding = (kernel_size - 1) // 2
input_padded = F.pad(signal, (padding, padding), mode='reflect')
# Unfold the input to create sliding windows
unfolded = input_padded.unfold(-1, kernel_size, stride)
# Compute the median along the last dimension
median = unfolded.median(dim=dim).values
return median
def init_weights(m, mean=0.0, std=0.01):
classname = m.__class__.__name__
lib/train/losses.py
+154 -24
View File
@@ -1,31 +1,161 @@
import torch
import torchaudio
import torch.nn.functional as F
from ..infer_pack.commons import median_pool1d
def mfcc_loss(original_audio, generated_audio, sample_rate, n_mfcc=13, **kwargs):
"""
Computes the loss between the original and generated audio based on MFCCs.
Parameters:
original_audio (Tensor): The original audio waveform.
generated_audio (Tensor): The generated audio waveform.
sample_rate (int): The sample rate of the audio.
n_mfcc (int): Number of MFCC coefficients to compute (default is 13).
Returns:
Tensor: The computed MFCC loss.
"""
mfcc_transform = torchaudio.transforms.MFCC(sample_rate=sample_rate,n_mfcc=n_mfcc,melkwargs=kwargs)
mfcc_transform.to(original_audio.device)
def compute_temporal_envelope(log_magnitude: torch.Tensor, kernel_size=31):
# Compute MFCCs for original and generated audio
original_mfcc = mfcc_transform(original_audio)
generated_mfcc = mfcc_transform(generated_audio)
# Apply median pooling to the generated audio
pooled_stft = median_pool1d(log_magnitude, kernel_size, stride=kernel_size//2+1)
# Compute the L2 loss between the MFCCs of the original and generated audio
loss = torch.nn.functional.smooth_l1_loss(generated_mfcc, original_mfcc)
# Compute differences on the median-pooled audio
diff = F.normalize(torch.diff(pooled_stft, dim=-1, n=1),dim=-1)
return loss
# Return the mean of the differences as the loss
return diff.sum(dim=-2).nan_to_num(0)
def compute_cepstrals(log_magnitude: torch.Tensor, kernel_size=31):
pooled_magnitude = median_pool1d(log_magnitude, kernel_size, stride=kernel_size//2+1)
# Compute the cepstrum
cepstral_coefficients = torch.fft.ifft(pooled_magnitude, dim=-2).real
return cepstral_coefficients.sum(dim=-2).nan_to_num(0)
def compute_sts_loss(gen_stft: torch.Tensor, org_stft: torch.Tensor, kernel_size=31):
# B?N?T
org_mag = torch.log1p(org_stft.abs())
gen_mag = torch.log1p(gen_stft.abs())
gen_temp = compute_temporal_envelope(gen_mag, kernel_size=kernel_size)
org_temp = compute_temporal_envelope(org_mag, kernel_size=kernel_size)
temporal_loss = 1-F.cosine_similarity(gen_temp, org_temp, dim=-1).mean()
gen_spec = compute_cepstrals(org_mag,kernel_size=kernel_size)
org_spec = compute_cepstrals(gen_mag,kernel_size=kernel_size)
spectral_loss = F.smooth_l1_loss(gen_spec, org_spec)
org_phase = org_stft.angle()
gen_phase = gen_stft.angle()
# temporal invariant phase
org_phase = median_pool1d(org_stft.angle(),3,stride=3).mean(-1)
gen_phase = median_pool1d(gen_stft.angle(),3,stride=3).mean(-1)
phase_loss = F.smooth_l1_loss(gen_phase,org_phase)+F.smooth_l1_loss(1-torch.cos(gen_phase-org_phase),torch.zeros_like(gen_phase))
return temporal_loss, spectral_loss, phase_loss.nan_to_num(0)
def compute_harmonics(stft_matrix: torch.Tensor, kernel_size=31):
# Calculate log-magnitude spectrograms
mag = torch.log1p(stft_matrix.abs())
# Apply median filtering to separate harmonics and percussives
harmonic = median_pool1d(mag, kernel_size, stride=kernel_size//2+1)
percussive = median_pool1d(mag.transpose(-2,-1), kernel_size, stride=kernel_size//2+1).transpose(-2,-1)
return harmonic.nan_to_num(0), percussive.nan_to_num(0)
def combined_aux_loss(
original_audio: torch.Tensor, generated_audio: torch.Tensor, sample_rate: int,
c_mfcc=1., c_lfcc=1., c_hd=1., c_sts=1.,
n_fft=1024, hop_length=320, win_length=1024,
n_mfcc=13, n_lfcc=13, n_filter=80,
f_min=0., f_max=None):
kernel_size = n_fft//hop_length+1
if kernel_size % 2 == 0: kernel_size+=1
# Compute STFT once
if c_hd+c_sts>0:
hann_window = torch.hann_window(win_length).to(
dtype=original_audio.dtype, device=original_audio.device
)
generated_stft = torch.stft(
generated_audio.view(-1,generated_audio.size(-1)),
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=hann_window,
return_complex=True,
center=False)
original_stft = torch.stft(
original_audio.view(-1,original_audio.size(-1)),
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=hann_window,
return_complex=True,
center=False)
# Harmonic Loss
if c_hd>0:
original_harmonics, original_percussives = compute_harmonics(original_stft,kernel_size)
generated_harmonics, generated_percussives = compute_harmonics(generated_stft,kernel_size)
# Define loss terms
harmonic_loss = F.smooth_l1_loss(generated_harmonics, original_harmonics)
harmonic_loss += F.smooth_l1_loss(generated_percussives, original_percussives)
harmonic_loss *= c_hd
harmonic_loss
else: harmonic_loss = 0
# Spectral Temporal Smoothness Loss
if c_sts>0:
temporal_loss, spectral_loss, phase_loss = compute_sts_loss(generated_stft,original_stft,kernel_size)
temporal_loss *= c_sts
spectral_loss *= c_sts
phase_loss *= c_sts
else: spectral_loss = temporal_loss = phase_loss = 0
# MFCC Loss
if c_mfcc>0:
mfcc_transform = torchaudio.transforms.MFCC(
sample_rate=sample_rate, n_mfcc=n_mfcc,
melkwargs=dict(n_fft=n_fft, hop_length=hop_length, win_length=win_length, n_mels=n_filter, f_min=f_min, f_max=f_max)
).to(original_audio.device)
original_mfcc = mfcc_transform(original_audio)
generated_mfcc = mfcc_transform(generated_audio)
mfcc_loss = F.smooth_l1_loss(generated_mfcc, original_mfcc) * c_mfcc
else: mfcc_loss = 0
# LFCC Loss
if c_lfcc>0:
lfcc_transform = torchaudio.transforms.LFCC(
sample_rate=sample_rate, n_lfcc=n_lfcc, n_filter=n_filter, f_min=f_min, f_max=f_max,
speckwargs=dict(n_fft=n_fft, hop_length=hop_length, win_length=win_length)
).to(original_audio.device)
original_lfcc = lfcc_transform(original_audio)
generated_lfcc = lfcc_transform(generated_audio)
lfcc_loss = F.smooth_l1_loss(generated_lfcc, original_lfcc) * c_lfcc
else: lfcc_loss = 0
return harmonic_loss, temporal_loss, spectral_loss, mfcc_loss, lfcc_loss, phase_loss
def gradient_norm_loss(original_audio: torch.Tensor, generated_audio: torch.Tensor, net_d):
loss=0
# Compute the gradient penalty
# Randomly interpolate between real and generated data
size = [1]*original_audio.ndim
size[0] = original_audio.size(0)
alpha = torch.rand(*size, device=original_audio.device)
interpolated = alpha * original_audio + (1 - alpha) * generated_audio
interpolated.requires_grad_(True)
# Get the discriminator output for the interpolated data
_, disc_interpolated_output, _, _ = net_d(original_audio, interpolated)
# Compute gradients of discriminator output w.r.t. interpolated data
for output in disc_interpolated_output:
gradient = torch.autograd.grad(
outputs=output,
inputs=interpolated,
grad_outputs=torch.ones(output.size(), device=original_audio.device),
create_graph=True,
only_inputs=True
)[0]
gradient = gradient.view(gradient.size(0), -1)
loss += torch.log(gradient.norm(2,dim=-1)).abs().mean() # force gradnorm=1
return loss/len(disc_interpolated_output)
def feature_loss(fmap_r, fmap_g):
loss = 0
@@ -50,7 +180,7 @@ def discriminator_loss(disc_real_outputs, disc_generated_outputs):
loss += r_loss + g_loss
r_losses.append(r_loss.item())
g_losses.append(g_loss.item())
# loss /= len(disc_generated_outputs) #average
return loss, r_losses, g_losses
@@ -62,7 +192,7 @@ def generator_loss(disc_outputs):
l = torch.mean((1 - dg) ** 2)
gen_losses.append(l)
loss += l
# loss /= len(disc_outputs) #average
return loss, gen_losses
training_cli.py
+51 -61
View File
@@ -29,7 +29,7 @@ from .lib.train.data_utils import (
TextAudioCollate,
DistributedBucketSampler,
)
from .lib.train.losses import generator_loss, discriminator_loss, feature_loss, kl_loss, mfcc_loss
from .lib.train.losses import combined_aux_loss, generator_loss, discriminator_loss, feature_loss, gradient_norm_loss, kl_loss
from .lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
def save_checkpoint(ckpt, name, epoch, hps, model_path=None):
@@ -327,7 +327,7 @@ def train_and_evaluate(
# Make new cache
for batch_idx, info in enumerate(train_loader):
# Unpack
if hps.if_f0 == 1:
if hps.if_f0:
(
phone,
phone_lengths,
@@ -353,7 +353,7 @@ def train_and_evaluate(
if torch.cuda.is_available():
phone = phone.cuda(rank, non_blocking=True)
phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
if hps.if_f0 == 1:
if hps.if_f0:
pitch = pitch.cuda(rank, non_blocking=True)
pitchf = pitchf.cuda(rank, non_blocking=True)
sid = sid.cuda(rank, non_blocking=True)
@@ -362,7 +362,7 @@ def train_and_evaluate(
wave = wave.cuda(rank, non_blocking=True)
wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
# Cache on list
if hps.if_f0 == 1:
if hps.if_f0:
cache.append(
(
batch_idx,
@@ -406,7 +406,7 @@ def train_and_evaluate(
for batch_idx, info in tqdm(data_iterator,desc=f"[Epoch {epoch}]: "):
# Data
## Unpack
if hps.if_f0 == 1:
if hps.if_f0:
(
phone,
phone_lengths,
@@ -424,7 +424,7 @@ def train_and_evaluate(
if (not hps.if_cache_data_in_gpu) and torch.cuda.is_available():
phone = phone.cuda(rank, non_blocking=True)
phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
if hps.if_f0 == 1:
if hps.if_f0:
pitch = pitch.cuda(rank, non_blocking=True)
pitchf = pitchf.cuda(rank, non_blocking=True)
sid = sid.cuda(rank, non_blocking=True)
@@ -434,7 +434,7 @@ def train_and_evaluate(
# Calculate
with autocast(enabled=hps.train.fp16_run):
if hps.if_f0 == 1:
if hps.if_f0:
(
y_hat,
ids_slice,
@@ -473,6 +473,7 @@ def train_and_evaluate(
hps.data.mel_fmax,
)
if hps.train.fp16_run: y_hat_mel = y_hat_mel.half()
wave_orig = wave.clone()
wave = commons.slice_segments(wave, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice
# Discriminator
@@ -480,46 +481,9 @@ def train_and_evaluate(
y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, gen_wave)
with autocast(enabled=False):
gradient_penalty = torch.Tensor([0.]).to(wave.device)
if hps.train.get("c_gp",0.)>0:
# Compute the gradient penalty
# Randomly interpolate between real and generated data
size = [1]*wave.ndim
size[0] = wave.size(0)
alpha = torch.rand(*size, device=wave.device)
interpolated = alpha * wave + (1 - alpha) * gen_wave
interpolated.requires_grad_(True)
# Get the discriminator output for the interpolated data
_, disc_interpolated_output, _, _ = net_d(wave, interpolated)
# Compute gradients of discriminator output w.r.t. interpolated data
for output in disc_interpolated_output:
gradient = torch.autograd.grad(
outputs=output,
inputs=interpolated,
grad_outputs=torch.ones(output.size(), device=wave.device),
create_graph=True,
only_inputs=True
)[0]
gradient = gradient.view(gradient.size(0), -1)
gradient_penalty += torch.log(gradient.norm(2,dim=-1)).abs().mean() # force gradnorm=1
gradient_penalty *= hps.train.c_gp/len(disc_interpolated_output)
loss_mfcc = torch.Tensor([0.]).to(wave.device)
if hps.train.get("c_mfcc",0)>0:
loss_mfcc = mfcc_loss(wave, gen_wave, SR_MAP[hps.sample_rate],
n_fft=hps.data.filter_length,
hop_length=hps.data.hop_length,
win_length=hps.data.win_length,
n_mels=hps.data.n_mel_channels,
norm="slaney",
f_min=hps.data.mel_fmin,
f_max=hps.data.mel_fmax
) * hps.train.c_mfcc
loss_mfcc = loss_mfcc.squeeze()
gradient_penalty = gradient_penalty.squeeze()
gradient_penalty = gradient_norm_loss(wave,gen_wave, net_d)*hps.train.c_gp if hps.train.get("c_gp",0.)>0 else 0
loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
loss_disc_all = loss_disc+gradient_penalty+loss_mfcc
loss_disc_all = loss_disc+gradient_penalty
optim_d.zero_grad()
scaler.scale(loss_disc_all).backward()
@@ -531,11 +495,25 @@ def train_and_evaluate(
# Generator
y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
with autocast(enabled=False):
loss_mel = F.smooth_l1_loss(y_mel, y_hat_mel) * hps.train.get("c_mel",45.)
loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.get("c_mel",45.)
loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.get("c_kl",1.)
loss_fm = feature_loss(fmap_r, fmap_g) * hps.train.get("c_fm",2.)
harmonic_loss, temporal_loss, spectral_loss, mfcc_loss, lfcc_loss, phase_loss = combined_aux_loss(
wave, y_hat, SR_MAP[hps.sample_rate],
c_lfcc=hps.train.get("c_lfcc",0.),
c_mfcc=hps.train.get("c_mfcc",0.),
c_hd=hps.train.get("c_hd",0.),
c_sts=hps.train.get("c_sts",0.),
n_fft=hps.data.filter_length,
hop_length=hps.data.hop_length,
win_length=hps.data.win_length,
n_filter=hps.data.n_mel_channels,
f_min=hps.data.mel_fmin,
f_max=hps.data.mel_fmax
)
loss_gen, losses_gen = generator_loss(y_d_hat_g)
loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
aux_loss = harmonic_loss + temporal_loss + spectral_loss + mfcc_loss + lfcc_loss + phase_loss
loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl + aux_loss
optim_g.zero_grad()
scaler.scale(loss_gen_all).backward()
@@ -548,13 +526,12 @@ def train_and_evaluate(
# if epoch>=max(hps.total_epoch*.2,20): #start saving best model after 20% or 20 epochs
if loss_gen_all+loss_disc_all<least_loss:
least_loss = loss_gen_all+loss_disc_all
logger.info(f"[lowest loss] {least_loss=:.3f}: {loss_disc=:.3f} {loss_mfcc=:.3f} {gradient_penalty=:.3f}, {loss_gen=:.3f}, {loss_fm=:.3f}, {loss_mel=:.3f}, {loss_kl=:.3f}")
logger.info(f"[lowest loss] {least_loss=:.3f}: {loss_disc=:.3f} {gradient_penalty=:.3f} <==> {loss_gen=:.3f} {loss_fm=:.3f} {loss_mel=:.3f} {loss_kl=:.3f}")
logger.info(f"[aux] {aux_loss=:.3f}: {harmonic_loss=:.3f}, {temporal_loss=:.3f}, {spectral_loss=:.3f}, {mfcc_loss=:.3f}, {lfcc_loss=:.3f} {phase_loss=:.3f}")
if hps.save_best_model:
if hasattr(net_g, "module"):
ckpt = net_g.module.state_dict()
else:
ckpt = net_g.state_dict()
if hasattr(net_g, "module"): ckpt = net_g.module.state_dict()
else: ckpt = net_g.state_dict()
best_model_name = f"{hps.name}_e{epoch}_s{global_step}_loss{least_loss:2.0f}" if hps.save_every_weights else f"{hps.name}_loss{least_loss:2.0f}"
status = save_checkpoint(ckpt,best_model_name,epoch,hps)
@@ -575,16 +552,22 @@ def train_and_evaluate(
scalar_dict = {
"total/loss/all": loss_gen_all+loss_disc_all,
"total/loss/gen_all": loss_gen_all,
"total/loss/aux": aux_loss,
"total/loss/disc_all": loss_disc_all,
"total/loss/gen": loss_gen,
"total/loss/disc": loss_disc,
"total/loss/fm": loss_fm,
"total/loss/mel": loss_mel,
"total/loss/kl": loss_kl,
"total/loss/mfcc": loss_mfcc,
"learning_rate": lr,
"gradient/grad_norm_d": grad_norm_d,
"gradient/grad_norm_g": grad_norm_g,
"aux/loss/harmonic": harmonic_loss,
"aux/loss/temporal": temporal_loss,
"aux/loss/spectral": spectral_loss,
"aux/loss/mfcc": mfcc_loss,
"aux/loss/lfcc": lfcc_loss,
"aux/loss/phase_loss": phase_loss,
"gradient/lr": lr,
"gradient/grad_norm_disc": grad_norm_d,
"gradient/grad_norm_gen": grad_norm_g,
"gradient/gradient_penalty": gradient_penalty,
**{f"loss/g/{i}": v for i, v in enumerate(losses_gen)},
**{f"loss/d_r/{i}": v for i, v in enumerate(losses_disc_r)},
@@ -597,9 +580,15 @@ def train_and_evaluate(
"all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
}
with torch.no_grad():
if hasattr(net_g, "module"): inference = net_g.module.infer
else: inference = net_g.infer
if hps.if_f0: wave_gen = inference(phone, phone_lengths, pitch, pitchf, sid)[0][0, 0].data
else: wave_gen = inference(phone, phone_lengths, sid)[0][0, 0].data
audio_dict = {
"slice/wave_org": wave.flatten(),
"slice/wave_gen": y_hat.flatten()
"slice/wave_org": wave_orig[0][0],
"slice/wave_gen": wave_gen
}
utils.summarize(
writer=writer,
@@ -637,14 +626,15 @@ def train_and_evaluate(
if rank == 0:
logger.info(f"====> Epoch {epoch} (Total Loss={(loss_disc_all+loss_gen_all).item():.3f}): {global_step=} {lr=:.2E} {epoch_recorder.record()}")
logger.info(f"|| {loss_disc_all=:.3f}: {loss_disc=:.3f}, {loss_mfcc=:.3f}, {gradient_penalty=:.3f}")
logger.info(f"|| {loss_disc_all=:.3f}: {loss_disc=:.3f}, {gradient_penalty=:.3f}")
logger.info(f"|| {loss_gen_all=:.3f}: {loss_gen=:.3f}, {loss_fm=:.3f}, {loss_mel=:.3f}, {loss_kl=:.3f}")
logger.info(f"|| {harmonic_loss=:.3f}, {temporal_loss=:.3f}, {spectral_loss=:.3f}, {mfcc_loss=:.3f}, {lfcc_loss=:.3f}, {phase_loss=:.3f}")
if epoch >= hps.total_epoch and rank == 0:
logger.info("Training is done. The program is closed.")
ckpt = net_g.module.state_dict() if hasattr(net_g, "module") else net_g.state_dict()
if hps.save_best_model:
if hps.save_best_model and os.path.isfile(loss_file):
with open(loss_file,"r") as f:
data = json.load(f)
best_model_name = data.get("best_model_name","")