deepafx-st / deepafx_st /system.py
mikeross's picture
Duplicate from nateraw/deepafx-st
c983126
import torch
import auraloss
import torchaudio
from itertools import chain
import pytorch_lightning as pl
from argparse import ArgumentParser
from typing import Tuple, List, Dict
import deepafx_st.utils as utils
from deepafx_st.utils import DSPMode
from deepafx_st.data.dataset import AudioDataset
from deepafx_st.models.encoder import SpectralEncoder
from deepafx_st.models.controller import StyleTransferController
from deepafx_st.processors.spsa.channel import SPSAChannel
from deepafx_st.processors.spsa.eps_scheduler import EpsilonScheduler
from deepafx_st.processors.proxy.channel import ProxyChannel
from deepafx_st.processors.autodiff.channel import AutodiffChannel
class System(pl.LightningModule):
def __init__(
self,
ext="wav",
dsp_sample_rate=24000,
**kwargs,
):
super().__init__()
self.save_hyperparameters()
self.eps_scheduler = EpsilonScheduler(
self.hparams.spsa_epsilon,
self.hparams.spsa_patience,
self.hparams.spsa_factor,
self.hparams.spsa_verbose,
)
self.hparams.dsp_mode = DSPMode.NONE
# first construct the processor, since this will dictate encoder
if self.hparams.processor_model == "spsa":
self.processor = SPSAChannel(
self.hparams.dsp_sample_rate,
self.hparams.spsa_parallel,
self.hparams.batch_size,
)
elif self.hparams.processor_model == "autodiff":
self.processor = AutodiffChannel(self.hparams.dsp_sample_rate)
elif self.hparams.processor_model == "proxy0":
# print('self.hparams.proxy_ckpts,',self.hparams.proxy_ckpts)
self.hparams.dsp_mode = DSPMode.NONE
self.processor = ProxyChannel(
self.hparams.proxy_ckpts,
self.hparams.freeze_proxies,
self.hparams.dsp_mode,
sample_rate=self.hparams.dsp_sample_rate,
)
elif self.hparams.processor_model == "proxy1":
# print('self.hparams.proxy_ckpts,',self.hparams.proxy_ckpts)
self.hparams.dsp_mode = DSPMode.INFER
self.processor = ProxyChannel(
self.hparams.proxy_ckpts,
self.hparams.freeze_proxies,
self.hparams.dsp_mode,
sample_rate=self.hparams.dsp_sample_rate,
)
elif self.hparams.processor_model == "proxy2":
# print('self.hparams.proxy_ckpts,',self.hparams.proxy_ckpts)
self.hparams.dsp_mode = DSPMode.TRAIN_INFER
self.processor = ProxyChannel(
self.hparams.proxy_ckpts,
self.hparams.freeze_proxies,
self.hparams.dsp_mode,
sample_rate=self.hparams.dsp_sample_rate,
)
elif self.hparams.processor_model == "tcn1":
# self.processor = ConditionalTCN(self.hparams.sample_rate)
self.hparams.dsp_mode = DSPMode.NONE
self.processor = ProxyChannel(
[],
freeze_proxies=False,
dsp_mode=self.hparams.dsp_mode,
tcn_nblocks=self.hparams.tcn_nblocks,
tcn_dilation_growth=self.hparams.tcn_dilation_growth,
tcn_channel_width=self.hparams.tcn_channel_width,
tcn_kernel_size=self.hparams.tcn_kernel_size,
num_tcns=1,
sample_rate=self.hparams.sample_rate,
)
elif self.hparams.processor_model == "tcn2":
self.hparams.dsp_mode = DSPMode.NONE
self.processor = ProxyChannel(
[],
freeze_proxies=False,
dsp_mode=self.hparams.dsp_mode,
tcn_nblocks=self.hparams.tcn_nblocks,
tcn_dilation_growth=self.hparams.tcn_dilation_growth,
tcn_channel_width=self.hparams.tcn_channel_width,
tcn_kernel_size=self.hparams.tcn_kernel_size,
num_tcns=2,
sample_rate=self.hparams.sample_rate,
)
else:
raise ValueError(f"Invalid processor_model: {self.hparams.processor_model}")
if self.hparams.encoder_ckpt is not None:
# load encoder weights from a pre-trained system
system = System.load_from_checkpoint(self.hparams.encoder_ckpt)
self.encoder = system.encoder
self.hparams.encoder_embed_dim = system.encoder.embed_dim
else:
self.encoder = SpectralEncoder(
self.processor.num_control_params,
self.hparams.sample_rate,
encoder_model=self.hparams.encoder_model,
embed_dim=self.hparams.encoder_embed_dim,
width_mult=self.hparams.encoder_width_mult,
)
if self.hparams.encoder_freeze:
for param in self.encoder.parameters():
param.requires_grad = False
self.controller = StyleTransferController(
self.processor.num_control_params,
self.hparams.encoder_embed_dim,
)
if len(self.hparams.recon_losses) != len(self.hparams.recon_loss_weights):
raise ValueError("Must supply same number of weights as losses.")
self.recon_losses = torch.nn.ModuleDict()
for recon_loss in self.hparams.recon_losses:
if recon_loss == "mrstft":
self.recon_losses[recon_loss] = auraloss.freq.MultiResolutionSTFTLoss(
fft_sizes=[32, 128, 512, 2048, 8192, 32768],
hop_sizes=[16, 64, 256, 1024, 4096, 16384],
win_lengths=[32, 128, 512, 2048, 8192, 32768],
w_sc=0.0,
w_phs=0.0,
w_lin_mag=1.0,
w_log_mag=1.0,
)
elif recon_loss == "mrstft-md":
self.recon_losses[recon_loss] = auraloss.freq.MultiResolutionSTFTLoss(
fft_sizes=[128, 512, 2048, 8192],
hop_sizes=[32, 128, 512, 2048], # 1 / 4
win_lengths=[128, 512, 2048, 8192],
w_sc=0.0,
w_phs=0.0,
w_lin_mag=1.0,
w_log_mag=1.0,
)
elif recon_loss == "mrstft-sm":
self.recon_losses[recon_loss] = auraloss.freq.MultiResolutionSTFTLoss(
fft_sizes=[512, 2048, 8192],
hop_sizes=[256, 1024, 4096], # 1 / 4
win_lengths=[512, 2048, 8192],
w_sc=0.0,
w_phs=0.0,
w_lin_mag=1.0,
w_log_mag=1.0,
)
elif recon_loss == "melfft":
self.recon_losses[recon_loss] = auraloss.freq.MelSTFTLoss(
self.hparams.sample_rate,
fft_size=self.hparams.train_length,
hop_size=self.hparams.train_length // 2,
win_length=self.hparams.train_length,
n_mels=128,
w_sc=0.0,
device="cuda" if self.hparams.gpus > 0 else "cpu",
)
elif recon_loss == "melstft":
self.recon_losses[recon_loss] = auraloss.freq.MelSTFTLoss(
self.hparams.sample_rate,
device="cuda" if self.hparams.gpus > 0 else "cpu",
)
elif recon_loss == "l1":
self.recon_losses[recon_loss] = torch.nn.L1Loss()
elif recon_loss == "sisdr":
self.recon_losses[recon_loss] = auraloss.time.SISDRLoss()
else:
raise ValueError(
f"Invalid reconstruction loss: {self.hparams.recon_losses}"
)
def forward(
self,
x: torch.Tensor,
y: torch.Tensor = None,
e_y: torch.Tensor = None,
z: torch.Tensor = None,
dsp_mode: DSPMode = DSPMode.NONE,
analysis_length: int = 0,
sample_rate: int = 24000,
):
"""Forward pass through the system subnetworks.
Args:
x (tensor): Input audio tensor with shape (batch x 1 x samples)
y (tensor): Target audio tensor with shape (batch x 1 x samples)
e_y (tensor): Target embedding with shape (batch x edim)
z (tensor): Bottleneck latent.
dsp_mode (DSPMode): Mode of operation for the DSP blocks.
analysis_length (optional, int): Only analyze the first N samples.
sample_rate (optional, int): Desired sampling rate for the DSP blocks.
You must supply target audio `y`, `z`, or an embedding for the target `e_y`.
Returns:
y_hat (tensor): Output audio.
p (tensor):
e (tensor):
"""
bs, chs, samp = x.size()
if sample_rate != self.hparams.sample_rate:
x_enc = torchaudio.transforms.Resample(
sample_rate, self.hparams.sample_rate
).to(x.device)(x)
if y is not None:
y_enc = torchaudio.transforms.Resample(
sample_rate, self.hparams.sample_rate
).to(x.device)(y)
else:
x_enc = x
y_enc = y
if analysis_length > 0:
x_enc = x_enc[..., :analysis_length]
if y is not None:
y_enc = y_enc[..., :analysis_length]
e_x = self.encoder(x_enc) # generate latent embedding for input
if y is not None:
e_y = self.encoder(y_enc) # generate latent embedding for target
elif e_y is None:
raise RuntimeError("Must supply y, z, or e_y. None supplied.")
# learnable comparision
p = self.controller(e_x, e_y, z=z)
# process audio conditioned on parameters
# if there are multiple channels process them using same parameters
y_hat = torch.zeros(x.shape).type_as(x)
for ch_idx in range(chs):
y_hat_ch = self.processor(
x[:, ch_idx : ch_idx + 1, :],
p,
epsilon=self.eps_scheduler.epsilon,
dsp_mode=dsp_mode,
sample_rate=sample_rate,
)
y_hat[:, ch_idx : ch_idx + 1, :] = y_hat_ch
return y_hat, p, e_x
def common_paired_step(
self,
batch: Tuple,
batch_idx: int,
optimizer_idx: int = 0,
train: bool = False,
):
"""Model step used for validation and training.
Args:
batch (Tuple[Tensor, Tensor]): Batch items containing input audio (x) and target audio (y).
batch_idx (int): Index of the batch within the current epoch.
optimizer_idx (int): Index of the optimizer, this step is called once for each optimizer.
The firs optimizer corresponds to the generator and the second optimizer,
corresponds to the adversarial loss (when in use).
train (bool): Whether step is called during training (True) or validation (False).
"""
x, y = batch
loss = 0
dsp_mode = self.hparams.dsp_mode
if train and dsp_mode.INFER.name == DSPMode.INFER.name:
dsp_mode = DSPMode.NONE
# proces input audio through model
if self.hparams.style_transfer:
length = x.shape[-1]
x_A = x[..., : length // 2]
x_B = x[..., length // 2 :]
y_A = y[..., : length // 2]
y_B = y[..., length // 2 :]
if torch.rand(1).sum() > 0.5:
y_ref = y_B
y = y_A
x = x_A
else:
y_ref = y_A
y = y_B
x = x_B
y_hat, p, e = self(x, y=y_ref, dsp_mode=dsp_mode)
else:
y_ref = None
y_hat, p, e = self(x, dsp_mode=dsp_mode)
# compute reconstruction loss terms
for loss_idx, (loss_name, recon_loss_fn) in enumerate(
self.recon_losses.items()
):
temp_loss = recon_loss_fn(y_hat, y) # reconstruction loss
loss += float(self.hparams.recon_loss_weights[loss_idx]) * temp_loss
self.log(
("train" if train else "val") + f"_loss/{loss_name}",
temp_loss,
on_step=True,
on_epoch=True,
prog_bar=False,
logger=True,
sync_dist=True,
)
# log the overall aggregate loss
self.log(
("train" if train else "val") + "_loss/loss",
loss,
on_step=True,
on_epoch=True,
prog_bar=False,
logger=True,
sync_dist=True,
)
# store audio data
data_dict = {
"x": x.cpu(),
"y": y.cpu(),
"p": p.cpu(),
"e": e.cpu(),
"y_hat": y_hat.cpu(),
}
if y_ref is not None:
data_dict["y_ref"] = y_ref.cpu()
return loss, data_dict
def training_step(self, batch, batch_idx, optimizer_idx=0):
loss, _ = self.common_paired_step(
batch,
batch_idx,
optimizer_idx,
train=True,
)
return loss
def training_epoch_end(self, training_step_outputs):
if self.hparams.spsa_schedule and self.hparams.processor_model == "spsa":
self.eps_scheduler.step(
self.trainer.callback_metrics[self.hparams.train_monitor],
)
def validation_step(self, batch, batch_idx):
loss, data_dict = self.common_paired_step(batch, batch_idx)
return data_dict
def optimizer_step(
self,
epoch,
batch_idx,
optimizer,
optimizer_idx,
optimizer_closure,
on_tpu=False,
using_native_amp=False,
using_lbfgs=False,
):
if optimizer_idx == 0:
optimizer.step(closure=optimizer_closure)
def configure_optimizers(self):
# we need additional optimizer for the discriminator
optimizers = []
g_optimizer = torch.optim.Adam(
chain(
self.encoder.parameters(),
self.processor.parameters(),
self.controller.parameters(),
),
lr=self.hparams.lr,
betas=(0.9, 0.999),
)
optimizers.append(g_optimizer)
g_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
g_optimizer,
patience=self.hparams.lr_patience,
verbose=True,
)
ms1 = int(self.hparams.max_epochs * 0.8)
ms2 = int(self.hparams.max_epochs * 0.95)
print(
"Learning rate schedule:",
f"0 {self.hparams.lr:0.2e} -> ",
f"{ms1} {self.hparams.lr*0.1:0.2e} -> ",
f"{ms2} {self.hparams.lr*0.01:0.2e}",
)
g_scheduler = torch.optim.lr_scheduler.MultiStepLR(
g_optimizer,
milestones=[ms1, ms2],
gamma=0.1,
)
lr_schedulers = {
"scheduler": g_scheduler,
}
return optimizers, lr_schedulers
def train_dataloader(self):
train_dataset = AudioDataset(
self.hparams.audio_dir,
subset="train",
train_frac=self.hparams.train_frac,
half=self.hparams.half,
length=self.hparams.train_length,
input_dirs=self.hparams.input_dirs,
random_scale_input=self.hparams.random_scale_input,
random_scale_target=self.hparams.random_scale_target,
buffer_size_gb=self.hparams.buffer_size_gb,
buffer_reload_rate=self.hparams.buffer_reload_rate,
num_examples_per_epoch=self.hparams.train_examples_per_epoch,
augmentations={
"pitch": {"sr": self.hparams.sample_rate},
"tempo": {"sr": self.hparams.sample_rate},
},
freq_corrupt=self.hparams.freq_corrupt,
drc_corrupt=self.hparams.drc_corrupt,
ext=self.hparams.ext,
)
g = torch.Generator()
g.manual_seed(0)
return torch.utils.data.DataLoader(
train_dataset,
num_workers=self.hparams.num_workers,
batch_size=self.hparams.batch_size,
worker_init_fn=utils.seed_worker,
generator=g,
pin_memory=True,
persistent_workers=True,
timeout=60,
)
def val_dataloader(self):
val_dataset = AudioDataset(
self.hparams.audio_dir,
subset="val",
half=self.hparams.half,
train_frac=self.hparams.train_frac,
length=self.hparams.val_length,
input_dirs=self.hparams.input_dirs,
buffer_size_gb=self.hparams.buffer_size_gb,
buffer_reload_rate=self.hparams.buffer_reload_rate,
random_scale_input=self.hparams.random_scale_input,
random_scale_target=self.hparams.random_scale_target,
num_examples_per_epoch=self.hparams.val_examples_per_epoch,
augmentations={},
freq_corrupt=self.hparams.freq_corrupt,
drc_corrupt=self.hparams.drc_corrupt,
ext=self.hparams.ext,
)
self.val_dataset = val_dataset
g = torch.Generator()
g.manual_seed(0)
return torch.utils.data.DataLoader(
val_dataset,
num_workers=1,
batch_size=self.hparams.batch_size,
worker_init_fn=utils.seed_worker,
generator=g,
pin_memory=True,
persistent_workers=True,
timeout=60,
)
def shutdown(self):
del self.processor
# add any model hyperparameters here
@staticmethod
def add_model_specific_args(parent_parser):
parser = ArgumentParser(parents=[parent_parser], add_help=False)
# --- Training ---
parser.add_argument("--batch_size", type=int, default=32)
parser.add_argument("--lr", type=float, default=3e-4)
parser.add_argument("--lr_patience", type=int, default=20)
parser.add_argument("--recon_losses", nargs="+", default=["l1"])
parser.add_argument("--recon_loss_weights", nargs="+", default=[1.0])
# --- Controller ---
parser.add_argument(
"--processor_model",
type=str,
help="autodiff, spsa, tcn1, tcn2, proxy0, proxy1, proxy2",
)
parser.add_argument("--controller_hidden_dim", type=int, default=256)
parser.add_argument("--style_transfer", action="store_true")
# --- Encoder ---
parser.add_argument("--encoder_model", type=str, default="mobilenet_v2")
parser.add_argument("--encoder_embed_dim", type=int, default=128)
parser.add_argument("--encoder_width_mult", type=int, default=2)
parser.add_argument("--encoder_ckpt", type=str, default=None)
parser.add_argument("--encoder_freeze", action="store_true", default=False)
# --- TCN ---
parser.add_argument("--tcn_causal", action="store_true")
parser.add_argument("--tcn_nblocks", type=int, default=4)
parser.add_argument("--tcn_dilation_growth", type=int, default=8)
parser.add_argument("--tcn_channel_width", type=int, default=32)
parser.add_argument("--tcn_kernel_size", type=int, default=13)
# --- SPSA ---
parser.add_argument("--plugin_config_file", type=str, default=None)
parser.add_argument("--spsa_epsilon", type=float, default=0.001)
parser.add_argument("--spsa_schedule", action="store_true")
parser.add_argument("--spsa_patience", type=int, default=10)
parser.add_argument("--spsa_verbose", action="store_true")
parser.add_argument("--spsa_factor", type=float, default=0.5)
parser.add_argument("--spsa_parallel", action="store_true")
# --- Proxy ----
parser.add_argument("--proxy_ckpts", nargs="+")
parser.add_argument("--freeze_proxies", action="store_true", default=False)
parser.add_argument("--use_dsp", action="store_true", default=False)
parser.add_argument("--dsp_mode", choices=DSPMode, type=DSPMode)
# --- Dataset ---
parser.add_argument("--audio_dir", type=str)
parser.add_argument("--ext", type=str, default="wav")
parser.add_argument("--input_dirs", nargs="+")
parser.add_argument("--buffer_reload_rate", type=int, default=1000)
parser.add_argument("--buffer_size_gb", type=float, default=1.0)
parser.add_argument("--sample_rate", type=int, default=24000)
parser.add_argument("--dsp_sample_rate", type=int, default=24000)
parser.add_argument("--shuffle", type=bool, default=True)
parser.add_argument("--random_scale_input", action="store_true")
parser.add_argument("--random_scale_target", action="store_true")
parser.add_argument("--freq_corrupt", action="store_true")
parser.add_argument("--drc_corrupt", action="store_true")
parser.add_argument("--train_length", type=int, default=65536)
parser.add_argument("--train_frac", type=float, default=0.8)
parser.add_argument("--half", action="store_true")
parser.add_argument("--train_examples_per_epoch", type=int, default=10000)
parser.add_argument("--val_length", type=int, default=131072)
parser.add_argument("--val_examples_per_epoch", type=int, default=1000)
parser.add_argument("--num_workers", type=int, default=16)
return parser