STAR / models /diffusion.py
Yixuan Li
first commit
4853fdc
raw
history blame
52.3 kB
from typing import Sequence
import random
from typing import Any
from pathlib import Path
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.nn.functional as F
import diffusers.schedulers as noise_schedulers
from diffusers.schedulers.scheduling_utils import SchedulerMixin
from diffusers.utils.torch_utils import randn_tensor
from models.autoencoder.autoencoder_base import AutoEncoderBase
from models.content_encoder.content_encoder import ContentEncoder
from models.content_adapter import ContentAdapterBase
from models.common import LoadPretrainedBase, CountParamsBase, SaveTrainableParamsBase
from utils.torch_utilities import (
create_alignment_path, create_mask_from_length, loss_with_mask,
trim_or_pad_length
)
from safetensors.torch import load_file
class DiffusionMixin:
def __init__(
self,
noise_scheduler_name: str = "stabilityai/stable-diffusion-2-1",
snr_gamma: float = None,
cfg_drop_ratio: float = 0.2
) -> None:
self.noise_scheduler_name = noise_scheduler_name
self.snr_gamma = snr_gamma
self.classifier_free_guidance = cfg_drop_ratio > 0.0
self.cfg_drop_ratio = cfg_drop_ratio
self.noise_scheduler = noise_schedulers.DDPMScheduler.from_pretrained(
self.noise_scheduler_name, subfolder="scheduler"
)
def compute_snr(self, timesteps) -> torch.Tensor:
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = self.noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod)**0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device
)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(
device=timesteps.device
)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[...,
None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma)**2
return snr
def get_timesteps(
self,
batch_size: int,
device: torch.device,
training: bool = True
) -> torch.Tensor:
if training:
timesteps = torch.randint(
0,
self.noise_scheduler.config.num_train_timesteps,
(batch_size, ),
device=device
)
else:
# validation on half of the total timesteps
timesteps = (self.noise_scheduler.config.num_train_timesteps //
2) * torch.ones((batch_size, ),
dtype=torch.int64,
device=device)
timesteps = timesteps.long()
return timesteps
def get_target(
self, latent: torch.Tensor, noise: torch.Tensor,
timesteps: torch.Tensor
) -> torch.Tensor:
"""
Get the target for loss depending on the prediction type
"""
if self.noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif self.noise_scheduler.config.prediction_type == "v_prediction":
target = self.noise_scheduler.get_velocity(
latent, noise, timesteps
)
else:
raise ValueError(
f"Unknown prediction type {self.noise_scheduler.config.prediction_type}"
)
return target
def loss_with_snr(
self, pred: torch.Tensor, target: torch.Tensor,
timesteps: torch.Tensor, mask: torch.Tensor,
loss_reduce: bool = True,
) -> torch.Tensor:
if self.snr_gamma is None:
loss = F.mse_loss(pred.float(), target.float(), reduction="none")
loss = loss_with_mask(loss, mask, reduce=loss_reduce)
else:
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
# Adapted from https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py#L1006
snr = self.compute_snr(timesteps)
mse_loss_weights = torch.stack(
[
snr,
self.snr_gamma * torch.ones_like(timesteps),
],
dim=1,
).min(dim=1)[0]
# division by (snr + 1) does not work well, not clear about the reason
mse_loss_weights = mse_loss_weights / snr
loss = F.mse_loss(pred.float(), target.float(), reduction="none")
loss = loss_with_mask(loss, mask, reduce=False) * mse_loss_weights
if loss_reduce:
loss = loss.mean()
return loss
def rescale_cfg(
self, pred_cond: torch.Tensor, pred_cfg: torch.Tensor,
guidance_rescale: float
):
"""
Rescale `pred_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
"""
std_cond = pred_cond.std(
dim=list(range(1, pred_cond.ndim)), keepdim=True
)
std_cfg = pred_cfg.std(dim=list(range(1, pred_cfg.ndim)), keepdim=True)
pred_rescaled = pred_cfg * (std_cond / std_cfg)
pred_cfg = guidance_rescale * pred_rescaled + (
1 - guidance_rescale
) * pred_cfg
return pred_cfg
class CrossAttentionAudioDiffusion(
LoadPretrainedBase, CountParamsBase, SaveTrainableParamsBase,
DiffusionMixin
):
def __init__(
self,
autoencoder: AutoEncoderBase,
content_encoder: ContentEncoder,
content_adapter: ContentAdapterBase,
backbone: nn.Module,
duration_offset: float = 1.0,
noise_scheduler_name: str = "stabilityai/stable-diffusion-2-1",
snr_gamma: float = None,
cfg_drop_ratio: float = 0.2,
):
nn.Module.__init__(self)
DiffusionMixin.__init__(
self, noise_scheduler_name, snr_gamma, cfg_drop_ratio
)
self.autoencoder = autoencoder
for param in self.autoencoder.parameters():
param.requires_grad = False
self.content_encoder = content_encoder
self.content_encoder.audio_encoder.model = self.autoencoder
self.content_adapter = content_adapter
self.backbone = backbone
self.duration_offset = duration_offset
self.dummy_param = nn.Parameter(torch.empty(0))
def forward(
self, content: list[Any], task: list[str], waveform: torch.Tensor,
waveform_lengths: torch.Tensor, instruction: torch.Tensor,
instruction_lengths: Sequence[int], **kwargs
):
device = self.dummy_param.device
num_train_timesteps = self.noise_scheduler.config.num_train_timesteps
self.noise_scheduler.set_timesteps(num_train_timesteps, device=device)
self.autoencoder.eval()
with torch.no_grad():
latent, latent_mask = self.autoencoder.encode(
waveform.unsqueeze(1), waveform_lengths
)
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
content, content_mask = content_output["content"], content_output[
"content_mask"]
instruction_mask = create_mask_from_length(instruction_lengths)
content, content_mask, global_duration_pred, _ = \
self.content_adapter(content, content_mask, instruction, instruction_mask)
global_duration_target = torch.log(
latent_mask.sum(1) / self.autoencoder.latent_token_rate +
self.duration_offset
)
global_duration_loss = F.mse_loss(
global_duration_target, global_duration_pred
)
if self.training and self.classifier_free_guidance:
mask_indices = [
k for k in range(len(waveform))
if random.random() < self.cfg_drop_ratio
]
if len(mask_indices) > 0:
content[mask_indices] = 0
batch_size = latent.shape[0]
timesteps = self.get_timesteps(batch_size, device, self.training)
noise = torch.randn_like(latent)
noisy_latent = self.noise_scheduler.add_noise(latent, noise, timesteps)
target = self.get_target(latent, noise, timesteps)
pred: torch.Tensor = self.backbone(
x=noisy_latent,
timesteps=timesteps,
context=content,
x_mask=latent_mask,
context_mask=content_mask
)
pred = pred.transpose(1, self.autoencoder.time_dim)
target = target.transpose(1, self.autoencoder.time_dim)
diff_loss = self.loss_with_snr(pred, target, timesteps, latent_mask)
return {
"diff_loss": diff_loss,
"global_duration_loss": global_duration_loss,
}
@torch.no_grad()
def inference(
self,
content: list[Any],
condition: list[Any],
task: list[str],
instruction: torch.Tensor,
instruction_lengths: Sequence[int],
scheduler: SchedulerMixin,
num_steps: int = 20,
guidance_scale: float = 3.0,
guidance_rescale: float = 0.0,
disable_progress: bool = True,
**kwargs
):
device = self.dummy_param.device
classifier_free_guidance = guidance_scale > 1.0
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
content, content_mask = content_output["content"], content_output[
"content_mask"]
instruction_mask = create_mask_from_length(instruction_lengths)
content, content_mask, global_duration_pred, _ = \
self.content_adapter(content, content_mask, instruction, instruction_mask)
batch_size = content.size(0)
if classifier_free_guidance:
uncond_content = torch.zeros_like(content)
uncond_content_mask = content_mask.detach().clone()
content = torch.cat([uncond_content, content])
content_mask = torch.cat([uncond_content_mask, content_mask])
scheduler.set_timesteps(num_steps, device=device)
timesteps = scheduler.timesteps
global_duration_pred = torch.exp(
global_duration_pred
) - self.duration_offset
global_duration_pred *= self.autoencoder.latent_token_rate
global_duration_pred = torch.round(global_duration_pred)
latent_shape = tuple(
int(global_duration_pred.max().item()) if dim is None else dim
for dim in self.autoencoder.latent_shape
)
latent = self.prepare_latent(
batch_size, scheduler, latent_shape, content.dtype, device
)
latent_mask = create_mask_from_length(global_duration_pred).to(
content_mask.device
)
if classifier_free_guidance:
latent_mask = torch.cat([latent_mask, latent_mask])
num_warmup_steps = len(timesteps) - num_steps * scheduler.order
progress_bar = tqdm(range(num_steps), disable=disable_progress)
for i, timestep in enumerate(timesteps):
# expand the latent if we are doing classifier free guidance
latent_input = torch.cat([latent, latent]
) if classifier_free_guidance else latent
latent_input = scheduler.scale_model_input(latent_input, timestep)
noise_pred = self.backbone(
x=latent_input,
x_mask=latent_mask,
timesteps=timestep,
context=content,
context_mask=content_mask,
)
# perform guidance
if classifier_free_guidance:
noise_pred_uncond, noise_pred_content = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (
noise_pred_content - noise_pred_uncond
)
if guidance_rescale != 0.0:
noise_pred = self.rescale_cfg(
noise_pred_content, noise_pred, guidance_rescale
)
# compute the previous noisy sample x_t -> x_t-1
latent = scheduler.step(noise_pred, timestep, latent).prev_sample
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and
(i + 1) % scheduler.order == 0):
progress_bar.update(1)
waveform = self.autoencoder.decode(latent)
return waveform
def prepare_latent(
self, batch_size: int, scheduler: SchedulerMixin,
latent_shape: Sequence[int], dtype: torch.dtype, device: str
):
shape = (batch_size, *latent_shape)
latent = randn_tensor(
shape, generator=None, device=device, dtype=dtype
)
# scale the initial noise by the standard deviation required by the scheduler
latent = latent * scheduler.init_noise_sigma
return latent
class SingleTaskCrossAttentionAudioDiffusion(CrossAttentionAudioDiffusion
):
def __init__(
self,
autoencoder: AutoEncoderBase,
content_encoder: ContentEncoder,
backbone: nn.Module,
pretrained_ckpt: str | Path = None,
noise_scheduler_name: str = "stabilityai/stable-diffusion-2-1",
snr_gamma: float = None,
cfg_drop_ratio: float = 0.2,
):
nn.Module.__init__(self)
DiffusionMixin.__init__(
self, noise_scheduler_name, snr_gamma, cfg_drop_ratio
)
self.autoencoder = autoencoder
for param in self.autoencoder.parameters():
param.requires_grad = False
self.backbone = backbone
if pretrained_ckpt is not None:
pretrained_state_dict = load_file(pretrained_ckpt)
self.load_pretrained(pretrained_state_dict)
self.content_encoder = content_encoder
#self.content_encoder.audio_encoder.model = self.autoencoder
self.dummy_param = nn.Parameter(torch.empty(0))
def forward(
self, content: list[Any], condition: list[Any], task: list[str], waveform: torch.Tensor,
waveform_lengths: torch.Tensor, loss_reduce: bool = True, **kwargs
):
loss_reduce = self.training or (loss_reduce and not self.training)
device = self.dummy_param.device
num_train_timesteps = self.noise_scheduler.config.num_train_timesteps
self.noise_scheduler.set_timesteps(num_train_timesteps, device=device)
self.autoencoder.eval()
with torch.no_grad():
latent, latent_mask = self.autoencoder.encode(
waveform.unsqueeze(1), waveform_lengths
)
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
content, content_mask = content_output["content"], content_output[
"content_mask"]
if self.training and self.classifier_free_guidance:
mask_indices = [
k for k in range(len(waveform))
if random.random() < self.cfg_drop_ratio
]
if len(mask_indices) > 0:
content[mask_indices] = 0
batch_size = latent.shape[0]
timesteps = self.get_timesteps(batch_size, device, self.training)
noise = torch.randn_like(latent)
noisy_latent = self.noise_scheduler.add_noise(latent, noise, timesteps)
target = self.get_target(latent, noise, timesteps)
pred: torch.Tensor = self.backbone(
x=noisy_latent,
timesteps=timesteps,
context=content,
x_mask=latent_mask,
context_mask=content_mask
)
pred = pred.transpose(1, self.autoencoder.time_dim)
target = target.transpose(1, self.autoencoder.time_dim)
diff_loss = self.loss_with_snr(pred, target, timesteps, latent_mask, loss_reduce=loss_reduce)
return {
"diff_loss": diff_loss,
}
@torch.no_grad()
def inference(
self,
content: list[Any],
condition: list[Any],
task: list[str],
scheduler: SchedulerMixin,
latent_shape: Sequence[int],
num_steps: int = 20,
guidance_scale: float = 3.0,
guidance_rescale: float = 0.0,
disable_progress: bool = True,
**kwargs
):
device = self.dummy_param.device
classifier_free_guidance = guidance_scale > 1.0
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
content, content_mask = content_output["content"], content_output[
"content_mask"]
batch_size = content.size(0)
if classifier_free_guidance:
uncond_content = torch.zeros_like(content)
uncond_content_mask = content_mask.detach().clone()
content = torch.cat([uncond_content, content])
content_mask = torch.cat([uncond_content_mask, content_mask])
scheduler.set_timesteps(num_steps, device=device)
timesteps = scheduler.timesteps
latent = self.prepare_latent(
batch_size, scheduler, latent_shape, content.dtype, device
)
num_warmup_steps = len(timesteps) - num_steps * scheduler.order
progress_bar = tqdm(range(num_steps), disable=disable_progress)
for i, timestep in enumerate(timesteps):
# expand the latent if we are doing classifier free guidance
latent_input = torch.cat([latent, latent]
) if classifier_free_guidance else latent
latent_input = scheduler.scale_model_input(latent_input, timestep)
noise_pred = self.backbone(
x=latent_input,
timesteps=timestep,
context=content,
context_mask=content_mask,
)
# perform guidance
if classifier_free_guidance:
noise_pred_uncond, noise_pred_content = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (
noise_pred_content - noise_pred_uncond
)
if guidance_rescale != 0.0:
noise_pred = self.rescale_cfg(
noise_pred_content, noise_pred, guidance_rescale
)
# compute the previous noisy sample x_t -> x_t-1
latent = scheduler.step(noise_pred, timestep, latent).prev_sample
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and
(i + 1) % scheduler.order == 0):
progress_bar.update(1)
waveform = self.autoencoder.decode(latent)
return waveform
class DummyContentAudioDiffusion(CrossAttentionAudioDiffusion):
def __init__(
self,
autoencoder: AutoEncoderBase,
content_encoder: ContentEncoder,
content_adapter: ContentAdapterBase,
backbone: nn.Module,
content_dim: int,
frame_resolution: float,
duration_offset: float = 1.0,
noise_scheduler_name: str = "stabilityai/stable-diffusion-2-1",
snr_gamma: float = None,
cfg_drop_ratio: float = 0.2,
):
"""
Args:
autoencoder:
Pretrained audio autoencoder that encodes raw waveforms into latent
space and decodes latents back to waveforms.
content_encoder:
Module that produces content embeddings (e.g., from text, MIDI, or
other modalities) used to guide the diffusion.
content_adapter (ContentAdapterBase):
Adapter module that fuses task instruction embeddings and content embeddings,
and performs duration prediction for time-aligned tasks.
backbone:
U‑Net or Transformer backbone that performs the core denoising
operations in latent space.
content_dim:
Dimension of the content embeddings produced by the `content_encoder`
and `content_adapter`.
frame_resolution:
Time resolution, in seconds, of each content frame when predicting
duration alignment. Used when calculating duration loss.
duration_offset:
A small positive offset (frame number) added to predicted durations
to ensure numerical stability of log-scaled duration prediction.
noise_scheduler_name:
Identifier of the pretrained noise scheduler to use.
snr_gamma:
Clipping value in min-SNR diffusion loss weighting strategy.
cfg_drop_ratio:
Probability of dropping the content conditioning during training
to support CFG.
"""
super().__init__(
autoencoder=autoencoder,
content_encoder=content_encoder,
content_adapter=content_adapter,
backbone=backbone,
duration_offset=duration_offset,
noise_scheduler_name=noise_scheduler_name,
snr_gamma=snr_gamma,
cfg_drop_ratio=cfg_drop_ratio,
)
self.frame_resolution = frame_resolution
self.dummy_nta_embed = nn.Parameter(torch.zeros(content_dim))
self.dummy_ta_embed = nn.Parameter(torch.zeros(content_dim))
def forward(
self, content, duration, task, is_time_aligned, waveform,
waveform_lengths, instruction, instruction_lengths, **kwargs
):
device = self.dummy_param.device
num_train_timesteps = self.noise_scheduler.config.num_train_timesteps
self.noise_scheduler.set_timesteps(num_train_timesteps, device=device)
self.autoencoder.eval()
with torch.no_grad():
latent, latent_mask = self.autoencoder.encode(
waveform.unsqueeze(1), waveform_lengths
)
# content: (B, L, E)
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
length_aligned_content = content_output["length_aligned_content"]
content, content_mask = content_output["content"], content_output[
"content_mask"]
instruction_mask = create_mask_from_length(instruction_lengths)
content, content_mask, global_duration_pred, local_duration_pred = \
self.content_adapter(content, content_mask, instruction, instruction_mask)
n_frames = torch.round(duration / self.frame_resolution)
local_duration_target = torch.log(n_frames + self.duration_offset)
global_duration_target = torch.log(
latent_mask.sum(1) / self.autoencoder.latent_token_rate +
self.duration_offset
)
# truncate unused non time aligned duration prediction
if is_time_aligned.sum() > 0:
trunc_ta_length = content_mask[is_time_aligned].sum(1).max()
else:
trunc_ta_length = content.size(1)
# local duration loss
local_duration_pred = local_duration_pred[:, :trunc_ta_length]
ta_content_mask = content_mask[:, :trunc_ta_length]
local_duration_target = local_duration_target.to(
dtype=local_duration_pred.dtype
)
local_duration_loss = loss_with_mask(
(local_duration_target - local_duration_pred)**2,
ta_content_mask,
reduce=False
)
local_duration_loss *= is_time_aligned
if is_time_aligned.sum().item() == 0:
local_duration_loss *= 0.0
local_duration_loss = local_duration_loss.mean()
else:
local_duration_loss = local_duration_loss.sum(
) / is_time_aligned.sum()
# global duration loss
global_duration_loss = F.mse_loss(
global_duration_target, global_duration_pred
)
# --------------------------------------------------------------------
# prepare latent and diffusion-related noise
# --------------------------------------------------------------------
batch_size = latent.shape[0]
timesteps = self.get_timesteps(batch_size, device, self.training)
noise = torch.randn_like(latent)
noisy_latent = self.noise_scheduler.add_noise(latent, noise, timesteps)
target = self.get_target(latent, noise, timesteps)
# --------------------------------------------------------------------
# duration adapter
# --------------------------------------------------------------------
if is_time_aligned.sum() == 0 and \
duration.size(1) < content_mask.size(1):
# for non time-aligned tasks like TTA, `duration` is dummy one
duration = F.pad(
duration, (0, content_mask.size(1) - duration.size(1))
)
n_latents = torch.round(duration * self.autoencoder.latent_token_rate)
# content_mask: [B, L], helper_latent_mask: [B, T]
helper_latent_mask = create_mask_from_length(n_latents.sum(1)).to(
content_mask.device
)
attn_mask = ta_content_mask.unsqueeze(
-1
) * helper_latent_mask.unsqueeze(1)
# attn_mask: [B, L, T]
align_path = create_alignment_path(n_latents, attn_mask)
time_aligned_content = content[:, :trunc_ta_length]
time_aligned_content = torch.matmul(
align_path.transpose(1, 2).to(content.dtype), time_aligned_content
) # (B, T, L) x (B, L, E) -> (B, T, E)
# --------------------------------------------------------------------
# prepare input to the backbone
# --------------------------------------------------------------------
# TODO compatility for 2D spectrogram VAE
latent_length = noisy_latent.size(self.autoencoder.time_dim)
time_aligned_content = trim_or_pad_length(
time_aligned_content, latent_length, 1
)
length_aligned_content = trim_or_pad_length(
length_aligned_content, latent_length, 1
)
# time_aligned_content: from monotonic aligned input, without frame expansion (phoneme)
# length_aligned_content: from aligned input (f0/energy)
time_aligned_content = time_aligned_content + length_aligned_content
time_aligned_content[~is_time_aligned] = self.dummy_ta_embed.to(
time_aligned_content.dtype
)
context = content
context[is_time_aligned] = self.dummy_nta_embed.to(context.dtype)
# only use the first dummy non time aligned embedding
context_mask = content_mask.detach().clone()
context_mask[is_time_aligned, 1:] = False
# truncate dummy non time aligned context
if is_time_aligned.sum().item() < batch_size:
trunc_nta_length = content_mask[~is_time_aligned].sum(1).max()
else:
trunc_nta_length = content.size(1)
context = context[:, :trunc_nta_length]
context_mask = context_mask[:, :trunc_nta_length]
# --------------------------------------------------------------------
# classifier free guidance
# --------------------------------------------------------------------
if self.training and self.classifier_free_guidance:
mask_indices = [
k for k in range(len(waveform))
if random.random() < self.cfg_drop_ratio
]
if len(mask_indices) > 0:
context[mask_indices] = 0
time_aligned_content[mask_indices] = 0
pred: torch.Tensor = self.backbone(
x=noisy_latent,
timesteps=timesteps,
time_aligned_context=time_aligned_content,
context=context,
x_mask=latent_mask,
context_mask=context_mask
)
pred = pred.transpose(1, self.autoencoder.time_dim)
target = target.transpose(1, self.autoencoder.time_dim)
diff_loss = self.loss_with_snr(pred, target, timesteps, latent_mask)
return {
"diff_loss": diff_loss,
"local_duration_loss": local_duration_loss,
"global_duration_loss": global_duration_loss
}
@torch.no_grad()
def inference(
self,
content: list[Any],
condition: list[Any],
task: list[str],
is_time_aligned: list[bool],
instruction: torch.Tensor,
instruction_lengths: Sequence[int],
scheduler: SchedulerMixin,
num_steps: int = 20,
guidance_scale: float = 3.0,
guidance_rescale: float = 0.0,
disable_progress: bool = True,
use_gt_duration: bool = False,
**kwargs
):
device = self.dummy_param.device
classifier_free_guidance = guidance_scale > 1.0
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
length_aligned_content = content_output["length_aligned_content"]
content, content_mask = content_output["content"], content_output[
"content_mask"]
instruction_mask = create_mask_from_length(instruction_lengths)
content, content_mask, global_duration_pred, local_duration_pred = \
self.content_adapter(content, content_mask, instruction, instruction_mask)
scheduler.set_timesteps(num_steps, device=device)
timesteps = scheduler.timesteps
batch_size = content.size(0)
# truncate dummy time aligned duration prediction
is_time_aligned = torch.as_tensor(is_time_aligned)
if is_time_aligned.sum() > 0:
trunc_ta_length = content_mask[is_time_aligned].sum(1).max()
else:
trunc_ta_length = content.size(1)
# prepare local duration
local_duration_pred = torch.exp(local_duration_pred) * content_mask
local_duration_pred = torch.ceil(
local_duration_pred
) - self.duration_offset # frame number in `self.frame_resolution`
local_duration_pred = torch.round(local_duration_pred * self.frame_resolution * \
self.autoencoder.latent_token_rate)
local_duration_pred = local_duration_pred[:, :trunc_ta_length]
# use ground truth duration
if use_gt_duration and "duration" in kwargs:
local_duration_pred = torch.round(
torch.as_tensor(kwargs["duration"]) *
self.autoencoder.latent_token_rate
).to(device)
# prepare global duration
global_duration = local_duration_pred.sum(1)
global_duration_pred = torch.exp(
global_duration_pred
) - self.duration_offset
global_duration_pred *= self.autoencoder.latent_token_rate
global_duration_pred = torch.round(global_duration_pred)
global_duration[~is_time_aligned] = global_duration_pred[
~is_time_aligned]
# --------------------------------------------------------------------
# duration adapter
# --------------------------------------------------------------------
time_aligned_content = content[:, :trunc_ta_length]
ta_content_mask = content_mask[:, :trunc_ta_length]
latent_mask = create_mask_from_length(global_duration).to(
content_mask.device
)
attn_mask = ta_content_mask.unsqueeze(-1) * latent_mask.unsqueeze(1)
# attn_mask: [B, L, T]
align_path = create_alignment_path(local_duration_pred, attn_mask)
time_aligned_content = torch.matmul(
align_path.transpose(1, 2).to(content.dtype), time_aligned_content
) # (B, T, L) x (B, L, E) -> (B, T, E)
time_aligned_content[~is_time_aligned] = self.dummy_ta_embed.to(
time_aligned_content.dtype
)
length_aligned_content = trim_or_pad_length(
length_aligned_content, time_aligned_content.size(1), 1
)
time_aligned_content = time_aligned_content + length_aligned_content
# --------------------------------------------------------------------
# prepare unconditional input
# --------------------------------------------------------------------
context = content
context[is_time_aligned] = self.dummy_nta_embed.to(context.dtype)
context_mask = content_mask
context_mask[
is_time_aligned,
1:] = False # only use the first dummy non time aligned embedding
# truncate dummy non time aligned context
if is_time_aligned.sum().item() < batch_size:
trunc_nta_length = content_mask[~is_time_aligned].sum(1).max()
else:
trunc_nta_length = content.size(1)
context = context[:, :trunc_nta_length]
context_mask = context_mask[:, :trunc_nta_length]
if classifier_free_guidance:
uncond_time_aligned_content = torch.zeros_like(
time_aligned_content
)
uncond_context = torch.zeros_like(context)
uncond_context_mask = context_mask.detach().clone()
time_aligned_content = torch.cat([
uncond_time_aligned_content, time_aligned_content
])
context = torch.cat([uncond_context, context])
context_mask = torch.cat([uncond_context_mask, context_mask])
latent_mask = torch.cat([
latent_mask, latent_mask.detach().clone()
])
# --------------------------------------------------------------------
# prepare input to the backbone
# --------------------------------------------------------------------
latent_shape = tuple(
int(global_duration.max().item()) if dim is None else dim
for dim in self.autoencoder.latent_shape
)
shape = (batch_size, *latent_shape)
latent = randn_tensor(
shape, generator=None, device=device, dtype=content.dtype
)
# scale the initial noise by the standard deviation required by the scheduler
latent = latent * scheduler.init_noise_sigma
num_warmup_steps = len(timesteps) - num_steps * scheduler.order
progress_bar = tqdm(range(num_steps), disable=disable_progress)
# --------------------------------------------------------------------
# iteratively denoising
# --------------------------------------------------------------------
for i, timestep in enumerate(timesteps):
# expand the latent if we are doing classifier free guidance
if classifier_free_guidance:
latent_input = torch.cat([latent, latent])
else:
latent_input = latent
latent_input = scheduler.scale_model_input(latent_input, timestep)
noise_pred = self.backbone(
x=latent_input,
x_mask=latent_mask,
timesteps=timestep,
time_aligned_context=time_aligned_content,
context=context,
context_mask=context_mask
)
if classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (
noise_pred_cond - noise_pred_uncond
)
if guidance_rescale != 0.0:
noise_pred = self.rescale_cfg(
noise_pred_cond, noise_pred, guidance_rescale
)
# compute the previous noisy sample x_t -> x_t-1
latent = scheduler.step(noise_pred, timestep, latent).prev_sample
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and
(i + 1) % scheduler.order == 0):
progress_bar.update(1)
progress_bar.close()
# TODO variable length decoding, using `latent_mask`
waveform = self.autoencoder.decode(latent)
return waveform
class DoubleContentAudioDiffusion(CrossAttentionAudioDiffusion):
def __init__(
self,
autoencoder: AutoEncoderBase,
content_encoder: ContentEncoder,
content_adapter: nn.Module,
backbone: nn.Module,
content_dim: int,
frame_resolution: float,
duration_offset: float = 1.0,
noise_scheduler_name: str = "stabilityai/stable-diffusion-2-1",
snr_gamma: float = None,
cfg_drop_ratio: float = 0.2,
):
super().__init__(
autoencoder=autoencoder,
content_encoder=content_encoder,
content_adapter=content_adapter,
backbone=backbone,
duration_offset=duration_offset,
noise_scheduler_name=noise_scheduler_name,
snr_gamma=snr_gamma,
cfg_drop_ratio=cfg_drop_ratio
)
self.frame_resolution = frame_resolution
def forward(
self, content, duration, task, is_time_aligned, waveform,
waveform_lengths, instruction, instruction_lengths, **kwargs
):
device = self.dummy_param.device
num_train_timesteps = self.noise_scheduler.config.num_train_timesteps
self.noise_scheduler.set_timesteps(num_train_timesteps, device=device)
self.autoencoder.eval()
with torch.no_grad():
latent, latent_mask = self.autoencoder.encode(
waveform.unsqueeze(1), waveform_lengths
)
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
length_aligned_content = content_output["length_aligned_content"]
content, content_mask = content_output["content"], content_output[
"content_mask"]
context_mask = content_mask.detach()
instruction_mask = create_mask_from_length(instruction_lengths)
content, content_mask, global_duration_pred, local_duration_pred = \
self.content_adapter(content, content_mask, instruction, instruction_mask)
# TODO if all non time aligned, content length > duration length
n_frames = torch.round(duration / self.frame_resolution)
local_duration_target = torch.log(n_frames + self.duration_offset)
global_duration_target = torch.log(
latent_mask.sum(1) / self.autoencoder.latent_token_rate +
self.duration_offset
)
# truncate unused non time aligned duration prediction
if is_time_aligned.sum() > 0:
trunc_ta_length = content_mask[is_time_aligned].sum(1).max()
else:
trunc_ta_length = content.size(1)
# local duration loss
local_duration_pred = local_duration_pred[:, :trunc_ta_length]
ta_content_mask = content_mask[:, :trunc_ta_length]
local_duration_target = local_duration_target.to(
dtype=local_duration_pred.dtype
)
local_duration_loss = loss_with_mask(
(local_duration_target - local_duration_pred)**2,
ta_content_mask,
reduce=False
)
local_duration_loss *= is_time_aligned
if is_time_aligned.sum().item() == 0:
local_duration_loss *= 0.0
local_duration_loss = local_duration_loss.mean()
else:
local_duration_loss = local_duration_loss.sum(
) / is_time_aligned.sum()
# global duration loss
global_duration_loss = F.mse_loss(
global_duration_target, global_duration_pred
)
# --------------------------------------------------------------------
# prepare latent and diffusion-related noise
# --------------------------------------------------------------------
batch_size = latent.shape[0]
timesteps = self.get_timesteps(batch_size, device, self.training)
noise = torch.randn_like(latent)
noisy_latent = self.noise_scheduler.add_noise(latent, noise, timesteps)
target = self.get_target(latent, noise, timesteps)
# --------------------------------------------------------------------
# duration adapter
# --------------------------------------------------------------------
# content_mask: [B, L], helper_latent_mask: [B, T]
if is_time_aligned.sum() == 0 and \
duration.size(1) < content_mask.size(1):
# for non time-aligned tasks like TTA, `duration` is dummy one
duration = F.pad(
duration, (0, content_mask.size(1) - duration.size(1))
)
n_latents = torch.round(duration * self.autoencoder.latent_token_rate)
helper_latent_mask = create_mask_from_length(n_latents.sum(1)).to(
content_mask.device
)
attn_mask = ta_content_mask.unsqueeze(
-1
) * helper_latent_mask.unsqueeze(1)
align_path = create_alignment_path(n_latents, attn_mask)
time_aligned_content = content[:, :trunc_ta_length]
time_aligned_content = torch.matmul(
align_path.transpose(1, 2).to(content.dtype), time_aligned_content
)
latent_length = noisy_latent.size(self.autoencoder.time_dim)
time_aligned_content = trim_or_pad_length(
time_aligned_content, latent_length, 1
)
length_aligned_content = trim_or_pad_length(
length_aligned_content, latent_length, 1
)
time_aligned_content = time_aligned_content + length_aligned_content
context = content
# --------------------------------------------------------------------
# classifier free guidance
# --------------------------------------------------------------------
if self.training and self.classifier_free_guidance:
mask_indices = [
k for k in range(len(waveform))
if random.random() < self.cfg_drop_ratio
]
if len(mask_indices) > 0:
context[mask_indices] = 0
time_aligned_content[mask_indices] = 0
pred: torch.Tensor = self.backbone(
x=noisy_latent,
timesteps=timesteps,
time_aligned_context=time_aligned_content,
context=context,
x_mask=latent_mask,
context_mask=context_mask,
)
pred = pred.transpose(1, self.autoencoder.time_dim)
target = target.transpose(1, self.autoencoder.time_dim)
diff_loss = self.loss_with_snr(pred, target, timesteps, latent_mask)
return {
"diff_loss": diff_loss,
"local_duration_loss": local_duration_loss,
"global_duration_loss": global_duration_loss,
}
@torch.no_grad()
def inference(
self,
content: list[Any],
condition: list[Any],
task: list[str],
is_time_aligned: list[bool],
instruction: torch.Tensor,
instruction_lengths: Sequence[int],
scheduler: SchedulerMixin,
num_steps: int = 20,
guidance_scale: float = 3.0,
guidance_rescale: float = 0.0,
disable_progress: bool = True,
use_gt_duration: bool = False,
**kwargs
):
device = self.dummy_param.device
classifier_free_guidance = guidance_scale > 1.0
content_output: dict[
str, torch.Tensor] = self.content_encoder.encode_content(
content, task, device=device
)
length_aligned_content = content_output["length_aligned_content"]
content, content_mask = content_output["content"], content_output[
"content_mask"]
instruction_mask = create_mask_from_length(instruction_lengths)
content, content_mask, global_duration_pred, local_duration_pred = \
self.content_adapter(content, content_mask, instruction, instruction_mask)
scheduler.set_timesteps(num_steps, device=device)
timesteps = scheduler.timesteps
batch_size = content.size(0)
# truncate dummy time aligned duration prediction
is_time_aligned = torch.as_tensor(is_time_aligned)
if is_time_aligned.sum() > 0:
trunc_ta_length = content_mask[is_time_aligned].sum(1).max()
else:
trunc_ta_length = content.size(1)
# prepare local duration
local_duration_pred = torch.exp(local_duration_pred) * content_mask
local_duration_pred = torch.ceil(
local_duration_pred
) - self.duration_offset # frame number in `self.frame_resolution`
local_duration_pred = torch.round(local_duration_pred * self.frame_resolution * \
self.autoencoder.latent_token_rate)
local_duration_pred = local_duration_pred[:, :trunc_ta_length]
# use ground truth duration
if use_gt_duration and "duration" in kwargs:
local_duration_pred = torch.round(
torch.as_tensor(kwargs["duration"]) *
self.autoencoder.latent_token_rate
).to(device)
# prepare global duration
global_duration = local_duration_pred.sum(1)
global_duration_pred = torch.exp(
global_duration_pred
) - self.duration_offset
global_duration_pred *= self.autoencoder.latent_token_rate
global_duration_pred = torch.round(global_duration_pred)
global_duration[~is_time_aligned] = global_duration_pred[
~is_time_aligned]
# --------------------------------------------------------------------
# duration adapter
# --------------------------------------------------------------------
time_aligned_content = content[:, :trunc_ta_length]
ta_content_mask = content_mask[:, :trunc_ta_length]
latent_mask = create_mask_from_length(global_duration).to(
content_mask.device
)
attn_mask = ta_content_mask.unsqueeze(-1) * latent_mask.unsqueeze(1)
# attn_mask: [B, L, T]
align_path = create_alignment_path(local_duration_pred, attn_mask)
time_aligned_content = torch.matmul(
align_path.transpose(1, 2).to(content.dtype), time_aligned_content
) # (B, T, L) x (B, L, E) -> (B, T, E)
# time_aligned_content[~is_time_aligned] = self.dummy_ta_embed.to(
# time_aligned_content.dtype
# )
length_aligned_content = trim_or_pad_length(
length_aligned_content, time_aligned_content.size(1), 1
)
time_aligned_content = time_aligned_content + length_aligned_content
# --------------------------------------------------------------------
# prepare unconditional input
# --------------------------------------------------------------------
context = content
# context[is_time_aligned] = self.dummy_nta_embed.to(context.dtype)
context_mask = content_mask
# context_mask[
# is_time_aligned,
# 1:] = False # only use the first dummy non time aligned embedding
# # truncate dummy non time aligned context
# if is_time_aligned.sum().item() < batch_size:
# trunc_nta_length = content_mask[~is_time_aligned].sum(1).max()
# else:
# trunc_nta_length = content.size(1)
# context = context[:, :trunc_nta_length]
# context_mask = context_mask[:, :trunc_nta_length]
if classifier_free_guidance:
uncond_time_aligned_content = torch.zeros_like(
time_aligned_content
)
uncond_context = torch.zeros_like(context)
uncond_context_mask = context_mask.detach().clone()
time_aligned_content = torch.cat([
uncond_time_aligned_content, time_aligned_content
])
context = torch.cat([uncond_context, context])
context_mask = torch.cat([uncond_context_mask, context_mask])
latent_mask = torch.cat([
latent_mask, latent_mask.detach().clone()
])
# --------------------------------------------------------------------
# prepare input to the backbone
# --------------------------------------------------------------------
latent_shape = tuple(
int(global_duration.max().item()) if dim is None else dim
for dim in self.autoencoder.latent_shape
)
shape = (batch_size, *latent_shape)
latent = randn_tensor(
shape, generator=None, device=device, dtype=content.dtype
)
# scale the initial noise by the standard deviation required by the scheduler
latent = latent * scheduler.init_noise_sigma
num_warmup_steps = len(timesteps) - num_steps * scheduler.order
progress_bar = tqdm(range(num_steps), disable=disable_progress)
# --------------------------------------------------------------------
# iteratively denoising
# --------------------------------------------------------------------
for i, timestep in enumerate(timesteps):
# expand the latent if we are doing classifier free guidance
if classifier_free_guidance:
latent_input = torch.cat([latent, latent])
else:
latent_input = latent
latent_input = scheduler.scale_model_input(latent_input, timestep)
noise_pred = self.backbone(
x=latent_input,
x_mask=latent_mask,
timesteps=timestep,
time_aligned_context=time_aligned_content,
context=context,
context_mask=context_mask
)
if classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (
noise_pred_cond - noise_pred_uncond
)
if guidance_rescale != 0.0:
noise_pred = self.rescale_cfg(
noise_pred_cond, noise_pred, guidance_rescale
)
# compute the previous noisy sample x_t -> x_t-1
latent = scheduler.step(noise_pred, timestep, latent).prev_sample
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and
(i + 1) % scheduler.order == 0):
progress_bar.update(1)
progress_bar.close()
# TODO variable length decoding, using `latent_mask`
waveform = self.autoencoder.decode(latent)
return waveform